ABSTRACT
Title of dissertation: RATIONALITY AND MORAL RISK: A
MODERATE DEFENSE OF HEDGING
Christian Tarsney, Doctor of Philosophy, 2017
Dissertation directed by: Professor Dan Moller
Department of Philosophy
How should an agent decide what to do when she is uncertain not just about
morally relevant empirical matters, like the consequences of some course of action,
but about the basic principles of morality itself? This question has only recently
been taken up in a systematic way by philosophers. Advocates of moral hedging
claim that an agent should weigh the reasons put forward by each moral theory in
which she has positive credence, considering both the likelihood that that theory is
true and the strength of the reasons it posits. The view that it is sometimes rational
to hedge for one’s moral uncertainties, however, has recently come under attack both
from those who believe that an agent should always be guided by the dictates of
the single moral theory she deems most probable and from those who believe that
an agent’s moral beliefs are simply irrelevant to what she ought to do. Among the
many objections to hedging that have been pressed in the recent literature is the
worry that there is no non-arbitrary way of making the intertheoretic comparisons
of moral value necessary to aggregate the value assignments of rival moral theories
into a single ranking of an agent’s options.
This dissertation has two principal objectives: First, I argue that, contra these
recent objections, an agent’s moral beliefs and uncertainties are relevant to what she
rationally ought to do, and more particularly, that agents are at least sometimes ra-
tionally required to hedge for their moral uncertainties. My principal argument for
these claims appeals to the enkratic conception of rationality, according to which the
requirements of practical rationality derive from an agent’s beliefs about the objec-
tive, desire-independent value or choiceworthiness of her options. Second, I outline
a new general theory of rational choice under moral uncertainty. Central to this the-
ory is the idea of content-based aggregation, that the principles according to which
an agent should compare and aggregate rival moral theories are grounded in the
content of those theories themselves, including not only their value assignments but
also the metaethical and other non-surface-level propositions that underlie, justify,
or explain those value assignments.
RATIONALITY AND MORAL RISK: A
MODERATE DEFENSE OF HEDGING
by
Christian Tarsney
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland, College Park in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
2017
Advisory Committee:
Professor Dan Moller, Chair/Advisor
Professor Samuel Kerstein
Professor Christopher Morris
Professor Eric Pacuit
Professor Karol Soltan
c© Copyright by
Christian Tarsney
2017
Acknowledgments
For helpful comments and/or discussion of various parts of this dissertation,
I thank Ron Aboodi, Heather Adair, Krister Bykvist, Joseph Carlsmith, Owen
Cotton-Barratt, Mathias Frisch, Andrew Fyfe, Andrew Garland, Hilary Greaves,
Quinn Harr, Amelia Hicks, Samuel Kerstein, Daniel Kokotajlo, William MacAskill,
Dan Moller, Christopher Morris, Toby Ord, Eric Pacuit, Caleb Pickard, Douglas
Portmore, Chelsea Rosenthal, Anders Sandberg, Julius Schoenherr, Andrew Sepielli,
Paul Shephard, Joshua Shepherd, Richard Teague, Sergio Tenenbaum, Taylor White,
and many anonymous reviewers. Thanks also to audiences at the University of Mary-
land, the Uehiro Centre for Practical Ethics, the Oxford Applied Ethics Graduate
Discussion Group, the 2014 Rocky Mountain Philosophy Conference at Colorado,
the 2014 Mark L. Shapiro Graduate Philosophy Conference at Brown, the 2014
Northern Illinois Graduate Philosophy Conference, the 2015 Long Island Philosophi-
cal Society Conference, the 2015 Rocky Mountain Ethics Congress, the 2015 “Nature
and Norms” conference at Johns Hopkins, the 2016 Central APA, the 2016 Virginia
Philosophical Association Conference, the 2017 Pacific APA, the 2017 Columbia-
NYU Graduate Philosophy Conference, and the 2017 British Society for Ethical
Theory Conference. Special thanks to my adviser, Dan Moller, for tireless reading
of drafts and many hours of fruitful discussion.
Personal thanks are also due to my mother, Juli, for unstinting editorial sup-
port; to my father, Jim, for passing along a regular stream of philosophical wit-
ticisms and irrefutable arguments for the existence of an absolute up; and to my
ii
sister Catherine and girlfriend Erika for assiduously reminding me that, to most
well-adjusted people, philosophy is not that interesting.
iii
Table of Contents
Acknowledgements ii
List of Tables vii
List of Figures viii
List of Abbreviations ix
1 The Problem of Moral Uncertainty 1
1.1 Moral Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Extant Approaches and Open Problems . . . . . . . . . . . . . . . . . 8
1.3 Reasons, Oughts, and Rationality . . . . . . . . . . . . . . . . . . . . 11
1.4 Two Conceptions of Rationality . . . . . . . . . . . . . . . . . . . . . 17
1.5 Moral Reasons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.6 Simplifying Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.7 Substantive Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.8 Moral Theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.9 Summary of the Dissertation . . . . . . . . . . . . . . . . . . . . . . . 36
2 Rationality and Moral Belief 39
2.1 Moral Belief and Reasons for Action: Two Theses . . . . . . . . . . . 41
2.2 The Prima Facie Case for WSR: The Enkratic Principle . . . . . . . 47
2.3 The Prima Facie Case for SSR: Parallels between Moral and Empir-
ical Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3 Challenges to the Rational Relevance of Moral Belief 67
3.1 Normative Externalism . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.2 Weatherson on Moral Fetishism . . . . . . . . . . . . . . . . . . . . . 69
3.3 Weatherson on Prudential Uncertainty . . . . . . . . . . . . . . . . . 79
3.4 Arpaly and Schroeder on Inverse Akrasia . . . . . . . . . . . . . . . . 83
3.5 Harman on Moral Ignorance and Blame . . . . . . . . . . . . . . . . . 88
3.6 Concluding Thoughts . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
iv
4 Dominance-Based Approaches 97
4.1 Moral Nihilism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.1.1 The Cyclicity Problem . . . . . . . . . . . . . . . . . . . . . . 103
4.1.2 Nihilistic Undermining . . . . . . . . . . . . . . . . . . . . . . 106
4.2 Supererogation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.2.1 “All-Things-Considered” Supererogation . . . . . . . . . . . . 110
4.3 Rejecting Supererogationism: Cyclicity and Undermining . . . . . . . 116
4.3.1 Cyclicity and Supererogation . . . . . . . . . . . . . . . . . . 116
4.3.2 Undermining and Supererogation . . . . . . . . . . . . . . . . 117
4.4 Practical Upshots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.5 Can Dominance Reasoning Stand on Its Own? . . . . . . . . . . . . . 129
5 Intertheoretic Value Aggregation 133
5.1 Two More Challenges to Hedging . . . . . . . . . . . . . . . . . . . . 133
5.2 Moral and Empirical Uncertainty, Revisited . . . . . . . . . . . . . . 140
5.3 Case Studies in Comparability . . . . . . . . . . . . . . . . . . . . . . 146
5.3.1 First Case Study: Pluralistic Consequentialism . . . . . . . . . 147
5.3.2 Second Case Study: Absolutist Deontology . . . . . . . . . . . 152
5.4 Is Incomparability Anywhere a Threat to Comparability Everywhere? 156
5.5 “Top-Down” vs. “Content-Based” Approaches to Intertheoretic Ag-
gregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6 Outline of a General Theory 168
6.1 Three Components of Comparability . . . . . . . . . . . . . . . . . . 169
6.2 Structural Compatibility . . . . . . . . . . . . . . . . . . . . . . . . . 172
6.2.1 Structural Enrichment and Structural Depletion . . . . . . . . 182
6.2.2 Multi-Stage Aggregation . . . . . . . . . . . . . . . . . . . . . 187
6.3 Scale Normalizability . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
6.3.1 Extending Content-Based Normalization . . . . . . . . . . . . 202
6.3.2 Normalization Failure: Incomparability/Binary Normalization 215
6.3.3 Top-down Normalization . . . . . . . . . . . . . . . . . . . . . 219
6.4 Decision-Theoretic Compatibility . . . . . . . . . . . . . . . . . . . . 225
6.5 Rational Choice under Normative Uncertainty: A Tentative General
Account . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
7 The Regress Problem 234
7.1 The Regress Argument for Externalism . . . . . . . . . . . . . . . . . 236
7.1.1 Fixed Point Solutions: Convergence or Contraction . . . . . . 239
7.1.2 Sepielli’s View . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
7.1.3 Enkratic Externalism . . . . . . . . . . . . . . . . . . . . . . . 246
7.1.4 Which Enkratic Principle? . . . . . . . . . . . . . . . . . . . . 250
7.2 Two Regress Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 254
7.3 The Problem of Ideal Rationality . . . . . . . . . . . . . . . . . . . . 255
7.3.1 Gappiness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
7.3.2 Normalization Uncertainty . . . . . . . . . . . . . . . . . . . . 260
v
7.4 The Problem of Non-Ideal Rationality . . . . . . . . . . . . . . . . . 262
7.4.1 Consequentialism and the “Heuristic Regress” . . . . . . . . . 262
7.4.2 Non-Ideal Deliberation under Moral Uncertainty: The Power
of Priors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
7.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
A Moral Uncertainty for Deontologists 273
A.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
A.2 Absolutist Deontology and Uncertainty . . . . . . . . . . . . . . . . . 276
A.3 Stochastic Dominance . . . . . . . . . . . . . . . . . . . . . . . . . . 283
A.4 Stochastic Dominance and Deontological Uncertainty . . . . . . . . . 286
A.5 Option Individuation and Ought Agglomeration . . . . . . . . . . . . 295
A.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
B What Grounds Intertheoretic Comparisons of Moral Value 305
B.1 The Easy Case: Decision-Theoretic Compatibility . . . . . . . . . . . 306
B.2 The Easy Case: Structural Compatibility and Scale Normalizability . 312
B.3 A Simple Argument for Hedging . . . . . . . . . . . . . . . . . . . . . 319
B.4 Generalizing (or Failing to Generalize) from the Easy Case . . . . . . 322
B.4.1 Indirect Comparability . . . . . . . . . . . . . . . . . . . . . . 323
B.4.2 Uncertainty about Value Independence . . . . . . . . . . . . . 327
B.4.3 “Infectious Incomparability” . . . . . . . . . . . . . . . . . . . 332
B.4.4 Incomparability Simpliciter . . . . . . . . . . . . . . . . . . . 333
B.5 What Grounds Credence in Comparisons? . . . . . . . . . . . . . . . 337
Bibliography 343
vi
List of Tables
6.1 Dorothy’s Dilemma, Stage 1 . . . . . . . . . . . . . . . . . . . . . . . 189
6.2 Dorothy’s Dilemma, Stage 2 . . . . . . . . . . . . . . . . . . . . . . . 190
6.3 George’s Gamble, Pre-Normalization . . . . . . . . . . . . . . . . . . 226
6.4 George’s Gamble, Normalized . . . . . . . . . . . . . . . . . . . . . . 227
A.1 Possible Promise v1, payoff matrix . . . . . . . . . . . . . . . . . . . 289
A.2 Possible Promise v1, probability of payoff ≥ x . . . . . . . . . . . . . 289
A.3 Possible Promise v2, payoff matrix . . . . . . . . . . . . . . . . . . . 289
A.4 Possible Promise v2, probability of payoff ≥ x . . . . . . . . . . . . . 290
A.5 Possible Promise v3, payoff matrix . . . . . . . . . . . . . . . . . . . 294
A.6 Possible Promise v3, probability of payoff ≥ x . . . . . . . . . . . . . 294
vii
List of Figures
7.1 Non-Ideal Practical Deliberation . . . . . . . . . . . . . . . . . . . . . 268
viii
List of Abbreviations
DNET Democritean normative error theory
ECS exclusionary conception of supererogation
EE enkratic externalism
EP enkratic principle
EPP egoistic permissions principle
FDOT Final Dominance over Theories
GDoT Genuine Dominance over Theories
GDoT∗ Strengthened Genuine Dominance over Theories
MEC maximize expected choiceworthiness
MFO My Favorite Option
MFT My Favorite Theory
PEMT Principle of Equity among Moral Theories
PES principle of equal say
PIVA problem of intertheoretic value aggregation
SCS satisficing conception of supererogation
SD stochastic dominance
SSR Strong Subjective Rationalism
UCBN universal content-based normalization
WSR Weak Subjective Rationalism
ix
Chapter 1: The Problem of Moral Uncertainty
Being a conscientious person means trying to do the right thing. But trying
to do the right thing, even trying one’s best, is no guarantee of success, for we often
face substantial obstacles to acting well or rightly. One kind of obstacle is weakness
of will, and challenges to the will like fear, temptation, or distraction. Another kind
of obstacle is ignorance or uncertainty about the good or the right, about what we
ought to do. Sometimes the source of uncertainty is empirical, e.g., uncertainty
concerning the consequences of some available course of action. But sometimes we
are also relevantly uncertain about the basic requirements of morality, e.g., about
whether or to what extent some feature of the world like happiness, beauty, or
knowledge has non-derivative value, or about whether the right thing to do is always
the thing that maximizes the sum of value in the world.
This latter form of uncertainty, and how we can try our best to do the right
thing in spite of it, is the topic of this dissertation. Following convention, I will use
the term moral uncertainty to refer exclusively to this sort of “purely moral” un-
certainty, uncertainty about basic moral principles as opposed to uncertainty about
morally relevant empirical facts. As I hope will become clear, moral uncertainty
even in this restricted sense is a pervasive feature of everyday life: Many basic
1
questions about the requirements of morality remain unresolved despite millennia
of philosophical effort. Moreover, many of the accounts of morality that philoso-
phers have found most plausible make demands of which we routinely fall short, and
that in many cases conflict with one another. Lying often seems likely to yield bet-
ter overall consequences than truth-telling, putting Kantian and utilitarian theories
in conflict. Commonsense morality often demands that we prioritize our friends or
family members in a way that impartial, universalistic theories like both Kantianism
and utilitarianism forbid. Egalitarian theories of justice demand that we support
the socialist candidate for public office while historical entitlement theories demand
that we support the libertarian. Therefore our uncertainty about morality is not
just theoretical but practical: Any agent, no matter how conscientious, must often
be unsure whether she is doing the right thing.
1.1 Moral Uncertainty
It may be tempting, in the face of such worries, to simply deny that we are
relevantly uncertain about the basic requirements of morality. Especially when we
focus on the most extreme demands of a given moral theory, or on cases where
the theory conflicts with our own strongly held moral or political beliefs, we may
be able to talk ourselves at least momentarily into the view that such a theory is
certainly false. The importance of knowing how to make decisions in the face of
moral uncertainty will depend very much, then, on how often we should really be
uncertain, and how uncertain we should be, about the demands of morality with
2
respect to the morally significant choices we face.
It is a trivial consequence of certain views about rational belief that we should
always be at least somewhat uncertain about basic moral questions. In particular,
if one accepts probabilism, according to which rational belief should come in de-
grees that obey the axioms of the probability calculus, and if one further accepts
the requirement of regularity, according to which the probability assigned to any
proposition that is not a logical truth or falsehood must be strictly greater than 0
and less than 1, then it seems we are rationally required to maintain at least some
degree of uncertainty about any non-trivial moral proposition.
As I will explain in §1.7, I am sympathetic to these claims. But they alone
are not enough to make the problem of moral uncertainty interesting, for our moral
uncertainties about the choices we face as moral agents might nevertheless be so
slight as to be practically irrelevant on any plausible account of rational choice
under moral uncertainty. For instance, accepting regularity commits one to non-zero
credence is the proposition that it is always morally obligatory to torture babies for
fun. Nevertheless, one’s degree of belief in this proposition may still be so close to
zero that the reasons for action it generates, if any, are vanishingly weak, to the
point that it is not worth attempting to explicitly account for them in practical
reasoning.
But whether or not one accepts the general commitments of probabilism and
regularity, there are more mundane, and more compelling, reasons for thinking that
we ought to be substantially uncertain about a great many fundamental moral ques-
tions, and in a way that is practically relevant to a great many of our choices. First is
3
the fact of widespread, intractable, and apparently reasonable disagreement among
intelligent, reflective, and well-motivated moral agents. Among philosophers, cen-
turies of debate between consequentialists, Kantians, virtue ethicists, et al have
so far failed to yield consensus, and indeed the variety of moral theories that can
claim active adherents in the philosophical literature continues to expand rather
than contract. Likewise, traditional religious moral codes differ considerably in
the requirements and prohibitions they enjoin (e.g. with respect to sexual morality,
consumption of intoxicants, the relative weight of obligations to the poor, family
members, ancestors, and religious institutions, etc), and these disagreements have
proven no more tractable. Further, as any philosopher who has taught an undergrad-
uate course on contemporary moral controversies can attest, even those without any
antecedent commitment to either a philosophical theory of morality or a religious
moral upbringing can differ greatly in their intuitive moral judgments concerning
our obligations to non-human animals, the distant poor, or future generations, or
concerning the permissibility of assisted suicide, killing civilians in war, abortion
and infant euthanasia, and too many other topics to name. Thus, whatever popu-
lation one takes to be the appropriate reference class of competent moral judges, it
must be conceded that one’s own moral convictions are far from universally shared
within that class, and indeed that one’s own comprehensive set of moral beliefs is
almost certainly shared by a tiny fraction of moral judges at best.
In addition to these ongoing moral disagreements, the degree to which moral
opinion has changed over time should challenge our confidence even in those moral
judgments that are widely shared among present-day moral agents. Practices that
4
have been generally accepted or encouraged in many times and places, like slavery,
infant exposure, pederasty, and arranged marriage are now widely or universally
viewed as morally repugnant, while requirements once seen as stringent, say with
respect to personal or familial honor or sexual chastity, have lost their stringency or
dropped out of our shared moral code altogether. There are good reasons to regard
many of these belief changes as moral progress—we need not be too troubled by
our disagreement with our ancestors regarding, say, the permissibility of declining
a duel. But reflection on this history should lead us to conclude, by pessimistic
induction, that many of our own moral beliefs will likely be seen as absurd and
perverse by our own descendants, and likely in some respects that even the most
morally reflective among us would be unable to guess.
Likewise, we ourselves may have experienced significant changes of moral belief
in our lifetimes. Indeed, it seems likely to me that anyone who has thought long
about questions of moral theory has experienced at least one such change in their
initial convictions. Here again, though we may have good reason to regard our past
opinions as wrongheaded, memories of misplaced confidence in those past opinions
should leave us open to the possibility of further significant revisions in future.1
Further grounds for modesty in our moral beliefs comes from the finding that
we are disposed toward overconfidence across a wide variety of doxastic domains.2
In general, where such assessment is possible, we find that people’s explicit degrees
of belief display significant calibration errors in the direction of excessive confidence.
1For remarks on these and related considerations, see Moller (2011, pp. 432-4).
2For a useful overview of recent research on miscalibration and overconfidence, see Part III of
Kahneman (2011).
5
For example, when faced with questions that have a well-ordered range of possible
answers (e.g. dates or quantities) and asked to produce, say, 90% or 95% confidence
intervals on this range (such that one is 90% or 95% confident that the correct
answer falls in between the lower bound and upper bound), most of us will produce
intervals that fail to cover the correct answer at far more than the expected rate
of 10% or 5% (Soll and Klayman, 2004). Likewise, when asked to assign degrees of
confidence to one’s answers to common-knowledge questions, our stated confidence
tends to substantially exceed our measured reliability (Alpert and Raiffa, 1982).
Of course, in the moral context we have no means of checking the reliability of a
given individual’s moral judgments, but insofar as the finding of overconfidence can
be inductively generalized to domains in which it cannot be directly confirmed, we
have reason to suspect that our confidence in our moral judgments may be prone to
this same bias.
A final reason for general uncertainty about questions of morality is the lack
of well-established methods or foundations for moral theorizing. Proposed starting
points for moral theory include conceptual truths about action and rational agency
(e.g. Korsgaard (1996)), self-evident axioms about the good and the right (e.g.
Sidgwick (1874), de Lazari-Radek and Singer (2014)), the considered case judgments
of reflective moral agents (e.g. Rawls (1951), Daniels (1979)), the predicted choices of
agents in idealized bargaining conditions (e.g. Gauthier (1986)), intentional design of
moral principles to meet the needs of human societies, anthropological observation of
the moral practices of existing societies, scientific study of the evolutionary function
of moral norms, and many more. The most popular method of investigation in
6
contemporary analytic moral philosophy, the method of reflective equilibrium based
on heavy appeal to intuitive judgments about cases, has come under concerted
attack and is regarded by many philosophers (e.g. Singer (2005), Greene (2008)) as
deeply suspect. Additionally, every major theoretical approach to moral philosophy
(whether at the level of normative ethics or metaethics) is subject to important and
intuitively compelling objections, and the resolution of these objections often turns
on delicate and methodologically fraught questions in other areas of philosophy
like the metaphysics of consciousness or personal identity (Moller, 2011, pp. 428-
432). Whatever position one takes on these debates, it can hardly be denied that
our understanding of morality remains on a much less sound footing than, say,
our knowledge of the natural sciences. If, then, we remain deeply and justifiably
uncertain about a litany of important questions in physics, astronomy, and biology,
we should certainly be at least equally uncertain about moral matters, even when
some particular moral judgment is widely shared and stable upon reflection.
The only plausible way one might deny that we should be substantially uncer-
tain about many important moral matters, it seems to me, is to deny that morality
is a matter of belief in the first place—that is, to adopt a non-cognitivist account
of moral judgment. Still, even non-cognitivists must account for the appearance of
moral uncertainty, for the fact that something very much like uncertainty is un-
questionably a part of our moral experience. Some philosophers (e.g. Smith (2002))
have argued that non-cognitivism must be rejected on the grounds that it cannot
adequately account for the uncertainty-like features of moral judgment. Others (e.g.
Sepielli (2012)) have proposed sophisticated versions of non-cognitivism meant to
7
address these objections and explain how non-doxastic attitudes can display fea-
tures closely analogous to uncertainty. I have nothing to add to these debates and
will simply assume, as others in the literature have done, that any adequate version
of non-cognitivism must allow for features of moral judgment structurally similar
enough to what the cognitivist calls “moral uncertainty” that discussions of the phe-
nomenon conducted in the language of cognitivism can be readily translated into
the non-cognitivist framework.
1.2 Extant Approaches and Open Problems
In light of the pervasiveness of moral uncertainty (or its non-cognitivist ana-
logue), and of compelling reasons for that uncertainty, it is more than a little remark-
able that the question of how agents should rationally resolve moral dilemmas in
the face of such uncertainty received almost no attention from analytic philosophers
until quite recently.3 The recent spate of philosophical interest in the problem dates
to the publication of Ted Lockhart’s Moral Uncertainty and Its Consequences (Lock-
hart, 2000), which developed an ambitious theory of expected value maximization in
the face of moral uncertainty.4 Unfortunately Lockhart’s proposal, which turns on a
3Speculatively, it may be that the increased focus in the second half of the 20th century on
moral dilemmas like the trolley problem, as well as questions of practical ethics like philanthropic
obligations to the distant poor and the moral status of animals, led moral philosophers to pay
more attention to the practical conflicts between prominent moral theories like Kantianism and
utilitarianism and hence to the need for practical guidance in the face of uncertainty concerning
these cases of practical conflict.
4The philosophical literature on choice under moral uncertainty before Lockhart comprises
a very few papers, chiefly Greenwell (1977), Lockhart (1977), Pfeiffer (1985), Hudson (1989),
Oddie (1994), and Gracely (1996). There is also, interestingly, a much earlier and fairly extensive
literature from Catholic moral theology dealing with a particular aspect of the problem of moral
uncertainty, namely, how confident one must be that an act is objectively permissible rather than
forbidden before one is subjectively permitted to perform it. This debate mainly pitted rigoristic
Jansenists (including Pascal (1657)), who demanded near-certainty before a morally risky act could
8
clever but counterintuitive account of how to make quantitative value comparisons
across rival moral theories, has been subjected to powerful criticisms (especially by
Sepielli (2006, 2013)) and has found no subsequent defenders in the literature.
In the years since Lockhart, other philosophers have suggested partial or gen-
eral solutions to the problem of intertheoretic value comparisons (e.g. Ross (2006),
Sepielli (2009)), while still others have offered arguments that the problem is unsolv-
able (e.g. Gustafsson and Torpman (2014), Nissan-Rozen (2015)). We will examine
these debates at length in Chapters 5-6.
Others, meanwhile, have explored the practical implications of moral uncer-
tainty with respect to particular moral dilemmas. Most of this literature has fo-
cused on cases of “moral risk” (actions that seem to be either forbidden or merely-
permissible), rather than cases of “moral conflict” (actions that may be either for-
bidden or obligatory), with the most widely discussed cases being abortion and veg-
etarianism. Prominent in this literature are Ross (2006), Guerrero (2007), Moller
(2011), and MacAskill (2013).
Finally, in the last few years, there has been concerted pushback against the
very idea that what an agent ought to do is sensitive to her uncertainties about
basic moral questions and hence that there is need for a theory of rational choice
under moral uncertainty. The extreme position in this literature (exemplified by
Weatherson (2014), Harman (2015), and Hedden (2016)) claims that what an agent
be contemplated, against the more laissez-faire Jesuits. (For a useful overview , see Sepielli (2010,
pp. 48-53).) Kant was a latecomer to this debate, opining in his Religion within the Limits of
Reason Alone that “it is a basic moral principle, which requires no proof, that one ought to hazard
nothing that may be wrong” (Kant, 1793, p. 173; cited in Kerstein, 2002, p. 214). (Thanks to Dan
Moller for bringing this passage to my attention.)
9
ought to do is simply insensitive to her moral beliefs, while a more moderate position
(exemplified by Gustafsson and Torpman (2014) and Nissan-Rozen (2015)) merely
denies that agents should engage in expected value-style weighing of the reasons
posited by rival moral theories. We will consider several of the concerns raised in
this literature, beginning in Chapter 3.
Despite this recent flurry of interest, most of the central problems associated
with moral uncertainty still lack any satisfactory resolution, as we will see. In
addition to the extreme lack of consensus regarding whether and in what ways
the normative requirements incumbent on an agent depend on her moral beliefs
(and, indeed, lack of clarity concerning the range of possible views on this question
and which normative concepts it centrally concerns), no approach to the problem of
intertheoretic value comparisons has achieved anything like widespread support even
among those philosophers who agree that such an approach is needed. In addition,
consideration of the problem of moral uncertainty draws attention to a deeper and
more general problem concerning the nature of subjective normativity, the so-called
“regress problem” which we will take up in Chapter 7, a problem that has been
widely recognized but to which the extant literature contains only one or two hints
at potential solutions.
Fortunately, recent developments in closely related philosophical literatures
have given us new tools for approaching many of the problems related to moral
uncertainty. Recent work on rationality, stemming from Broome (1999, 2001) and
Kolodny (2005) inter alia, has furnished more sophisticated ways of thinking about
both the nature and the content of rational requirements. Attention among metaethi-
10
cists to the relationship and relative priority of objective (“fact-relative”), subjective
(“belief-relative”), and prospective (“evidence-relative”) normativity (Zimmerman
(2008) and Parfit (2011), among many others) has sharpened these distinctions
and brought clarity to long-standing debates between “objectivists” and “subjec-
tivists” about moral obligation. A closely related debate on the relationship between
moral ignorance and culpability, stemming largely from work by Michael Zimmer-
man (1997) and Gideon Rosen (2002, 2004) has helped to illuminate questions of the
normative force of moral belief from a different direction. And work on value incom-
mensurability/incomparability, driven forward in large part by the essays contained
in Chang (1997a), has yielded new technical approaches to the apparent incompa-
rability of disparate ethical considerations that is the central difficulty for theories
of choice under moral uncertainty.
Despite its many difficulties, then, it is reasonable to hope that we will be able
to make substantial progress on the problem of moral uncertainty in the coming
years, at least to the point of offering more complete and more plausible statements
of the various rival views than are presently available. In the pages that follow, I
aim to make some contribution toward this goal.
1.3 Reasons, Oughts, and Rationality
It will be useful to spend some time setting out a repertoire of basic con-
cepts to which I will regularly appeal in later chapters. I take the domain of
practical reasoning to be characterizable in terms of four fundamental concepts:
11
agents, options, objective reasons/oughts, and subjective reasons/oughts. To the
extent that other practical concepts like rationality/rational requirement, obliga-
tion/permission/prohibition, moral rightness/wrongness, praiseworthiness/blame-
worthiness, and so on are meaningful, it seems to me, they must be either interde-
finable with or definable in terms of these four basic concepts. Short of examining
the other candidates for fundamental practical concepts one at a time, I can defend
this meager conceptual base only by an appeal to parsimony and by demonstrating
in subsequent chapters how much can be said in the terms in offers.
Agents we can characterize simply as the sort of beings capable of choosing
between options: the conceptual and empirical criteria for practical agency have
little affect on the questions we will be considering. About options, a bit more
needs to be said. Options are potential objects of choice for agents—that is, things
that an agent can choose.5 I will generally speak of “options” rather than “actions”
as the objects of practical choice since an agent’s options can include omissions,
sequences of actions, and perhaps also general strategies/plans/dispositions to act
that don’t correspond to any specifically contemplated action. Nevertheless, actions
and courses of action are paradigmatic examples of practical options.
In general I will assume that we are concerned with choice between maximal
options, that is, options that are fully specified from an agent’s perspective, having
no more specific variants that the agent is capable of choosing between. For instance,
going to the store is an option of which walking to the store and driving to the store
5One might, I suppose, take the notion of “choosing” to be fundamental and the notion of an
“option” derivative, but this won’t matter for our purposes.
12
are variants, so going to the store is non-maximal. If the agent anticipates, as she
mulls over her options, that she can drive to the store via by way of either Oak St
or Peach St, then “driving to the store” will not count as a maximal option for her
either, since it has these more specific variants. It is very difficult to give any exact
criterion for what an agent’s options are at a given moment, and hence what her
maximal options are, but we can safely leave this question unanswered.
Choosing an option is importantly different from intending to choose or to
perform an option. By “choosing option O” I mean something like “attempting to
perform option O”: It is useful to talk of “choosing O” rather than “O-ing” since
an agent may choose an option (e.g., paradigmatically, an action) and yet fail to
perform it, and may perhaps perform an option (“ ”) accidentally or involuntarily.
Forming an intention to perform an option O means making up one’s mind that one
will, at least a short time in the future, perform option O. Choosing O is, as I am
using the expression, just the first step in the process of actually performing O. For
example, an agent who manages to best Kavka’s toxin puzzle (Kavka, 1983) intends
to drink the toxin at midnight, but does not choose to drink the toxin unless and
until she orders her hand to begin reaching for the vial. It is cases like this that
highlight the need for practical principles to distinguish between intending an option
and choosing/performing that option.6
6What I have said does not rule out rational requirements to form intentions. Rather, it simply
distinguishes between a rational requirement to choose option O (which might follow from a belief
that one ought to choose option O) and a rational requirement to intend to choose option O
(which might follow from a belief that one ought to intend to choose option O). Very often these
requirements will coincide—perhaps they will always coincide when the action in question is to be
performed immediately. But they can also come apart, not just in anomalous cases like Kavka’s,
but also in the ordinary case where one believes that one is (now) rationally required to intend
to perform some action in the future, but does does believe that one is (now) rationally required
13
Where merely schematic description is called for, I will denote agents by the
terms A, B, C... or A1, A2, A3... as the situation demands, and likewise to options
by the terms O, P, Q... or O1, O2, O3...
The two remaining elements in our repertoire of fundamental concepts, about
which most needs to be said, are objective reasons/oughts and subjective rea-
sons/oughts. I say that these are, together, just two fundamental concepts rather
than four because I take reasons and oughts to be interdefinable. Roughly, an
(objective/subjective) “ought” claim asserts that an agents has on-balance or all-
things-considered (objective/subjective) reason to choose a particular option from
some set of alternatives. Conversely, an (objective/subjective) reason to choose O
is just a consideration that counts in favor of its being the case that the agent (ob-
jectively/subjectively) ought to choose O.7 Whether one takes reasons or oughts as
the more fundamental notion out of each pair is, I think, not particularly important.
But for expository purposes, it is perhaps a bit easier to start with reasons.8
to choose/attempt to perform that action. For a more extended argument to similar effect see
Reisner (2013), who suggests that all rational requirements must take the form of “matching
attitude requirements” (requiring that an agent’s attitudes match her beliefs about what attitudes
she ought to have) and that therefore a rational requirement to intend can follow only from a belief
about what one ought to intend, not from a belief about what one ought to do.
7I can only acknowledge in passing that these interdefinitions contradict the views of various
philosophers (e.g. Foot (1972), Horty (2014)) according to whom true “ought” claims can fail to
correspond to reasons for action. But this dispute may be merely verbal: it seems to me that I
have in mind a different, more restrictive sense of “ought” than these philosophers.
8I don’t want to say a great deal about the relationship between reasons and oughts, as I don’t
have a great deal to say and the details will generally be unimportant for our purposes. But I
will take for granted that the relationship is monotonic: That is to say, both reasons and oughts
express two-place relations between agents and options, and if given a certain set of reasons it
is the case that agent A ought to choose option O, then the addition to A’s set of reasons of a
further reason for A to choose O cannot make it no longer the case that A ought to choose O,
except insofar as the addition of that reason to choose O requires the withdrawal of other reasons
to choose O, the addition of reasons not to choose O, or the addition of reasons for/withdrawal of
reasons against choosing some alternative to O. (Reasons against choosing O can be thought of as
reasons for choosing the non-maximal option ¬O, or as reasons for choosing each of the maximal
alternatives to O.)
14
If the concepts of objective and subjective reasons are indeed basic, then it
is of course no use attempting to give genuine definitions for them. The best I
can do, in the style of Wittgenstein’s characterization of “ethics” (Wittgenstein,
1965), is to attempt a few roughly synonymous expressions of each concept, and
hope that the reader ends up with the intended concept in mind. Reasons are, then,
“considerations that count in favor of actions.” Objective reasons are considerations
that count in favor of action objectively, in light of the way the world really is and
independent of the agent’s beliefs about the world (except in cases where these
beliefs are relevant features of the world, e.g., when an agent has objective reason
to take some course of action that would remedy her ignorance or false beliefs).
Very roughly, objective reasons count in favor of action “from the standpoint of
omniscience,” and correspond in general (though not as a matter of definition or
conceptual truth) to the considerations that would influence an agent’s epistemically
(and perhaps motivationally) idealized counterpart in wanting or advising the agent
herself to adopt a certain course of action. In Parfit’s well-known idiom, therefore,
objective reasons are “fact-relative” rather than “belief-relative” (Parfit, 2011, p.
150).
As objective reasons count in favor of action “from the standpoint of omni-
science,” so subjective reasons count in favor of action from an agent’s own stand-
point, and in particular from her doxastic standpoint. In Parfit’s idiom, they are
“belief-relative” (and indeed, as it will be the major task of the coming chapters
to argue, moral -belief-relative). In the classic case from Williams (1979), an agent
who desires to drink from a glass that she believes to contain gin but that in fact
15
contains gasoline has objective reason not to drink from the glass, in virtue of the
fact that it contains gasoline, but has subjective reason to drink from the glass, in
virtue of her belief that it contains gin. More particularly, an agent’s subjective
reasons depend either on her (“de dicto”) beliefs about what objective reasons she
has, or on her (“de re”) beliefs about the properties that underwrite or subvene on
her subjective reasons (of which distinction we will have much more to say in the
coming chapters).
It is important to be clear that this objective/subjective distinction is not the
distinction between agent-relative and non-agent-relative reasons. While objective
reasons are (generally) independent of an agent’s beliefs, they need not be inde-
pendent of her desires or other aspects of her mental state. For instance, in the
Williams case, the agent’s objective reason not to drink from the glass does not de-
pend on her belief but may well depend (as Williams would have us believe) on her
desires, e.g., to avoid suffering and death. So talk of “objective reasons” does not
by itself commit us to the existence of objective-in-the-sense-of-desire-independent
value/disvalue, rightness/wrongness, etc. (This point will be further clarified in the
next section when we discuss the Humean conception of rationality.) Moreover, even
desire-independent objective reasons may still be agent-relative, e.g., each agent may
have special reason to pursue her own objective welfare, where this consists of a list
of objective goods whose contents are independent of her desires.
To reiterate the assumed relationship between reasons and oughts, an agent A
objectively ought to choose option O iff she has on-balance, all-things-considered ob-
jective reason to choose O, and subjectively ought to choose O iff she has on-balance,
16
all-things-considered subjective reason to choose O. I take subjective oughts to be
synonymous with rational requirements : Agents are, always and only, rationally re-
quired to do what they have on-balance subjective reason to do. To be rational is to
non-accidentally satisfy one’s rational requirements: that is, to be sensitive to one’s
reasons in such a way that one generally does what one has on-balance subjective
reason to do. I will frequently use “A subjectively ought to choose O” and “A is
rationally required to choose O” interchangeably, as two ways of saying the same
thing.9
1.4 Two Conceptions of Rationality
The concepts of reasons and oughts admit of two very different conceptions,
each of which has been defended in various forms in the philosophical literature.
According to what I will call the Humean conception of rationality: (i) An agent
A has objective reason to choose an option O iff choosing O would in some way
9Kolodny (2005) has influentially argued that there are no general reasons to comply with
rational requirements and that the apparent normativity of rational requirements stems from the
fact that, whenever an agent is rationally required to choose option O, she is so required because she
believes herself to have reason of some more mundane sort for choosing option O. I am generally in
sympathy with Kolodny’s account of rationality, and I take our disagreement with respect to the
relationship between reasons and rational requirements to be merely verbal: Kolodny does not draw
a distinction between objective and subjective reasons, reserving the term “reason” exclusively for
what I have been calling “objective reasons.” But everything I have said in the preceding section
can be paraphrased in Kolodny’s idiom if we simply take a “subjective reason” to choose option O
as a positive degree of belief that one has some objective reason to choose O, on the assumption
that such degrees of belief contribute monotonically toward generating a rational requirement to
choose O. (Strictly, the account of rational requirement articulated in Kolodny (2005) implies
that only full belief that one has objective reason to choose O can contribute toward generating
rational requirements, but this feature of the view clearly requires amendment: even very partial
belief that, say, my drink has been poisoned and that I therefore have objective reason not to
drink it can generate a rational requirement against drinking.) On this way of putting things,
then, one feature of Kolodny’s view is that reasons are fundamental in the domain of objective
normativity (that is, more fundamental than objective “oughts”/obligations/requirements), which
rational requirements are fundamental in the domain of subjective normativity (that is, more
fundamental than “subjective reasons”). But this is compatible with the framework I have adopted.
17
contribute to the satisfaction of her desires (or other motivational states); (ii) A
objectively ought to choose O iff O satisfies her desires, on balance, better than any
alternative option; and (iii) A has subjective reason to choose O iff she has positive
degree of belief that choosing O would in some way contribute to the satisfaction of
her desires (“ ”).
The notion of subjective oughts/rational requirements requires a bit more care.
As I have said, I take it to be a conceptual truth that an agent subjectively ought to
do what she has most subjective reason to do, so we may say that on the Humean
conception of rationality, A subjectively ought to choose O iff her degrees of belief
concerning the satisfaction or dissatisfaction of her various desires conditional on
choosing each of the options available to her favor O above the available alterna-
tives. But this leaves open the question how degrees of subjective-reason-giving
weight are assigned to an agent’s degrees of belief concerning the satisfaction or dis-
satisfaction of her desires conditional on choosing a particular option. The standard
answer to this question is that an agent subjectively ought to maximize the expected
satisfaction of her desires, i.e., that the weight of a subjective reason for/against
choosing O is simply the product of the strength of the underlying desire times the
agent’s degree of belief that choosing O will result in the satisfaction/dissatisfaction
of that desire. But, as we will see, other theories of rational requirement are possible
within the general framework of the Humean conception of rationality.
In contrast, according to what I will call the enkratic conception of rationality:
(i) An agent A has objective reason to choose an option O iff O is in some respect
good, valuable, or choiceworthy; (ii) A objectively ought to choose O iff O is better,
18
more valuable, or more choiceworthy than any alternative option; and (iii) A has
subjective reason to choose O iff she has positive degree of belief that O is in some
respect good, valuable, or choiceworthy.
As on the Humean conception, the enkratic conception of rationality admits
some variation when it comes to subjective oughts/rational requirements. I have
chosen to call this conception of rationality “enkratic” because, unlike the Humean
conception, it validates the enkratic principle, which asserts (in the idiolect I have
introduced) that if an agent fully believes that she objectively ought to choose an
option O, then she is rationally required to choose option O. (Analogously, on the
Humean conception, an agent who fully believes that choosing O will best satisfy her
desires is rationally required to choose O.) But this is only a limiting case of rational
requirement, that must be strengthened to yield a complete theory. Paralleling
the standard Humean view on which agents are rationally required to maximize
expected desire satisfaction, one version of the enkratic conception holds that agents
are rationally required to maximize expected choiceworthiness (Wedgwood, 2013).
But as with the Humean conception, other variations are possible, as we will see in
Chapters 2 and 7.10
The Humean conception of objective reasons/oughts emerges as a limiting case
of the enkratic conception, if one believes that the goodness, value, or choiceworthi-
ness of an option consists exclusively in its capacity for desire satisfaction. But one
10Just as the enkratic conception of rational requirements will, in one way or another, generalize
the enkratic principle, so it will take as the object of rational prohibition the classic philosophical
notion of akrasia. To act akratically is to act contrary to one’s beliefs about one’s objective reasons
for acting, meaning most centrally to do something that one is certain one objectively ought not
do, or fail to do something that one is certain one objectively ought to do.
19
can also hold the Humean conception while acknowledging the existence of desire-
independent goodness, value, or choiceworthiness, if one holds that the reason-giving
force for an agent of these properties of her options is mediated by her desires.11
Likewise, the Humean conception of subjective reasons/oughts emerges as a
limiting case of the enkratic conception, if one considers an agent who believes
that the goodness, value, or choiceworthiness of her options consists exclusively
of their capacity to satisfy her desires. For such an agent, her beliefs about the
choiceworthiness of her options will always coincide with her beliefs about their
capacity for desire satisfaction, and so the two conceptions of rationality will agree
about the rational requirements to which she is subject.
Despite this potential for convergence, the Humean and enkratic conceptions
of rationality are incompatible. Even if the Humean is in fact correct that all
objective reasons for action consist in the action’s contributing to the satisfaction
of an agent’s desires, an agent may certainly believe otherwise, and may believe
that she has most objective reason to choose option O while believing that some
incompatible option P would best satisfy her desires, if she herself is not a Humean.
11At the level of objective reasons/oughts, therefore, one who adopts the enkratic conception but
does not believe in any form of goodness, value, or choiceworthiness that does not directly depend
on an agent’s desires must find their view collapse either to the Humean conception or to nihilism
about objective reasons. For this reason, I will generally assume that the enkratic conception of
rationality is accompanied by a belief in desire-independent goodness, value, or choiceworthiness
and a belief that these properties constitute or furnish desire-independent reasons for action.
While this rules out the strong desire-dependence entailed by the Humean conception, it does
not rule out lesser forms of agent-relativity. I might have objective reason to choose O because O is
(or will result in a state of affairs that is) good for me (relative to me, from my perspective) rather
than good simpliciter, and O might be good for me in virtue of comporting with my values, plans,
projects, etc. Just how much agent-relativity the enkratic conception can allow before collapsing
into a version of the Humean conception is a question I will leave to the side. The important point
is that, on any interesting version of the enkratic conception, the mere fact that one of my desires
or other motivational states would be satisfied by choosing O is not ipso facto a reason for me to
choose O. For a defense of this claim and a representative presentation of the enkratic conception
of rationality, see for instance Quinn (1993).
20
A would-be religious ascetic might sincerely believe that he has most objective reason
to fast in the desert even while believing what would best satisfy his desires is a
life of drunken revelry. On the enkratic conception of rationality, this agent’s belief
that he has objective reason to fast in the desert gives him subjective reason to fast
in the desert, even if this belief is utterly unmotivating and even if there is no part
of his motivational set that favors fasting in the desert (that is to say, even if he is
in this instance fully akratic). Conversely, on the Humean conception, the agent’s
belief that drunken revelry would satisfy some of his desires provides some reason
for him to choose drunken revelry, even if he believes that there is nothing good
or choiceworthy about this option and therefore that he has no objective reason to
choose it.
Further, suppose the agent recognizes in himself some desires that would be
satisfied by fasting in the desert, but believes that his desires would overall by
better satisfied by drunken revelry. And conversely, he recognizes that he has some
objective reasons to choose drunken revelry, but believes that his objective reasons
to fast in the desert are much stronger. In this case, all else being equal, the enkratic
conception of rationality implies that this agent has more subjective reason to fast
in the desert while the Humean conception implies that he has more subjective
reason to choose drunken revelry. Thus, the two conceptions will disagree about
what course of action this agent is rationally required to adopt.
It should be obvious that, even at the coarse-grained level at which I have
so far described these two views, there is much that one might dispute internal to
either view. That is, I do not claim that everyone who takes a broadly Humean
21
view of rationality subscribes to every claim I have packed into the Humean con-
ception, and likewise for the enkratic conception. What I do claim, admittedly
without much argument, is that all philosophical theories of reasons and rationality
represent elaborations of or variations on these two basic conceptions. This claim
will be substantiated, if at all, mainly by the use to which I am able to put the
Humean/enkratic distinction in the following pages. But to lend it some initial
plausibility, let me make three brief observations.
First: It is often said that the conceptual core of rationality is coherence,
a notion which in some way generalizes the more limited notion of (mere logical)
consistency. Theories of rationality, then, will be defined by what combinations of
mental states apart from inconsistent belief they identify as incoherent. It seems to
me that there are just two alleged paradigms of incoherence in the practical domain:
1. I have desires that induce a certain ranking of states of affairs - I believe that
choosing option O will result in a more highly desired state of affairs than if
I choose incompatible option P - and yet I choose (or form an intention to
choose/perform) option P.
2. I believe that option O is better, more choiceworthy, favored by stronger
objective reasons than option P - and yet I choose (or form an intention to
choose/perform) option P.
These two forms of alleged incoherence correspond to the Humean and enkratic
conceptions of rationality respectively. The Humean identifies certain triples of
〈motivational state, belief about what will best satisfy my motivational state, choice/intention〉
22
as incoherent, while the enkratic theorist identifies certain pairs of 〈normative/evaluative
belief, choice/intention〉 as incoherent. It is of course possible to hold that both of
these represent forms of incoherence, but only on pain of concluding that someone
like our imagined ascetic, who believes that one course of action would best satisfy
his desires while believing that he has most objective reason to perform another
course of action, is doomed to incoherence whatever course of action he chooses.
The best way to argue for one conception of rationality over the other—and the
method I will adopt when I argue for the enkratic conception in the next chapter—
is to argue that in cases like these, one resolution of the practical dilemma looks
more coherent than the other.
Second: The enkratic/Humean distinction emerges naturally in another way,
from the idea of direction of fit. A Humean sees desires as exhibiting predominantly
world-to-mind direction of fit. While the normative standards that govern desires
will of course often call on me to revise my particular desires in light of new in-
formation about the world (e.g. the information that the substance in this glass is
gasoline rather than gin), my basic, non-instrumental desires are subject to revision
only in light of other desires, to meet standards of internal coherence and consis-
tency. Once that set of basic desires has been made maximally coherent, my job as a
practical agent is to shape the world so as to best fit those desires. According to the
enkratic conception of rationality, on the other hand, desires and other motivational
states are governed by norms that have mind-to-world direction of fit: that is, there
are normative standards governing my basic, non-instrumental desires beyond the
standards of internal coherence, standards that direct me to desire certain things
23
in the world in virtue of properties that have nothing to do with the relationship
between those things and my existing desires. “Value” or “choiceworthiness” are
merely convenient shorthands for these properties, whatever they turn out to be,
that direct me to form a basic, non-instrumental desire for something.
Third and finally: The distinction between Humean and enkratic views of ra-
tionality tracks and, I believe, refines the more standard division between internalist
and externalist theories of practical reasons, a distinction famously introduced by
Williams (1979). The reason internalist holds that all reasons for action are con-
tingent, perhaps in very complex ways, on the contents of an agent’s motivational
set, while the reason externalist holds that some reasons for action are not contin-
gent in this way. The Humean/enkratic distinction is more fundamental than the
internal/external distinction, it seems to me, because a reason for action can exist
in virtue of an agent’s desires without being contingent on the particular content
of those desires. Williams, for instance, defends reason internalism in a way that
largely takes for granted the Humean conception of reasons (or at least, that ignores
the possibility of the enkratic conception): He claims that for an external reason
statement to be true, it would have to be the case that any agent who deliberated
well enough would eventually come to be motivated to act in a certain way. And
this is indeed (at least to a first approximation) what it would take for an external
reason statement to be true, if the Humean conception of reasons is right. But
one who held, for instance, that any agent who deliberated well enough would de-
velop at least some motivation to avoid physical pain, despite counting as a reason
externalist under Williams’ schema, would be better classified with Humeans like
24
Williams than with those who hold that reasons for action need have nothing at all
to do with an agent’s actual or potential motivations.
To summarize: On the Humean conception, objective reasons for action exist
in virtue of an agent’s desires or other motivational states, and agents are rationally
required to (for instance) maximize expected desire satisfaction. On the enkratic
conception, objective reasons for action exist in virtue of desire-independent eval-
uative features of the world (goodness, value, choiceworthiness...), and agents are
rationally required (for instance) to maximize the expected satisfaction of their ob-
jective reasons, desire-based or otherwise, and thus in particular to choose options
that they are certain they objectively ought to choose, even when they are also
certain that the option in question would not best satisfy their desires.
The divide between the Humean and enkratic conceptions of rationality is a
theme to which we will repeatedly return. Whether these views indeed represent
rival conceptions/theories of the same basic concepts, or instead represent funda-
mentally different concepts (e.g., each laying claim to the expression “objective
reason” to pick out essentially unrelated concepts of “contributing to desire satis-
faction” and “value/choiceworthiness,” respectively) is an important question that
we will take up in Chapter 6. But for the moment I want my vocabulary, at least,
to remain neutral between the two views and so I will treat them as rival concep-
tions of one underlying concept. (The possibility of understanding the two views as
rival elaborations of an underlying idea of rationality-as-coherence provides at least
provisional justification for this choice.)
I have of course left a great deal unsaid about the nature of reasons and
25
oughts. Some of the details will be filled in by subsequent chapters, others will
not. I can only hope that the questions I leave unanswered—in particular, the
great many questions concerning kinds of reasons, defeators to reasons, and the
ways in which reasons can interact to generate on-balance or all-things-considered
reasons/oughts/requirements—will not affect the substance of my arguments in ways
I’ve failed to anticipate.
1.5 Moral Reasons
Fundamentally, the topic of the coming chapters will be the problem of nor-
mative uncertainty, which can be roughly characterized as uncertainty about one’s
objective reasons that is not a result of some underlying empirical uncertainty (un-
certainty about the state of concretia). However, I will confine myself almost ex-
clusively to questions about moral uncertainty: uncertainty about one’s objective
moral reasons that is not a result of etc etc. This is in part merely a matter of
vocabulary: “moral uncertainty” is a bit less cumbersome than “normative uncer-
tainty,” a consideration that bears some weight when the chosen expression must
occur dozens of times per chapter. It is also in part because the vast majority of
the literature on normative uncertainty deals specifically with moral uncertainty,
and because moral uncertainty provides more than enough difficult problems and
interesting examples, so that there is no need to venture outside the moral domain.
Additionally, however, focusing on moral uncertainty is a useful simplification
that allows us to avoid difficult questions about the relationship between moral and
26
non-moral reasons (though I am hopeful that the theoretical framework I develop can
be applied straightforwardly to normative uncertainties of a non-moral kind). For
myself, I have no taste for the moral/non-moral distinction: To put it as crudely
and polemically as possible, it seems to me that all objective reasons are moral
reasons. But this view depends on substantive normative ethical commitments that
it is well beyond the scope of this dissertation to defend. So I will make no use of
this claim, or indeed of any claims about the relationship between moral and non-
moral reasons (excepting some relatively modest claims required by the discussion
of supererogation in Chapter 4).
If one does think that all reasons are moral reasons, or that moral reasons
always override non-moral reasons, then a complete account of how agents ought
to act under moral uncertainty can be given without any discussion of non-moral
reasons (Lockhart, 2000, p. 16). To the extent that one does not share either of
these assumptions, theories of choice under moral uncertainty must generally be
qualified with “insofar as there are no relevant non-moral considerations.”
1.6 Simplifying Assumptions
One simplifying assumption that will generally be in force as we proceed, then,
is that non-moral reasons are inert in whatever cases we consider.
Another is that the practical agents we consider have no relevant empirical
uncertainties. For example, if we consider an agent who has some degree of belief in a
consequentialist moral theory, it will generally be assumed that she is not relevantly
27
uncertain about the consequences of her various options. This is in general an
innocuous simplification, though we will encounter a few cases in which potential
interactions between moral and empirical uncertainties require examination.
Most importantly, we will generally assume that the decision-relevant beliefs of
the agents we consider—both their empirical and normative beliefs—are conclusively
justified, i.e. epistemically/theoretically rational and more rational than any alterna-
tive beliefs the agent might have on the subject matter in question. This is in order to
sidestep the debate between subjectivism and evidentialism/prospectivism about ra-
tional requirements and the more recent debate between “narrow-scope” and “wide-
scope” principles of rational requirement.12 When an agent A has an unjustified
belief that she objectively ought to choose option O, the subjectivist/narrow-scoper
says that A is rationally required to choose O, while the evidentialist/wide-scoper
says that A is not rationally required to choose O but is instead rationally required
to revise her beliefs (and, presumably, to then choose whatever practical option is
favored by her new, justified doxastic state).13
12I take these debates to be closely related. If one accepts a narrow-scope principle of practical
rational requirement like “If A believes that she objectively ought to choose O, then she is rationally
required to choose O,” one is left with the subjectivist conclusion that what an agent is rationally
required/subjectively ought to do depends simply on her beliefs and not on the evidence for
or against those beliefs. If on the other hand one accepts only the wide-scope principle “It is
rationally required that, if A believes that she objectively ought to choose O, she chooses O,” then
the evidentialist conclusion seems natural: If A’s evidence does not support her belief that she
objectively ought to choose O but rather supports the belief that she objectively ought to choose
P, the best way of satisfying her rational requirements, it seems, is to abandon the belief that she
objectively ought to choose O, adopt the belief that she objectively ought to choose P, and then
choose (or form the intention to choose) P.
13I confess myself somewhat skeptical whether the debates between subjectivists vs. eviden-
tialists on the one hand, and narrow- vs. wide-scopers on the other hand, are substantive. The
evidentialist/wide-scoper wants principles that express the joint requirements of theoretical and
practical rationality, while the subjectivist/narrow-scoper prefers to keep principles of practical
rationality separate from principles of theoretical rationality. But both parties will agree that an
agent with unjustified beliefs should, rationally, revise her beliefs, and that if she is ideally theo-
retically and practically rational, she will revise her beliefs and then act rationally in light of her
28
When the antecedent of a principle of rational requirement concerns an agent’s
beliefs about her objective reasons (which, on the enkratic view for which I will argue
in the next chapter, is the form taken by all principles of practical rationality),
the wide-scope formulation of that principle will collapse into the narrow-scope
formulation if we assume that the agent has conclusive reason to hold the belief
mentioned in the antecedent. For instance, consider the wide-scope principle: “It is
rationally required that, if A believes she objectively ought to choose option O, then
she chooses option O.” This principle can be equivalently stated in disjunctive form:
“It is rationally required that either A does not believe that she objectively ought
to choose option O or A chooses option O.” But if this principle holds, and it is also
the case that A is rationally required to believe that she objectively ought to choose
option O (because she has conclusive evidence for that belief), then it follows that
she is rationally required to intend to choose option O, since forming this intention
is the only way for her to satisfy all of the rational requirements that apply to her.
Thus the wide-scope principle implies the following narrow-scope principle: “If A
has conclusive reason to believe that she objectively ought to choose option O, then
she is rationally required to choose option O.” Thus, by limiting our attention to
cases where the agent has conclusive justification for her relevant beliefs, we can
generally sidestep debates about the scope of rational requirements. I will therefore
typically adopt narrow-scope formulations for the sake of convenience, with the
understanding that these can be seen as limiting cases of wide-scope principles if
new, justified beliefs. So the views do not differ in terms of their ultimate practical implications.
29
that is to one’s taste.14
Each of these simplifying assumptions (no relevant non-moral reasons, cer-
tainty about all relevant empirical matters, and conclusively justified normative
and empirical beliefs) will be dropped at various points, as needed, but it should be
assumed that they are in force unless otherwise specified.
1.7 Substantive Assumptions
Along with these simplifications, I should acknowledge at the outset two im-
portant philosophical commitments of which I will make regular use, even though
they will go largely undefended. The first, which has already been introduced, is the
assumption that agents subjectively ought/are rationally required to do what they
have all-things-considered subjective reason to do. I take this to be a conceptual
truth, but it is not uncontroversial, and defending it lies beyond the scope of our
discussion.
The other important philosophical assumption, to which I have so far only
alluded, is a standard Bayesian picture of rational belief: namely that (i) belief
comes in cardinal degrees, (ii) it is a requirement of theoretical rationality that
an agent’s degrees of belief satisfy the probability calculus, and (iii) agents should
14To the extent that there is a substantive choice between narrow-scopism and wide-scopism
about rational requirement, I incline toward the narrow-scope view, for reasons much like those
given in Kolodny (2005): Principles of rational requirement ought to specify the ways in which
an agent can respond-by-reasoning to her situation, and while an agent can reason her way from
the belief that she objectively ought to choose option O to an intention to choose O (or, I think,
simply to choosing O), she cannot reason her way from her lack of an intention to choose O (or
her failure to choose O) to giving up the belief that she objectively ought to choose O. If the most
rational thing for her to do is to give up this belief, it is in light of a different rational requirement,
a requirement of theoretical rationality that concerns how she ought to respond to her evidence or
her beliefs about evidence.
30
update their beliefs by some form of Bayesian conditionalization (Talbott, 2015).
The first two components of the Bayesian picture will be crucial to the view I
eventually defend. The third is important only insofar as it lends support to the
requirement of regularity, i.e., that a rational agent should not assign minimal or
maximal degrees of belief (0 or 1) to any propositions apart from logical truths.15
I will refer to degrees of belief interchangeably as degrees of belief, credences,
or subjective probabilities. I assume that an agent’s degree of belief in a proposition
corresponds to something like the odds she is willing to accept in betting on the
truth of that proposition. In its naive form, this “betting interpretation” of sub-
jective probabilities is subject to several powerful objections (for an overview, see
§3.3 of Ha´jek (2012)), and the refinements needed to meet these objections require
strong idealizing assumptions. Indeed, it must be conceded that the Bayesian pic-
ture in general is at best an enormous idealization that ordinary human agents can
only hope to very roughly approximate in the best of circumstances. But I do think
that it’s the right ideal to approximate, and I know of no alternative that is more
psychologically realistic while also representing a plausible normative ideal. So in
attempting to develop an ideal theory of rational choice under moral uncertainty,
I will help myself to a manifestly ideal(ized) theory of the nature of that uncer-
tainty. (At the end of Chapter 7, I will briefly how non-ideal agents can attempt to
15Often, regularity is characterized more loosely, as assigning 0 and 1 only to necessary or a
priori truths. But insofar as we should be uncertain about the basic principles of morality, and
basic normative principles generally, even though the truths of these domains are presumably
necessary rather than contingent and arguably a priori rather than a posteriori, it is useful for our
purposes to make the more demanding regularity assumption.
For a typical defense of regularity as a requirement of rational belief, see for instance Ha´jek
(2003, pp. 31-2).
31
approximate the requirements of ideal rationality.)
1.8 Moral Theories
Throughout the following chapters I will speak mainly of uncertainty concern-
ing moral theories, rather than moral principles or particular moral dilemmas. By
“theory,” however, I mean simply a maximal consistent set of propositions in a given
domain—hence, a moral theory is a maximal consistent set of moral propositions.
Understood in this way, I think, the language of theories is innocuous, and does not
exclude or bias the playing field against “anti-theoretical” views like common-sense
morality or moral particularism.16 An agent who is uncertain, for instance, whether
it is permissible to tell a white lie to her friend in the particular choice situation
divides her belief between at least two moral theories, one of which includes the
proposition that it is permissible and another of which includes the proposition that
it is not. If she is uncertain about many moral questions, then she divides her beliefs
between many moral theories.17
16Thanks to Christopher Morris for pressing me on this point.
17Granted, there is a hefty dose of Bayesian idealization in the claim that, say, a particularist
distributes her moral belief over maximal consistent sets of moral propositions. The particularist
denies that morality is axiomatizable, i.e., that there is any finite (or at any rate, reasonably
small) set of true moral propositions of which all the rest are logical consequences. Therefore
she will presumably hold that the true moral theory (i.e., the maximal consistent set of true
moral propositions) is not the sort of thing that any human agent could consciously entertain, and
perhaps even that it is too complex to be represented in the unconscious resources of the human
mind (e.g., in the form of non-occurrent beliefs). Nevertheless, throughout the dissertation (except
briefly in Chapter 7) my focus will be on ideal rationality, setting aside the further question of how
bounded, finite human agents are to approximate that ideal. Therefore I will assume that agents
assign degrees of belief to all propositions that their conceptual resources enable them to entertain,
including moral propositions. As long as there are only finitely many atomic moral propositions
that such an agent can represent with her conceptual resources, she will assign degrees of belief to
maximal consistent conjunctions of literals (atomic moral propositions or their negations), which
amounts to an assignment of belief to moral theories. (It is worth reiterating here that I am helping
myself freely to talk of “truth” on the assumption that those who deny the truth-aptness of moral
32
In fact, if we want a general theory of rational choice under uncertainty that
treats normative, empirical, mathematical, and other forms of uncertainty in a con-
tinuous, unified manner, then that general theory will make reference not to moral
theories but simply to theories of the world, i.e., maximal consistent sets of propo-
sitions without any domain restriction. Since we will generally be operating under
the simplifying assumption that the agents we consider are only uncertain about
moral matters, and experience no uncertainty (or at least, no relevant uncertainty)
about any non-moral matters, I will mostly set this point aside and speak simply
of moral theories. But in Chapter 6 and Appendix B we will have some reason to
think of agents as distributing belief over complete theories of the world that include
propositions of every kind, moral and otherwise.
Finally: In Chapter 2, I will defend the enkratic conception of rationality
as more plausible than the Humean conception, and thereafter it will sometimes
be convenient to speak of “value” or “choiceworthiness” in place of “objective rea-
sons.” In the idiolect of subsequent chapters, unless otherwise specified, the value or
choiceworthiness of an option is just the degree to which it is supported by objective
reasons, relative to other options.18 I will also sometimes speak of the “subjective
choiceworthiness” of an option as a convenient locution for the degree to which that
propositions will nonetheless supply some truth-surrogate that permits us to say things analogous
to what we can say in the language of truth.)
18There is a narrow sense of terms like “value” or “choiceworthiness” on which these are, as
a conceptual matter, desire-independent features of the world or properties of practical options,
and it was to this sense that I appealed above in attempting to distinguish the enkratic from the
Humean conception of rationality. But a reader who remains unconvinced by the anti-Humean
arguments of the next chapter should be able to understand these terms, thereafter, as simply
synonyms for “objective reason,” leaving open the possibility that all objective reasons involve
and depend on the satisfaction of desires.
33
option is supported by subjective reasons. The value scale of a moral (or other
normative) theory is the structure provided by that theory for comparing degrees
of objective reason: a set of possible values that the theory assigns to practical
options, together with a more or less enriched ordering on that set. For instance,
classical utilitarianism has a cardinal value scale, with practical options assigned
real-numbered values over which are defined not only the relation ≥ but also the
operations + and −.19 But one way of conceiving certain deontological theories,
for instance, is as ordinally structured, ranking options as better or worse while
denying that there are meaningful quantitative answers to the question “How much
better/worse?”20
A theory’s value assignment, then, is its mapping from practical options to
points on its value scale. Where convenient, I will denote the value assigned by a
theory T to a practical option O by the expression V T (O). It will also occasionally
be convenient to speak of the value assigned to payoffs (increments of some value-
bearing phenomenon, like a gain or loss of one hedon) using the same formalism
19Whether multiplication and division are defined depends on whether the utilitarian value scale
is an interval scale (giving cardinal degrees of difference between values but having no privileged or
non-arbitrary zero point and hence not giving ratios between absolute degrees of choiceworthiness)
or a ratio scale (giving both cardinal degrees of difference between values and a non-arbitrary zero
point, and hence permitting talk of ratios between values). For my part it seems most natural to
conceive of consequentialist theories as having ratio structure, since if those theories are aggregative
(determining the value of the states of affairs that might result from an agent’s actions by summing
the values instantiated at multiple, independent value locations, e.g. persons or spacetime regions),
then there is a non-arbitrary zero point on the value scale, namely, an empty universe that does
not contain any value locations at all. For a risk-neutral expected value maximizer, however, only
the interval structure of a normative theory is relevant, so it will generally make no difference
whether we conceive of these theories as interval or ratio.
20For detailed discussion of the distinction between ordinal, interval, and ratio structures in
the context of moral theories and its relevance to rational choice under moral uncertainty, see
MacAskill (2014), pp. 14-15, 52ff. We will return to these questions in greater detail in Chapter 6,
where I argue that many other structures are possible as well—indeed, all possible enrichments of
the very simplest structure a normative theory can possess, namely, a binary structure that simply
classifies options as permissible or impermissible.
34
(e.g., V T (one hedon)). Strictly speaking, such expressions should be understood to
denote the difference in value between practical options that are equivalent in all
morally relevant respects but for the payoff in question (e.g., a pair of options that
produce all the same consequences, except that one produces an additional hedon
that the other does not).21
In addition to an objective value assignment, a normative theory may also in-
clude one or more additional value assignments that are “subjectivized” to various
degrees, i.e., relativized to various aspects of an agent’s credal state. For instance,
hedonistic utilitarianism gives an objective value assignment according to which
the value of any practical option is equal to the net hedonic value of its actual
(or counterfactual) consequences, but also a partially subjective value assignment
according to which the value of any practical option is equal to the expected net
hedonic value of its consequences.22 As we will see, a large part of the problem of
decision-making under moral uncertainty has to do with the aggregation of value
assignments, i.e., the method by which an agent goes from a probability distribu-
tion over value assignments to a new value assignment that represents the overall
degrees of subjective reason for choosing particular practical options in light of that
probability distribution.
21This latter locution is intelligible, and will be used, only in the context of theories that recognize
value differences between options more fine-grained than “better” or “worse”—e.g., in the context
of cardinally structured theories but not in the context of merely-ordinal theories.
22While our primary interest will be in the fully subjective value assignment that compares an
agent’s subjective reasons for choosing her various options in light of the entirety of her credal
state, I will argue in Chapter 6 for a multi-stage aggregation procedure on which we have reasons
to be interested in partially subjective value assignments, e.g., the value assignment to an agent’s
options conditional on a particular moral theory in which she has positive credence, which takes
account of her empirical uncertainties but not her moral uncertainties. We will also consider value
assignments that are conditioned on disjunctions of moral theories, which play an important role
in the multi-stage aggregation process.
35
Again it is worth emphasizing that talk of value scales and value assignments
does not presuppose any tendentiously theoretical conception of ethics or rule out
anti-theoretical views like particularism. All that is presupposed is that a maxi-
mal consistent set of moral propositions should, in one way or another, assess and
compare an agent’s (moral) reasons for or against particular practical options.
1.9 Summary of the Dissertation
Having at least roughly described the conceptual framework and philosophical
background assumptions on which I will draw in the coming chapters, we are now
ready to engage with the problem of moral uncertainty proper.
I begin in the next chapter by making the positive case for the relevance of
moral uncertainty to rational decision-making. I will describe and defend two prin-
ciples of rationality, one a strengthening of the other, but either of which is enough
to establish the need for a theory of rational choice that is sensitive to an agent’s
moral beliefs (on the weaker principle) and/or degrees of belief (on the stronger
principle). Chapter 3 then attempts to answer some challenges that have recently
been forwarded against this conclusion by philosophers like Brian Weatherson and
Elizabeth Harman who hold that what an agent ought to do, in every important
sense, depends on the moral facts and the agent’s empirical beliefs but not on her
moral beliefs or degrees of belief.
Having addressed these challenges, we can take up the question of what princi-
ples of rational choice should guide an agent who is uncertain about the requirements
36
of morality. We will begin in Chapter 4 by considering the capabilities and limi-
tations of approaches based on dominance reasoning, which have recently received
significant attention from philosophers. I will conclude that, contra worries in this
recent literature, interesting normative consequences concerning moral nihilism, su-
pererogationism, and certain controversial moral topics like meat consumption can
be derived from relatively weak and uncontroversial dominance principles, but also
that these dominance principles only have interesting scope for application if they
can be construed as limiting cases of some more general (e.g. expectational) theory
of choice under moral uncertainty.
Chapter 5 then embarks on the task of describing this more general theory.
The chief obstacle to this task is the problem of finding a non-arbitrary basis for
intertheoretic comparisons of moral value. I argue, however, that rather than looking
for a single, maximally general principle to govern such comparisons (as Lockhart
and others have attempted), we should take a bottom-up approach that starts from
what I will call “comparability classes” of moral theories, within which there are clear
and rationally perspicuous bases for intertheoretic comparison. Any general theory
of moral uncertainty, I suggest, must be compatible with the clear and compelling
results we find in these more limited cases of uncertainty between closely related
and structurally similar accounts of morality.
Chapter 6 will then attempt to flesh out the general approached introduced
in Chapter 5 into the outline of a general theory of rational choice under moral
uncertainty. I will distinguish three obstacles to aggregation of rival moral theories
under uncertainty, enumerate possible responses to each of these obstacles, and
37
argue for my own preferred response to each. From this will emerge a novel picture
of decision-making under moral uncertainty, though with many important details
still left to be filled in.
Finally, in Chapter 7, I take up one of the most difficult problems for any
approach to moral uncertainty. This is the so-called “regress problem,” which stems
from the observation that, just as an agent can be uncertain about first-order norma-
tive principles, she can likewise be uncertain about second-order principles of choice
under uncertainty like those proposed in Chapters 5 and 6. Therefore, it seems,
the principles of rational choice that demand sensitivity to moral uncertainty in the
first place force a rational agent to deliberate not just on the basis of second-order
principles that accommodate her first-order uncertainty, but also on the basis of
third-order principles that accommodate her second-order uncertainty, fourth-order
principles that accommodate her third-order uncertainty, and so on ad infinitum.
In response to this challenge, I argue that at least some normative principle(s) must
possess “external,” belief-independent normative force, but that, contra first-order
moral externalists like Brian Weatherson, this external status is best attributed to
a fundamental principle of rationality : the enkratic principle, correctly formulated.
I then distinguish two aspects of the regress problem, one having to do with ideal
rationality and the other with non-ideal rationality, and argue that this “enkratic
externalist” position enables solutions to both problems.
38
Chapter 2: Rationality and Moral Belief
In this chapter and the next, I aim to establish the need for a theory of rational
decision-making under moral uncertainty. To do this, I will defend a weaker and
then a stronger thesis: first, that an agent’s full moral beliefs give her (subjective)
reasons for action and second, that her partial moral beliefs (i.e., positive degrees of
moral belief) give her (subjective) reasons for action. If the weak thesis is correct,
then a complete theory of rational choice must take account of an agent’s moral
beliefs. If the strong thesis is correct, then a complete theory of rational choice
must take account of her moral uncertainties. This establishes the project to be
pursued in the remainder of the dissertation as one worth pursuing.
In this chapter I spell out the two theses more precisely, then describe what I
take to be prima facie compelling arguments for each thesis. In the next chapter,
I will consider arguments for the practical irrelevance of moral belief and/or moral
uncertainty raised by Brian Weatherson, Nomy Arpaly and Timothy Schroeder, and
Elizabeth Harman, and argue that none of these challenges is as compelling as the
prima facie case for the two theses.
Throughout, my goal will be to establish that an agent’s moral beliefs and
degrees of belief are relevant to what she subjectively ought to do or is rationally
39
required to do, in the sense of these terms introduced in the previous chapter. By
contrast, I do not aim to show that an agent’s moral beliefs and uncertainties are
relevant to what it is right for her to do, or what she would be praiseworthy or
blameworthy for doing, or to what it would be virtuous for her to do, etc. One or
more of these notions may turn out to have some necessary connection to subjective
oughts/rational requirements, such that what I say about the latter concepts will
have implications concerning the conditions for rightness, blameworthiness, virtue,
or the like. But I will make no claim to this effect.
By making this stipulation, I run some risk of talking past my philosophical
adversaries, who as we will see often appear to be concerned in the first instance
with characterological notions or with praise and blame rather than with subjective
rationality. But my goal is simply to vindicate the project of looking for princi-
ples of decision-making that account for an agent’s moral beliefs and uncertainties,
and insofar as moral uncertainty is relevant to subjective oughts/rational require-
ments, this project is vindicated. Conversely, insofar as the philosophers to whom
I respond are rejecting this project (and not merely declaring it irrelevant to some
other project, like theories of blame and responsibility), we are in substantive dis-
agreement. As will become clear in the next chapter, Weatherson and Harman at
least do seem to reject the project of looking for moral-belief-relative norms, and
hence reject the claim that moral beliefs and degrees of belief are relevant to rational
requirement, so we can focus on the concept of rational requirement without missing
the point of their arguments.
40
2.1 Moral Belief and Reasons for Action: Two Theses
These, then, are the two principles that I will defend in this chapter and the
next:
Weak Subjective Rationalism (WSR) If an agent A fully believes that she is
objectively morally required to choose option O, and is conclusively justified
in so believing, then A thereby has some subjective reason to choose O.
Strong Subjective Rationalism (SSR) If an agent A partially believes that she
is objectively morally required to choose option O (i.e., has positive degree
of belief that she is so required), and is conclusively justified in so believing,
then A thereby has some subjective reason to choose O, with the strength of
this reason covarying monotonically with A’s degree of belief.
Put intuitively, WSR asserts the rational relevance of moral belief and SSR
further asserts the rational relevance of moral uncertainty. Assuming that full belief
can be understood as any degree of belief that equals or exceeds some “Lockean
threshold” L, WSR asserts that an agent whose credence that she is objectively
required to choose option O is greater than or equal to L has thereby at least some
subjective reason to choose O, meaning among other things that ceteris paribus (ab-
sent any other subjective reasons for or against choosing O or any of its alternatives)
she is rationally required to choose O.1 SSR asserts that an agent who assigns any
1This Lockean threshold view, it seems to me, is the best way of accommodating the notion
of full belief within a Bayesian framework. Nevertheless, is carries certain counterintuitive impli-
cations, most importantly that rational belief does not agglomerate: a fully rational agent may
41
positive probability to the proposition that she is objectively required to choose O
has some subjective reason to choose O, which increases in strength as that probabil-
ity increases, meaning among other things that an agent who is uncertain whether
she is objectively morally required to choose option O or objectively morally re-
quired to choose incompatible option P thereby has both some subjective reason to
choose O and some subjective reason to choose P, reasons which must be accounted
for in determining what she subjectively ought to do, all things considered.
Note that neither WSR nor SSR denies that there may be other sources of
subjective reasons entirely unrelated to an agent’s beliefs about her objective moral
obligations and even entirely unrelated to an agent’s beliefs generally. Thus, these
principles only assert the rational relevance of full/partial moral belief, not that
such belief is the only determinant of rational requirement.
Given this limitation, the weaker principle WSR may appear so trivial that no
one could deny it. But in fact, Harman (2015) does pretty clearly deny WSR, and it’s
possible that Weatherson does as well.2 More importantly, though, even WSR “gets
us in the game” of looking for norms of rational choice that are sensitive to an agent’s
moral beliefs, contra the spirit of Weatherson’s and Harman’s critiques, both of
which criticize the project of seeking such norms (rather than, for instance, claiming
believe P, believe Q, but fail to believe P & Q. Rejecting agglomeration is, to my mind, the most
attractive solution to the lottery paradox: Assuming that one wishes to allow for full belief in
anything other than the limiting case of absolute certainty, then given a fair lottery with a billion
tickets, a rational agent ought to believe that Ticket 1 will not win, ought to believe that Ticket
2 will not win, ... but ought not believe that none of the tickets will win. Defending this view is
beyond the scope of the present discussion, but nothing I have to say in defense of WSR will, as
far as I can tell, turn on the exact nature of full belief.
2See the quoted statement of Harman’s view that opens §3.5 in the next chapter. Weatherson’s
view is complex and difficult to situate relative to my theses, but I make the attempt in footnote
1 of Chapter 3.
42
that these norms should be understood only as norms of rational requirement and
not of some other normative concept like blameworthiness).3
Strictly speaking, WSR must be conjoined with some minor auxiliary premises
to establish the relevance of moral beliefs to rational requirements. If agents are
rationally required to do what they have most subjective reason to do, then given
WSR the reason to choose O generated by an agent’s full belief that she is objectively
morally required to choose O will sometimes produce a rational requirement to
choose O—when no other reasons are present, or when the balance of other reasons
is against choosing O, but close enough that the reason postulated by WSR tips
the scales in its favor. This could fail to be the case only if reasons of some other
kind are always present and always decisive: for instance, one might in principle
allow that an agent’s moral beliefs give her subjective reasons, but hold that the
combination of her empirical beliefs with the true moral theory always gives her
stronger subjective reasons, such that the reasons given by her moral beliefs can
never make the difference in deciding what she is rationally required to do. While
this view is consistent, I can see no motivation for defending it and know of no
3In a more recent article, Brian Hedden (2016) gives the clearest articulation of the anti-WSR
position to date. On the initial statement of his view, “There is no normatively interesting sense
of ought in which what you ought to do depends on your uncertainty about normative facts” (p.
3). It is clear from the arguments Hedden marshals, however, that he means to deny the relevance
(to any “normatively interesting sense of ought”) not just of moral uncertainty but of moral belief
generally. Indeed, later in the paper he confirms that he is defending this stronger position, in
contrasting his view with Weatherson’s: “Weatherson is arguing that moral beleifs do not affect
what you morally ought to do, but denying that moral beliefs affect what you morally ought to do
is compatible with thinking that there is some other sense of ought (e.g., not the primary moral
ought, but rather a super-subjective ought), in which what you ought to do depends on your moral
beliefs. It is this latter claim that I am denying” (p. 23n). This distinction notwithstanding, it
seems clear to me that Weatherson as well as Hedden and Harman would be unwilling to concede
that the all-things-considered rational ought that I have been calling the “subjective ought” is
sensitive to an agent’s moral beliefs or uncertainties.
43
one who has tried. So while my defense of WSR and SSR will allow for subjective
reasons of other kinds that are insensitive to an agent’s moral beliefs, I will take for
granted that these other reasons are not always rationally decisive, and hence that
if WSR/SSR are true, then an agent’s moral beliefs/degrees of belief will sometimes
generate rational requirements.
As WSR implies that moral belief is relevant to rational requirements, so SSR
implies that moral uncertainty is relevant to rational requirements. For instance, if
SSR is true, then an agent’s worry that it might be morally impermissible to eat
meat will provide her with some subjective reason to avoid eating meat, and where
no other reasons are present or the balance of those other reasons does not strongly
favor eating meat, this reason may prove decisive, generating a rational requirement
against eating meat.
WSR and SSR are, by design, somewhat independent of the gradations of
viewpoint most commonly drawn in the literature on moral uncertainty. Among
those who accept that an agent’s moral beliefs are relevant to what she subjectively
ought to do, the most modest position in the standard taxonomy is the view Lock-
hart (2000) calls “My Favorite Theory” (MFT): An agent subjectively ought to do
whatever is prescribed by the moral theory to which she assigns the highest degree
of belief.4 The next most modest position is a view sometimes called “My Favorite
Option” (MFO): An agent subjectively ought to take whatever action she judges
most likely to be objectively permissible (where various moral theories that judge
4This view is defended under another name by Gracely (1996), and as My Favorite Theory by
Gustafsson and Torpman (2014).
44
the action permissible might contribute to her overall degree of belief that it is so).5
Finally, there are “hedging” views, according to which what an agent subjectively
ought to do is sensitive not only to her degrees of belief about objective rightness
and wrongness, but to the degrees of potential rightness and wrongness associated
with her various options. Unlike My Favorite Theory and My Favorite Option, for
instance, hedging views imply that if an agent believes to degree .8 that she is under
a weak obligation to keep a promise to a friend by donating some money to a high
school reunion fund, but believes to degree .2 that she is under a much stronger
obligation to donate that money to GiveDirectly instead, the comparative strength
of her potential obligation to donate to GiveDirectly can make it the rationally
required option.
Both WSR and SSR are compatible with any of these three views (though
SSR combines very uncomfortably with My Favorite Theory). Importantly, even
the stronger SSR does not imply that one must adopt a hedging view, for it does
not assert that the strength of the subjective reasons given by partial moral beliefs is
in any way related to the degrees of rightness or wrongness that those beliefs assign
to an agent’s options. While the argument offered for SSR in the next section can
be easily seen as an argument for hedging, objections to hedging views in particular
(especially, arguments stemming from the problem of intertheoretic value compar-
isons) might lead one instead to a view like MFO, even if one accepts the argument
for SSR. Though in chapters 5-6 I will argue that the problem of intertheoretic
5MFO is considered and rejected under this name by Gustafsson and Torpman (2014) and
MacAskill (2014). Lockhart considers and rejects the same view, which he calls “PR2” (Lockhart,
2000, p. 26ff).
45
value comparisons can be at least sometimes overcome and therefore that hedging
is at least sometimes the rational response to moral uncertainty, my concern for the
moment is not to defend anything more than SSR.
The opponents of the moral uncertainty project, in denying both WSR and
SSR, and adopt a position even more conservative than MFT. If what an agent
ought to do is determined simply by the true moral theory plus her empirical be-
liefs/evidence, then the belief that one is or might be objectively required to choose
O does not in itself generate any reason for action. For instance, a Kantian who
believes (even with certainty) that she is morally required to tell a harsh truth
rather than a harmless white lie nevertheless has no subjective reason at all to tell
the truth or avoid the lie, if in fact utilitarianism is the right moral theory. For
lack of established terminology, I will refer to this position as irrelevantism: the
irrelevantist holds that moral beliefs and degrees of belief are irrelevant to what any
agent subjectively ought to do.6
In examining the arguments on offer for the irrelevantist position, I will do my
best to distinguish between those that challenge only SSR and those that go further
and challenge WSR. But first, I will describe what I take to be the compelling
prima facie basis for accepting the two theses and hence the rational relevance of
moral uncertainty. The case for WSR and the relevance of moral belief turns on its
connection to the enkratic principle and the enkratic conception of rationality. The
further case for SSR and the relevance of moral uncertainty, as we will see in §2.3,
6Harman, in advocating this view, dubs it “actualism,” but I avoid this terminology since it
risks confusion with the somewhat-related metaethical debate between actualism and possibilism.
46
turns on a close analogy between moral and empirical uncertainty.
2.2 The Prima Facie Case for WSR: The Enkratic Principle
The enkratic principle, in its most plain-vanilla formulation, asserts the fol-
lowing:
EP1 If an agent A believes that she ought to (perform) O, then A is rationally
required to intend to (perform) O.
In light of claims made and apparatus introduced in the preceding chapter, we
can immediately substitute a slightly revised formulation:
EP2 If an agent A believes that she objectively ought to choose option O, and is
conclusively justified in so believing, then A is rationally required to choose
option O.
EP2 improves on EP1 in three ways: first, by clarifying that it is beliefs about ob-
jective rather than subjective reasons/oughts that generate rational requirements;7
7The relationship between the subjective oughts/rational requirements to which an agent is
subject and the agent’s beliefs about the subjective oughts/rational requirements to which she
is subject poses difficult questions, which we will say something about in Chapter 7. For the
moment, however, it is worth noting that endorsing principles that derive rational requirements
from an agent’s justified beliefs about objective reasons does not involve any commitment with
respect to subjectivism vs. evidentialism about subjective oughts. It does rule out a certain sort
of subjectivism on which an agent always ought to do whatever she believes that she subjectively
ought to do (at least, it rules out this view on the assumption that an agent could, e.g. via a
failure of rationality, believe that she objectively ought to choose option O and yet believe that
she subjectively ought to choose some incompatible option P rather than O). But it says nothing
about the rational requirements to which an agent is subject when her beliefs about what she
objectively ought to do are unjustified, i.e., it leaves open both the subjectivist possibility that
the agent subjectively ought to act in accordance with her unjustified belief and the evidentialist
possibility that she subjectively ought to act in accordance with the evidence that supports a
different belief.
47
second, by limiting our attention to justified beliefs and therefore setting aside the
subjectivist vs. evidentialist and narrow-scope vs. wide-scope debates;8 and third,
by framing the consequent rational requirement in terms of choice rather than per-
formance or intention.
At least one more revision is required, however, to produce a fully plausible
formulation of the enkratic principle. As Wedgwood (2013) points out, a principle
like EP2 seems to wrongly imply that if I believe (with, say, 99.9% confidence)
that choosing option O will save one life and have no other morally significant
consequences, but recognize a 0.1% risk that choosing option O will result in the
destruction of a medium-sized city and the deaths of a million innocent people, and
am certain that the only alternative option P would have no morally significant
consequences at all, I am rationally required to choose option O (since I believe, just
with something less than certainty, that O will in fact have better consequences than
P and is therefore the option I have most objective reason to choose). If the notion
of “belief” that figures in EP2 is taken to be full belief in the Lockean threshold
sense, as in WSR, then this unwanted consequence follows straightforwardly.9
8A stronger condition that might set these debates even more conclusively to the side would
require that the beliefs in question be conclusively justified, i.e., that the agent be epistemically
required rather than merely permitted to hold those beliefs in light of her evidence. This distinction
will be important, for instance, if one draws a more basic distinction between justifying reasons and
requiring reasons in the epistemic domain (Gert, 2003). The issues raised by this distinction would
be a distraction in the context of the present discussion, but in any case the various principles
defended in this chapter suffice to establish the rational relevance of moral belief/uncertainty even
with the more rigorous condition of conclusively justified beliefs, so long as one accepts that it is
possible for an agent to have conclusive justification for believing a false moral theory.
9Accepting this as a counterexample to standard formulations of EP does not require one to
accept (begging the question of the present chapter) that purely moral uncertainty can affect what
an agent is rationally required to do, but only the more or less uncontroversial point that morally
relevant empirical uncertainty can affect what an agent is rationally required to do. For anyone
who doubts this latter claim, some defense of it will be given in the next section.
48
Wedgwood proposes replacing the standard formulation of EP with a principle
asserting that agents are rationally required to form sets of intentions that maximize
expected choiceworthiness, and he takes this revised principle to be “the fundamental
principle of rationality” (Wedgwood, 2013, p. 505). (MacAskill (2014) also defends
a somewhat restricted version of this view, which he dubs MEC.) Adapted to our
idiolect, Wedgwood’s version of EP can be stated as follows:
EP3/MEC Assuming that all of agent A’s relevant beliefs are conclusively justified,
then in any choice situation, A is rationally required to choose an option with
maximal expected choiceworthiness (i.e., such that the probability-weighted
sum of possible objective reasons to choose that option is equal to or greater
than the equivalent sum for any alternative option).
Such a principle would certainly be suitable to my purposes, lending clear
support not only to WSR but also to SSR and indeed to moral hedging. But
we need not go nearly so far ourselves to accommodate Wedgwood’s objection to
standard formulations of EP.
We might also, for instance, simply weaken EP2 by confining its application
to cases of moral certainty rather than mere belief:
EP4 If an agent A believes with certainty (i.e., probability 1) that she objectively
ought to choose option O, and is conclusively justified in so believing, then A
is rationally required to choose option O.
Of course, if we accept the assumption of regularity then this restricted prin-
ciple will strike us as entirely uninteresting, since an agent is never justified in
49
believing with probability 1 that she ought to choose option O. As a compromise
between the ambitious EP3 and the overly-modest EP4, then, we might adopt the
following ceteris paribus formulation of EP:
EP5 If an agent A fully believes (in the Lockean threshold sense) that she objec-
tively ought to choose option O, and is conclusively justified in so believing,
then A is rationally required to choose option O, unless her beliefs about her
objective reasons conditional on its not being the case that she objectively
ought to choose option O generate subjective reasons against choosing O that
are strong enough to defeat her subjective reasons for choosing O.
EP4 and EP5 can be seen as limiting cases of EP3, or of some other similarly strong
formulation of EP that posits, for instance, a more complex attitude toward risk in
place of Wedgwood’s risk-neutral, expectation-maximizing view.
Anyone who accepts any of these formulations of EP, though, should also
accept WSR. Let’s focus on EP5 as the most interesting case. Deriving WSR from
EP5 requires just two auxiliary premises: first, that an agent who believes that she is
objectively morally required to choose option O believes that she objectively ought
to choose O, and second, that the ceteris paribus rational requirement to choose
O described in the consequent of EP5 stems from subjective reasons to choose O
which are overridden, rather than undercut or otherwise caused to disappear, in
cases where the risks associated with option O prevent it from being rationally
required.
I see no reason to doubt the latter assumption. With respect to the former,
50
note that this does not require the controversial claim that moral reasons always
override non-moral reasons. One might hold, with Portmore (2008), that non-moral
reasons can sometimes prevent moral reasons from generating moral requirements
(as when the prudential costs of some altruistic act make it morally permissible not
to perform that act, even though the moral goods that would result from the act
are such that it would be morally obligatory if there were no prudential cost). The
claim is only that moral requirements are a species of requirement simpliciter, i.e.
that to be objectively morally required to choose option O is just to be objectively
required to choose option O, for moral reasons.
This should at least make plausible the general claim that, if we accept any
plausible formulation of the enkratic principle, we should also accept WSR. Indeed,
even without examining particular formulations of EP, it should be clear that WSR
is hard to resist if one accepts the general spirit of the enkratic conception of ratio-
nality. On the enkratic conception, the paradigmatic form of incoherence and hence
irrationality in the practical domain is a mismatch between one’s beliefs about what
one objectively ought to do and one’s intentions or choices. Even allowing for Wedg-
wood’s point that full belief in the Lockean threshold sense does not always generate
rational requirements, an agent who believes in this sense that she ought to choose
option O and yet fails to experience even pro tanto motivation to choose O clearly
manifests incoherence in this sense (i.e., manifests akrasia). To say that a fully ra-
tional agent in this situation would feel at least some motivation to choose O is just
to say that an agent in this situation has some subjective reason to choose option
O, which is what WSR asserts.
51
But whether we should accept any formulation of the enkratic principle de-
pends on whether we should accept the broader enkratic conception of rationality.
I attempted to show in the previous chapter that there is a basic divide between
theories of practical rationality that elaborate the enkratic, normative-belief-based
conception of coherence in the practical domain and those that elaborate the ri-
val Humean, desire-based conception. If this is so, then why should we accept the
enkratic conception over its Humean competitor?
This is of course too large a debate to settle in a few pages. So I will only
gesture at what I take to be the most compelling reason to prefer the enkratic
conception. (I will then close the section by arguing that even one who prefers the
Humean conception of rationality should accept WSR or something like it.)
The well-trodden path of argument against Humean theories of practical ratio-
nality involves agents with bizarre motivational states, like Parfit’s “Future Tuesday
Indifference” (Parfit, 1984, pp. 123ff) or Quinn’s imagined basic disposition toward
turning on radios (Quinn, 1993, pp. 236ff). But among other standard rejoinders,
the Humean may simply maintain that in such alien cases it is difficult for us to
appreciate the reason-giving force that such an agent’s desires would have for her.
Let’s instead consider the case I described in the previous chapter as a test case for
the Humean and enkratic conceptions: the would-be religious ascetic—let’s now call
him Simon—who believes that one course of action (fasting and repentance in the
desert) would be most objectively good or choiceworthy while believing that another
(endless debauched partying) would best satisfy his desires. The crucial question is,
of these two choices, which would manifest a lack of internal coherence on Simon’s
52
part, given the motivational and belief states we have ascribed to him.
It seems intuitively clear to me that the choice to fast in the desert coheres
better with Simon’s overall internal state that the choice to revel in the city. One
obvious way in which this is so is that Simon will feel that he can explain or justify
to himself the choice to fast, despite the sacrifices it entails, in a way that he could
not explain or justify to himself the choice to revel. Whichever choice he makes,
he may feel some sense of ongoing internal struggle after the choice has been made,
and even a species of regret. But if he chooses to revel, he will have reason to
believe that he made an all-things-considered mistake, that he was weak-willed, that
he cannot justify or defend what he is doing, whereas if he chooses to fast he will
have no reason to think any of these things (assuming his beliefs about the objective
choiceworthiness of fasting remain unchanged).
Of course, there is still the following puzzle: If Simon does in fact successfully
choose to fast in the desert, doesn’t this very choice show that fasting-in-the-desert
was in fact what he had most desire to do, and therefore that, despite appearance, he
was acting rationally on the Human conception as well as the enkratic conception
of rationality? Three remarks on this puzzle: (i) A point of divergence among
Humean theories of rationality is whether rationality concerns only the satisfaction
of present desires/motivational states or of anticipated future desires/motivational
states as well. If one adopts the latter view, then there is no puzzle, since an agent
may easily act against the sum of his present and anticipated future desires. (ii) In
the last chapter, I assumed that the Humean accepts a picture of subjective reasons
and hence of rationality on which these things are determined not by what will
53
best satisfy an agent’s desires but by his beliefs about what will best satisfy those
desires. In seems inevitable that the Humean must say something like this, since
in the absence of perfect knowledge about what will satisfy our desires we can only
be guided by our beliefs. If this is right, and if we keep the objective/subjective
distinction clearly in view, then there is no puzzle: Objectively, of course, Simon
might choose to fast even though reveling will in fact best satisfy his desires—
suppose he simply underestimates the degree to which his desire to avoid physical
suffering will be dissatisfied by choosing to fast. And subjectively it is equally clear
that Simon might choose an option that he believes will not best satisfy his desires—
he might, for instance, regard desire as inherently sinful and hold that in choosing to
fast in the desert he is actuated by some entirely different source of motivation, like
divine grace (or a Kantian sense of duty). (iii) The Humean may still, if they like,
insist that their view does not entail drunken revelry to be the more rational choice
for Simon. They might say that, in any event, he must believe when he chooses
to fast that this is the choice most consistent with his overall motivational state at
the moment, even if he believes that motivational state to include things other than
desires, and that this is enough to make his choice rationally permissible. Or they
might give their conception of subjective rationality a de re twist and say that, even
if Simon does not believe at the moment of choice that fasting in the desert will best
satisfy his desires, nevertheless the fact that he chooses it means that he believes
it will best satisfy the objectives that he in fact most strongly desires, even though
he would deny those objectives to be objects of desire. But either of these views
has the obvious consequence of making the Humean theory trivial, implying that
54
any or virtually any option that an agent in fact chooses must ipso facto have been
rationally permissible for that agent.
To sum up, then: If the Humean view is to avoid triviality, it must claim
that Simon is rationally required to revel in the city rather than fast in the desert.
Therefore if one judges, as I do, that fasting is in fact the more rational option
for Simon given his various beliefs and motivations, one must reject the Humean
conception of rationality in favor of the enkratic conception (and therefore accept
some version of the enkratic principle, which will entail WSR).10
But in any case, it seems to me that even Humeans about rationality should
accept WSR or something very much like it, and hence should accept the relevance
of moral belief to rational requirement. The tightness of the connection between
10Niko Kolodny (2005) offers a different line of argument for what I have been calling the enkratic
conception of rationality. He argues at length that we do not have (objective) reasons to comply
with rational requirements as such—that is, when in light of my beliefs and/or evidence I am
rationally required to choose option O, although in fact there is no objective reason to choose
O (e.g. because, contrary to my beliefs and evidence, O will in fact produce exclusively bad
consequences), the mere fact that rationality requires it of me is not an objective reason for me to
choose O. But, he further claims, rationality is apparently normative, and he proposes to explain
this apparent normativity by what he calls the “Transparency Account”: Rational requirements are
apparently normative because, whenever one is subject to a rational requirement to choose option
O, it is in virtue of the fact that one believes oneself to have reasons of some more ordinary kind
(e.g., based on consequences) for choosing O that one is so required. Hence, from a first-personal
standpoint, rational requirements always line up with (because they represent) the normativity of
objective reasons. But, the argument continues, this Transparency Account only succeeds if there
is an appropriate conceptual connection between rational requirements and one’s beliefs about
objective reasons. Kolodny proposes, to this end, that all rational requirements can be derived
from what he calls the Core Requirements. In his preferred idiom of forming attitudes rather than
choosing options, these requirements go as follows.
C+ If one believes that one has conclusive reason to have [attitude] A, then one is rationally
required to have A.
C− If one believes that one lacks sufficient reason to have [attitude] A, then one is rationally
required not to have A. (Kolodny, 2005, p. 557)
C+ and C−, of course, closely resemble standard articulations of the enkratic principle, and like the
articulations I have offered above, C+ implies that Simon is rationally required to choose/intend
to fast in the desert, while C− implies that he is rationally prohibiting from choosing/intending
to revel in the city.
55
Humeanism and WSR depends on whether one accepts motivational internalism or
externalism about moral judgments. On the internalist view, an agent who sincerely
believes (or judges, on non-cognitive accounts) that she is morally required to choose
option O must possess at least some pro tanto motivation toward choosing O. For
the Humean, such motivation is at least prima facie reason-giving. Therefore, agents
who believe that they are obligated to choose O will generally have some objective
reason to choose O, and provided they recognize their motivation toward choosing
O, will have some subjective reason as well, just as WSR asserts.
The motivational externalist, on the other hand, holds that it is possible in
principle for an agent to judge that she is morally required to choose O and yet
to experience no motivation of any kind toward choosing O. Nevertheless, even
externalists accept that as a matter of fact, nearly all our moral judgments are
accompanied by at least some degree of motivation, and therefore the Humean
motivational externalist will still accept that in nearly all cases, agents who believe
that they are morally required to choose an option O have at least some reason to
choose O. Thus, while the proponent of the enkratic conception of rationality should
(as I have argued) accept WSR outright, the Humean should accept WSR at least
as a defeasible generalization holding true in most cases (defeasible in that an agent
may fail to exhibit any motivation corresponding to her full moral belief that she
is required to choose option O and/or in that this motivation may fail to generate
subjective reasons if the agent does not recognize it in herself or if other, deeper
features of her motivational state generate undercutting defeaters to the prima facie
56
reason for choosing O).11
Summarizing the argument of this section: WSR asserts that an agent who
fully believes, with conclusive justification, that she is objectively morally required
to choose option O has thereby some subjective reason to choose O. To accept WSR
is to accept (contra critics like Harman, Weatherson, and Hedden) that an agent’s
moral beliefs affect the rational requirements to which she is subject. There is a
strong prima facie case for accepting WSR and therefore the rational relevance of
moral belief: (i) WSR is plausibly entailed by the enkratic conception of rationality,
which I have argued is more plausible than the rival Human conception; and (ii)
even one who accepts the Humean conception of rationality should accept WSR at
least as a defeasible generalization, and therefore should accept that moral belief is
generally reason-giving and hence relevant to rational requirements.
2.3 The Prima Facie Case for SSR: Parallels between Moral and
Empirical Uncertainty
The case for SSR might also be made in terms of rationality-as-coherence:
If you believe that option O might be more objectively choiceworthy than any
alternative but experience no corresponding pro tanto motivation to choose O, this is
arguably also an instance of irrational incoherence. Likewise, if your degree of belief
that O is objectively most choiceworthy increases but your motivation to choose
11An analogous argument could be made that the Humean should accept SSR, if one believes
that partial as well as full moral belief is necessarily or generally motivational. These claims,
however, are more controversial than the corresponding claims with respect to full moral belief.
57
O does not correspondingly increase, this might be seen as a form of diachronic
irrationality.
But rather than pursuing this line of thought, I will rely instead on an argu-
ment commonly offered in support of moral hedging, that also supports the some-
what weaker SSR. This argument takes the form of a challenge: If, as nearly every-
one agrees, empirical risks and uncertainties are relevant to what agents subjectively
ought to do, then why not also moral risks and uncertainties? What is the relevant
difference between uncertainty about the empirical state of the world and uncer-
tainty about the basic principles of morality, such that the former but not the latter
are relevant to the requirements of practical rationality?
In the next chapter, we will examine some arguments that implicitly or ex-
plicitly endeavor to meet this challenge. But first, let’s try to cast the argument in
the following more positive light: Whatever arguments can be given that rational
requirements should be sensitive to an agent’s empirical uncertainties can be closely
paralleled by arguments that rational requirements should likewise be sensitive to
her moral uncertainties.
The case for the rational relevance of empirical uncertainty is most commonly
made by appeal to intuition. Two famous and structurally analogous thought exper-
iments, Frank Jackson’s drug case and Derek Parfit’s mineshaft case, serve as highly
effective pumps for this intuition. In both cases, an agent has three options, O, P,
and Q. She knows that, of O and P, one is the best of her options and the other
is catastrophic, but she has no way of knowing which is which and assigns equal
likelihood to either possibility. Q, meanwhile, she knows for certain to be nearly as
58
good as the best option (either O or P) and much better than the worst. In Jack-
son’s case, O, P, and Q are drugs that a doctor may prescribe to her patient. Of O
and P, one is a complete cure and the other a deadly poison, while Q is known to
be a partial cure that leaves the patient with only a few minor symptoms (Jackson,
1991, pp. 462f). In Parfit’s case, a mine is flooding and one hundred miners are
trapped in one of two shafts. The agent may either (O) block the lefthand shaft,
(P) block the righthand shaft, or (Q) block neither shaft. If she blocks the shaft
in which the miners are trapped, the flood waters will not enter that shaft and the
miners will be saved, but if she blocks the other shaft, all the water will be directed
into the shaft containing the miners, all of whom will be killed. If she takes option
Q, blocking neither shaft, the flooding will be moderate and only ten miners—at
the bottom of the occupied shaft—will be killed (Parfit, 2011, p. 159). In cases
of this sort, often collectively referred to as “Jackson cases,” it seems obvious to
most people that the agent is rationally required to choose option Q, as implied by
expected utility theory.
The equivalent of irrelevantism with respect to empirical beliefs is objectivism
about rational requirements: the view that an agent is always required to perform
the action that is objectively best, e.g., the action that has the best consequences.
But this is incompatible with the judgment that, in Jackson cases, agents are ra-
tionally required to choose option Q, since Q by stipulation does not have the best
consequences (and in fact, is the one option that the agent knows for certain will
not have the best consequences). Even if O, for instance, is the optimific action (in
Jackson’s original case, the drug that will completely cure the patient’s condition),
59
the agent’s uncertainty about the consequences of O should lead her to choose Q
instead.12
The moral of Jackson cases is that agents are not always rationally required
to choose the action that is objectively best, or even an action that has maximal
probability of being objectively best. Rather, rational requirements depend crucially
on an agent’s uncertainties and the balance of risk, at least with respect to the
empirical facts. If anyone is inclined to resist this conclusion, the intuition may be
driven home as follows: (1) Suppose you find yourself in the position of Jackson’s
doctor, and must prescribe your patient one of the drugs O, P, or Q. Which drug
would you in fact choose to prescribe? (2) Assuming your answer is drug Q, do you
view this choice as a mere idiosyncratic preference? An arational foible? Something
akin to an involuntary twitch? Or was your choice guided by reasons? It is very
hard to resist the conclusion that, in practice, we choose options like Q over options
like O and P without any hesitation, and that in so doing we view these choices
as rationally prescribed or required. This seems to commit us to the view that
rational requirements are sensitive to degrees of belief. Call this two-step argument
the Practice Test. In a moment, I will try to show that the same test can be applied,
with nearly the same force, to cases of moral uncertainty in service of the analogous
conclusion.
But first, let’s note a more abstract argument that might be given in favor of
12Even philosophers who defend an objective understanding of moral requirements typically do
not defend objectivism about rational requirements in the above sense. Rather, they attempt to
explain how the practical conclusion that a rational agent should choose option Q is compatible
with the claim that she ought to or is morally required to choose whichever of O or P is in fact
optimific. For one example of this strategy, see Oddie and Menzies (1992).
60
choosing option Q in Jackson cases: namely, an appeal to Kant’s famous principle
that “‘ought’ implies ‘can’.” To one who asserts that an agent subjectively ought (=
is rationally required) to choose whichever option of O and P in fact has the best
consequences, we are inclined to point out that the agent in the Jackson case cannot
comply with this prescription, because she cannot reliably identify the optimific
option. Of course, there is a sense in which the agent can choose the optimific
option: namely, she can choose O and she can choose P, and one or the other of
those options is optimific. But the true force of “‘ought’ implies ‘can”’ is clearly
stronger than this. It is not enough that an agent might manage to comply with
a prescription by luck or accident. For that prescription to represent a genuine
rational requirement, she must be able to comply with it rationally, that is to say,
reliably and on the basis of reasons. Her ability to do what she ought to do must not
be contingent on a lucky guess. Principles of rational requirement that are sensitive
to an agent’s degrees of belief in normative or evaluative propositions can satisfy
this constraint; norms that are sensitive only to facts about the world to which the
agent may lack epistemic access (e.g., which drug is in fact the perfect cure) cannot.
Now let’s consider an analogue of Jackson’s case, where empirical uncertainty
is replaced with moral uncertainty, and see what the “‘ought’ implies ‘can’” princi-
ple tells us about this case. Suppose you have just been elected World Dictator for
Life, and must choose a path for your loyal subjects to follow. Your moral beliefs
are evenly divided between hedonistic and perfectionist forms of average consequen-
tialism. You have a group of learned advisers who have crafted three possible social
schemes for you to consider, and have run (highly reliable) projections of the average
61
hedonic and perfectionist welfare levels that would result from each scheme.
• On the Spartan scheme, citizens will live in communal barracks, subsist on
a high-protein bean paste, and develop their moral and intellectual virtues
by 18-hour days of work, reading, and exercise. This plan carries an expected
value of 100 perfectons per person, but zero net hedons (hedons minus dolors).
• On the transhumanist scheme, citizens will live inside machines that simulate
pleasant experience while regulating their hormone levels to maintain a hedonic
optimum. (The machines will be serviced in perpetuity by an army of highly
dependable automata.) This plan carries an expected value of 100 net hedons
per person, but zero perfectons.
• Finally, on the eudaemonist scheme, citizens will live as participants in a
free society with a strong social safety net. They will be taught the value of
work, education, and exercise, and also given access to the same prohedonic
drugs that would be used in the experience machines. Because some in this
society will eschew work etc, and others will eschew drugs, the eudaemonist
scheme cannot generate hedonic or perfectionist optima. But it still carries an
expected value of 95 perfectons and 95 net hedons per person.13
13For a somewhat more down-to-earth moral-uncertainty-based Jackson case, see MacAskill’s
“Susan and the Medicine II” (MacAskill, 2014, pp. 9-11).
Nota bene that the assignment of cardinal values does not imply commensurability between the
hedonistic and perfectionist theories. We could, for instance, change the size of a perfecton so
that the three schemes generate 1,000,000/0/950,000 perfectons per person respectively, without
changing the substance of the thought experiment. But for the thought experiment as it stands,
we should assume that the size of a perfecton and a hedon are such that five perfectons and five
hedons both represent (what we would intuitively describe as) a minor but not trivial change in
someone’s welfare.
62
Given your distribution of moral belief, you are certain that either the Spartan
or the transhumanist scheme is objectively best. Nevertheless, it seems intuitively
rational to choose the eudaemonist scheme. And it seems likely that this is the
scheme that nearly anybody would choose, if they had the imagined distribution
of credence and accepted their advisers’ projections. As with Jackson’s original
case, we can apply the Practice Test: If you would select the eudaemonist scheme,
given your stipulated doxastic state, and if you view this choice as non-arbitrary
and guided by reasons, then you are committed to the conclusion that degrees of
moral belief are relevant to rational requirement (that is, your are committed to
SSR). Though you lack a full belief that the eudaemonist scheme is better than,
say, the transhumanist scheme, your partial belief (in the moral theory that implies
this to be the case) is, under the circumstances, sufficient to generate a rational
requirement, regardless of which moral theory ultimately turns out to be true.
Likewise, as with Jackson’s case, we can run an “‘ought’ implies ‘can’” argu-
ment: To say that you ought to choose whichever option is objectively best, as the
irrelevantist would have us say, is to give a prescription with which you have no ca-
pacity to (non-accidentally) comply. Just as in Jackson’s case you cannot discover
which of drugs O and P is the cure and which the poison, so in our moral ana-
logue you cannot discover which of the Spartan and transhumanist schemes would
be morally ideal and which a moral disaster. So a rational requirement to choose
the scheme that is in fact morally ideal would be a rational requirement with which
reason alone does not enable you to comply.
“‘Ought’ implies ‘can’” furnishes a negative argument against objectivist/irrelevantist
63
views of rational requirement. But one might also try to offer a positive explanation
for the rationality of risk aversion in the empirical domain, and here too, the most
natural explanations seem to transfer seamlessly from empirical risks (e.g. uncer-
tainty about the consequences of your actions) to moral risks (the risk that the true
moral theory is one that implies your action to be substantially morally wrong).
Consider, for instance, the following passage from Oddie and Menzies (1992), in
which they attempt to explain how an objective-utilitarian theory of value can jus-
tify subjective expected-utility maximization as a “selection procedure” for actions.
The regulative ideal for the moral agent is to maximize objective value.
But we have already noted two features of the ideal. First, it is an
ideal which admits of degrees of realization, and there is an obvious
and natural ordering of options with respect to the realization of the
ideal...Failing to achieve the ideal admits of degrees, and it is objectively
preferable for the agent to perform an act with more objective value
rather than less in those choices which do so fail. Second, the agent is not
usually in the position of knowing which option fully realizes the ideal. If
we put these two features together, then it is clear that in attempting to
realize the ideal the agent is not justified in selecting the option which, by
her lights, has the greatest probability of fully realizing the ideal, since
such an option may also have a significant probability of missing out
by a large margin. Rather, the agent is justified in selecting the option
which overall has the best prospects for objective value enhancement.
64
(Oddie and Menzies, 1992, pp. 530-1)
Reasoning along these lines plausibly explains our ordinary, obvious responses
to empirical risk: for instance, why I should not drink (or serve) coffee that I believe
with subjective probability .1 might contain a deadly poison, and why the doctor
in Jackson’s case should not prescribe either Drug O or Drug P. And this reasoning
requires little if any amendment to act as an argument for aversion to moral risk: I
know that the correct regulative ideal for the moral agent is the true moral theory,
or the features of the world to which that theory attributes moral significance. But
I don’t know what features those are (hedons? friendships? the good will?), insofar
as I am uncertain what moral theory is correct. Rather, the target I know I must
aim at is moral value as such, and in aiming at this ideal I will naturally be led to
account for all those features of the world that might have moral significance. In
other words, given that I recognize objective moral value as the goal towards which
I should act, I can do no better than to minimize the degree to which I deviate from
or fall short of this standard in expectation, and this requires that I be sensitive in
my practical deliberations to moral uncertainty and moral risk.
In summary, three of the most natural arguments for the claim that rational
requirements are sensitive to an agent’s empirical uncertainties and degrees of belief
also suggest that rational requirements are sensitive to an agent’s moral uncertain-
ties and degrees of belief. We can (i) construct intuition-pumping thought exper-
iments analogous to Jackson’s drug case and Parfit’s mineshaft case (augmenting
the intuition-pumping force of these thought experiments by means of the Practice
65
Test), (ii) appeal to the same negative “‘ought’ implies ‘can’” argument to explain
the intuitively falsity of the objectivist/irrelevantist response to these dilemmas, and
(iii) offer the same sort of positive explanation for the rationality of attending to
moral risk that we would naturally give with respect to empirical risk.14
There is, then, at least a prima facie analogy between empirical and moral
uncertainty. We need norms of rational choice that are sensitive to an agent’s
morally relevant empirical uncertainties, and there is no obvious reason why these
same norms should not extend to at least some cases of “pure,” non-empirical moral
uncertainty. Thus we have at least prima facie reason to accept SSR, which asserts
that rational requirements are sensitive to an agent’s moral degrees of belief. Of
course, it is still only a prima facie case, and in the next chapter we will examine
arguments that purport to defeat the analogy between these two kinds of uncertainty
and to show a fundamental distinction between the deliberative roles of empirical
and moral belief.
14It might seem that the arguments from Jackson cases and from “expected shortfall” principles
go further than justifying SSR and act as arguments for moral hedging or even a full-blown expected
value theory of rationality under moral uncertainty (a` la “maximize expected choiceworthiness”).
But one can accept the analogy between empirical and moral uncertainty without being driven
to any conclusion of a consequentialist or expected-value-maximizing flavor. In Chapter 5, for
instance, we will examine a case in which a non-expectational “threshold principle” for deciding
what to do when one is empirically uncertain whether a given course of action would violates a
deontological moral constraint seems to generalize naturally from empirical to moral uncertainty.
Thus the analogy between the two types of uncertainty can support various extensions of SSR, not
all of which will be forms of hedging.
66
Chapter 3: Challenges to the Rational Relevance of Moral Belief
There is prima facie reason to believe that an agent’s moral beliefs and degrees
of belief are relevant to the requirements of practical rationality. But these reasons
are not indefeasible. In this chapter we will examine some of the arguments that
have been recently offered for the irrelevance of an agent’s moral belief states to
what she subjectively ought to do or what rationality requires of her.
3.1 Normative Externalism
I begin by setting aside a debate that might seem—but, I claim, is not—crucial
to the defense of WSR and SSR. Weatherson frames his arguments for the irrelevance
of moral uncertainty as part of a broader defense of normative externalism, the view
that “[w]hat one should do, or should believe, in a particular situation is independent
of what one thinks one should do or believe, and (in some key respects) of what
one’s evidence suggests one should do or believe” (Weatherson, 2014, p. 141). The
externalist, as I will understand the position, claims that some norms of (theoretical
and/or practical) rationality are incumbent on an agent, have rational requiring
force with respect to her, whether or not she believes those norms are incumbent on
her or has any reason to believe that they are incumbent on her.
67
As we will see in Chapter 7, there is a compelling (and, I think, ultimately
successful) argument for normative externalism based on the threat of infinite regress
that the rival internalist position faces. The internalist claims that the norms that
apply to an agent are the norms that she accepts (or ought to accept, on the basis
of her evidence), and that these norms apply to her, roughly, to the extent that she
accepts them. Thus, when an agent is uncertain what “first-order” norm objectively
governs a particular choice, the various norms across which she distributes her belief
each contribute some subjective reasons for action, and (it seems) she must then go
searching for some “second-order” norm to tell her how to compare these reasons.
But what this second-order norm should be is itself an uncertain matter, so she will
no doubt end up distributing her belief over various, sometimes conflicting second-
order norms. When these norms do conflict she must, by the internalist’s lights, go
searching for some third -order norm that tells her how to deal with her uncertainty
about the second-order norms, and so on ad infinitum. Thus, it is claimed, the
internalist is trapped in a regress that paralyzes rational choice.
But even if, in order to block this regress, we must concede that some norma-
tive principle has its force independent of an agent’s beliefs or evidence, this does
not force us to accept the irrelevantist’s claim that first-order moral norms have this
sort of belief-independent normative force. Rather, as I will argue in Chapter 7, the
most plausible way of halting the regress described in the last paragraph is to accept
a basic principle of rationality—namely, the enkratic principle—as a constitutive or
definitional condition of rational agency and therefore as incumbent on agents re-
gardless of their beliefs or evidence. And since, as I argued in the last chapter,
68
the enkratic principle implies that what an agent subjectively ought to do depends
at least on her full moral beliefs, and plausibly on her partial beliefs as well, this
externalist conclusion is actively incompatible with irrelevantism. The irrelevantist,
therefore, cannot prove her case merely by arguing for normative externalism in
general, but must offer some separate argument to establish that first-order moral
norms have belief-independent normative force. More particularly, if she accepts
the rational relevance of empirical uncertainty, she must offer some argument for
a dividing line between “basic” moral norms (that apply to an agent regardless of
her empirical circumstances) and “derivative” moral norms (that apply to an agent
in virtue of the basic moral norms plus her empirical circumstances), such that the
former but not the latter have belief-independent authority. The regress argument
alone cannot discharge this burden of proof, but the arguments we consider in the
rest of this chapter attempt to do so.
3.2 Weatherson on Moral Fetishism
Weatherson (2014) states his states his position on moral uncertainty as fol-
lows.
Being uncertain about the physical consequences of your actions should
matter both to what you do, and how you are assessed. The red light
runner is immoral, even if she never actually harms anyone, because she
endangers morally significant humans. But the meat eater [who believes
but is not certain that meat eating is morally permissible] cannot be
69
condemned on the same grounds. If she is wrong that meat eating is
morally acceptable, that would be one thing. But a mere probability
that meat eating is immoral should not change one’s actions, or one’s
evaluations of meat eaters. (142-3)
I take this to be a denial of SSR and the rational relevance of moral uncertainty.1
Weatherson offers two arguments for this view, which deserve separate treat-
ment. The first claims that an agent who is sensitive to moral risks must exhibit a
defective motivational state that Weatherson, following Michael Smith, calls “moral
fetishism” (Smith, 1994). Though Smith introduces the idea of fetishism as part of a
very different argument, aimed against motivational externalism about moral judg-
ments, we will shortly see that the details of his argument bear a surprising relevance
to Weatherson’s. It is therefore worth a brief digression to set out Smith’s argument.
Smith argues that the motivational externalist—who holds that the sincere moral
judgment, e.g., that option O is morally required need not be accompanied by any
motivation to choose O—must explain why morally good agents nevertheless are
motivated to act in accordance with their moral judgments. This means that the
externalist must claim that a morally good agent is possessed of desires that lead
1Beyond this, I am somewhat uncertain how Weatherson’s view should be classified. At points
in Weatherson (2014), he seems to suggest a full-on denial even of WSR, for example when he
says: “I think the stronger, prima facie implausible, view is true: rightness and wrongness as such
shouldn’t even be part of our motivation” (p. 160). But he also concedes that an agent who acts
against her full moral beliefs can be convicted of “hypocrisy,” and seems to think that agents have
reason to avoid hypocrisy, which would amount to an acceptance of WSR (pp. 157-8). He also
regards it as at least acceptable for agents to be motivated by moral risks understood in terms of
“thick” moral properties, e.g., the risk that eating meat you are personally unwilling to slaughter
exhibits a vice of cowardice. If any case of moral uncertainty can be recast in such “thick” terms,
then Weatherson has for all intents and purposes accepted SSR and should then accept the need
for norms that guide agents’ choices under moral uncertainty—with the thought, perhaps, that
casting those norms in terms of thick moral properties will have substantial and interesting effects
on their content. If this is in fact his view, then so much the better.
70
her to act morally. According to Smith, these desires could be of only two kinds:
(i) “de re” motivations to take actions with the sorts of non-moral properties that
the agent judges to subvene on moral rightness or (ii) “de dicto” motivations to
perform morally right actions as such, i.e., to perform whichever actions turn out
to be morally right. An agent whose moral judgments are utilitarian, for example,
exhibits de re moral motivation if she desires to perform actions that promote the
general happiness independent of her judgment that these actions are morally right,
while she exhibits de dicto moral motivation if she desires to perform actions that
promote the general happiness precisely because she judges that those actions are
morally right.
Smith’s argument against externalism then takes the form of a dilemma: De
re moral motivation cannot be the mark of a morally good agent because it fails
to track changes in moral judgments. For example, if our morally good utilitarian
finally gets around to reading the Groundwork and is persuaded to abandon her
utilitarian convictions in favor of Kantianism, then as a morally good agent she
should become motivated to tell the truth and keep her promises even when doing
so will not promote the general happiness. But her de re motivations will not support
this shift, since they only impel her to promote the general happiness and are not
conditioned on her judgments about moral rightness. De re moral motivations do
not follow an agent’s moral judgments and hence, Smith claims, cannot be the
defining feature of a morally good agent.
If the morally good agent, as the externalist conceives her, cannot be char-
acterized by de re moral motivations, then she must instead be morally motivated
71
de dicto, i.e., motivated to perform right actions as such, because they are right.
But this, Smith claims, is not a feature of morally good agents at all, but rather a
vice, which he calls “moral fetishism.” Here Smith draws on Bernard Williams’ fa-
mous “one thought too many” objection to Kantianism and utilitarianism (Williams,
1981). Williams describes a man who must leap into a river to save his drowning wife
rather than another, equally drowning stranger. Theories like Kantianism and utili-
tarianism that make one’s personal attachments subordinate and accountable to the
universalistic and impartial demands of the moral law, Williams thinks, attribute
“one thought too many” to a morally good agent in this circumstance: Rather than
thinking “It’s my wife, so I must save her,” he instead thinks “It’s my wife, and the
maxim of saving one’s wife instead of a stranger can be consistently universalized”
or “It’s my wife, and in this circumstance saving her would be at least no worse for
the general happiness than saving the stranger.” It seems to Williams that such a
motivational scheme, that places the impersonal demands of morality ahead of per-
sonal commitments and values, is a character defect and not a virtue. Likewise, it
seems to Smith and Weatherson that an agent who is motivated to save her spouse,
or help her friends, or keep her promises, or donate to GiveDirectly merely on the
abstract and impersonal grounds that these things are morally right or good has
“made a fetish out of morality.”
Weatherson adapts Smith’s fetishism argument against the externalist’s de
dicto model of moral motivation into an argument for the permissibility of what he
calls “moral recklessness.” To be morally reckless is to ignore moral risks, e.g., to eat
meat despite one’s suspicions that eating meat might be morally wrong. Weatherson
72
argues that an agent who is motivated to avoid moral recklessness must be guilty
of moral fetishism—for instance, she avoids eating meat because it might be wrong.
If an agent who avoids moral recklessness must exhibit moral fetishism, and moral
fetishism is a vice, then Weatherson thinks it must follow that there is nothing wrong
with moral recklessness. This, then, is an argument against SSR: A moral agent
whose practical deliberations are sensitive to any positive degree of moral belief, for
instance that a certain action might be morally required or prohibited, is thereby
criticizable and so, presumably, cannot merely be acting as rationality requires.
There are a lot of things wrong with this argument. First, I confess that
I find Williams’, Smith’s, and Weatherson’s case intuitions paper thin and badly
misleading. To note the obvious, an agent can be responsive to “de dicto” moral
reasons of the sort postulated by WSR and SSR without needing to think, each
time she acts for those reasons, “...and I’m doing this because it is (or might be)
morally required of me.” Just as an airline pilot performing preflight safety checks
can be motivated by concern for the safety of her passengers without thinking, at
each step, a sentence like “By checking the deicing fluid levels, I help to keep my
passengers safe,” so an agent who acts to promote the general happiness because she
believes it to be her moral obligation need not stop to think to herself, as she rushes
into Singer’s pond to save the drowning child, “The Axiom of Universal Benevolence
demands it!” The “one thought too many” intuition that underlies the fetishism
objection, then, is misleading in part because it is based on a misrepresentation of
what it means for agents to be reasons-responsive and therefore what the defender
of de dicto moral motivation is committed to.
73
For my part, I don’t think that an agent who did reason through her moral
decisions in this hyper-explicit way would thereby necessarily be guilty of anything
worse than a little cognitive inefficiency. But contingently, human nature being what
it is, we would be rightly put off by anyone who always had to think, explicitly,
“...and doing such things is morally right” to summon up any motivation for right-
doing. Our moral psychology does not easily beget characters like Kant’s sorrowing
philanthropist. Rather, it seems almost impossible for a human being consistently
to act well without possessing the kinds of automatic sympathetic responses to the
needs of others that obviate the need for conscious reflection on the demands of the
moral law.2
Notably, also, it is easy to describe forms of de dicto motivation that do not
intuitively feel fetishistic or otherwise objectionable. Someone who, for example,
returns an accidentally over-generous tax refund to IRS “because it was the right
thing to do” seems to express laudable, not objectionable, motivations. (Indeed,
Weatherson admits that he must bite a bullet here and claim that people who say
such things, who feel themselves to be motivated by “conscience” or “a sense of
right and wrong,” are mistaken about their own motivations (Weatherson, 2014, p.
2I suspect, speculatively, that the underlying “one thought too many” intuition gets its force in
part from a superficial resemblance to a different, more persuasive argument, namely the “clean
hands” objection to deontology. When the deontologist says that she will not commit one murder
to prevent five, the objection goes, she places a concern for her own moral purity above a concern for
the rights or interests of others. She is more concerned with earning morality’s stamp of approval
than with the substance of morality, viz., those rights and interests. Likewise, when we consider
the sorrowing philanthropist or the man who pauses to think of the moral law before jumping in
the river to save his wife, we are apt to worry that their concern with doing the right thing has
become dangerous unmoored from the underlying substance of morality. And, in human beings as
we are, this may often be the case. But this is not a necessary pitfall of de dicto moral motivation:
It is perfectly possible, in principle, for an agent to be motivated de dicto while keeping clear sight
of the fact that doing the right thing requires attending to the interests of others, and not mere
mechanical adherence to a set of precepts.
74
161).)
But if one does find oneself compelled by the “one thought too many” intuition
and the claim that pure de dicto motivation is a vice, there is more to be said. Sepielli
(2016) offers three further objections to Weatherson’s fetishism argument. Each of
Sepielli’s objections strikes me as compelling, but I will confine my attention to the
first, namely, that consciously avoiding moral recklessness does not require an agent
to exhibit de dicto moral motivation.3 An agent might, Sepielli suggests, avoid
ordering steak at a restaurant because she is motivated to treat the cow with proper
respect, and she is morally uncertain whether eating the cow’s meat is consistent
with this goal. Here the agent exhibits something in between de re and de dicto
motivation, since her desires track a “thick” moral property, namely, treating beings
with proper respect. This sort of motivation seems to trigger the “fetishism” intuition
much less strongly than do purely de dicto motives, as Weatherson himself concedes
(Weatherson, 2014, p. 159).
Alternatively, Sepielli claims, an agent who avoids recklessly eating meat might
be motivated by purely de re concerns, like concern for the cow’s suffering, despite
her uncertainty whether the cow’s suffering has moral significance. Here, I think,
lies the most compelling response to Weatherson’s fetishism argument, though it
will take some effort yet to suss out. Note, to begin with, that in availing himself
of the second horn of Smith’s dilemma for externalists (the fetishism objection to
3Sepielli’s other objections are, in brief, that (ii) de dicto motivation in cases of moral uncer-
tainty is compatible with exhibiting desirable de re motivations in other sorts of cases, and so even
if de dicto motives were needed to avoid moral recklessness, these motives would not be objec-
tionable; and (iii) even if acting on norms that are sensitive to an agent’s moral uncertainties did
require that agent to exhibit objectionable motives, that fact would not constitute a refutation of
the norms themselves.
75
de dicto moral motivation), Weatherson impales himself pretty badly on the first
horn. Smith takes it for granted that a morally good agent must have motivations
that track her sincerely held moral judgments, and this assumption furnishes his
argument against the de re externalist model of moral motivation. But Weatherson’s
use of the fetishism argument commits him to denying this tracking claim: a morally
good agent, he claims, will be motivated de re, not by the features of the world that
she judges to have moral significance, but by the features of the world that in fact
have moral significance—and these, of course, do not change with an agent’s moral
beliefs.
This is a substantial intuitive cost to Weatherson’s view, and moreover it as
an unnecessary cost. If one wishes to avoid attributing de dicto motivations to
morally good agents, one would do better to adopt Smith’s own internalist view,
according to which sincere moral judgments are internally motivating, except in
agents who are suffering from akrasia or other rational failings (Smith, 1994, pp.
61f). Such a view clearly vindicates WSR, and contradicts Weatherson’s view that
an agent’s moral beliefs have no bearing on what she subjectively ought to do. A full
belief that option O is morally required constitutes, presumably, a judgment that
one is required to choose O, and thus on a Smith-ian motivational-internalist view
must give an agent at least pro tanto motivation to choose O, on pain of practical
irrationality.
Moreover, it seems plausible that this sort of motivational-internalist view
should vindicate SSR as well. Our ordinary moral judgments are beliefs to which we
assign, presumably, high probability but something less than certainty. For instance,
76
my judgment that I am morally required to rescue drowning children from shallow
ponds is about as strong as any moral judgment I hold, but it falls short of certainty
since I assign some credence, for instance, to moral error theories according to which
all statements of moral requirement are false. But if my (say) .9 credence that I am
morally required to save drowning children from ponds must be motivating in me,
on pain of practical irrationality, why should the same requirement not apply to my
.6 credence that I am required to avoiding harming distant future generations, or
my .4 credence that it is morally good, ceteris paribus, to bring additional happy
people into existence, or my .2 credence that it is good to preserve natural beauty
even when no one will benefit from the aesthetic experience of it? An internalist
view like Smith’s will end up endorsing SSR (i.e., that agents should be responsive
to their partial moral beliefs), unless it attributes undue moral significance to some
“Lockean threshold” for full moral belief/judgment.4 And given that this internalist
view escapes the fetishism objection and also capture the intuitive claim that a
morally good agent’s motivations should track her moral judgments, it is unclear
why Weatherson should not prefer this to his own de re-externalist view.
But moreover—and here we pick up once again the main thread of Sepielli’s
objection—there are externalist models of moral motivation that avoid both horns of
Smith’s argument for internalism, and that explain how an agent might avoid moral
4And in fact, as Sepielli points out, Smith himself is fairly explicit in accepting this implication:
“In his (2002), Smith tells us that for a rational agent, the motivatinoal force of a moral judgment
is a function of two features—‘importance’ (i.e. the moral significance the judgment assigns to
an action), and ‘certitude’ (i.e. the degree of confidence the judgment manifests). (pp. 309-
310) Nor does he regard motivation arising from uncertain judgments as any more fetishistic than
motivation arising from certain ones. (p. 213) Smith would agree, then, that an agent may be
ultimately motivated by things she thinks might be morally significant, even if...she is not sure
that they are.” (Sepielli, 2016, pp. 2957-8)
77
recklessness on the basis of purely de re motives. Jamie Dreier (2000) suggests a
model on which a morally good agent is characterized by a second-order motivation
to develop de re first-order moral motivations: that is to say, she desires to care
about the things that have moral significance for their own sake, and so for instance,
if she comes to believe that animal suffering has moral significance, will attempt to
cultivate in herself a de re concern for animal suffering. As Dreier points out, this
second-order motive is not simply de dicto motivation or moral fetishism by another
name, since the de re first-order desires at which the morally good agent aims are
not conditioned on her belief that the phenomena in question (e.g. animal suffering)
have moral significance, nor does this belief need to play any “maintenance role”
once the motivation is established. Rather, she comes to care about animal welfare
for its own sake, just as Weatherson would have her do.
Like Smith’s internalism, Dreier’s second-order model escapes concerns about
moral fetishism, explains the tracking relationship between moral judgment and
motivation, and in so doing clearly vindicates WSR: An agent who fully judges
that it is wrong to eat meat should develop (at least some) motivation against
eating meat. And again like Smith’s internalism, it seems natural that Dreier’s
model should vindicate SSR as well, for if a good agent has a second-order desire
to cultivate de re concern for the morally significant features of the world, then
her judgment that animal suffering might have moral significance will presumably
motivate her, at least a little bit, to develop a de re concern for animal suffering.
This, then, explains how a non-fetishistic agent might avoid moral recklessness: Her
judgment that eating meat might be wrong leads her to cultivate at least a limited de
78
re concern for animal interests that then imbues her with at least some motivation
to avoid meat consumption.
Indeed, it seems to me that Weatherson’s view itself requires one to adopt
something like Dreier’s second-order moral motive: For if you judge, as Weatherson
does, that a morally good agent is one who is motivated de re by the features of
the world that have moral significance, and you judge that some feature F has or
might have moral significance (e.g., animal welfare interests), it seems incoherent
(i.e., irrational) to experience no consequent motivation to develop in yourself a de
re concern for feature F. An agent who judges that, to be a good person, she must
have these de re concerns would at least desire and aim for de re concerns for those
features of the world that she fully judges to have moral significance, and again, it
is hard to see why this motivation should not extend (in proportionally diminished
degree) to partial judgments as well.
3.3 Weatherson on Prudential Uncertainty
Weatherson has a second argument for irrelevantism, to which I have already
alluded. Rather than analogizing moral uncertainty to empirical uncertainty, he
claims, we should instead draw an analogy with uncertainty about prudential val-
ues and norms, in which domain it seems counterintuitive that agents should be
motivated by partial belief. To pump this intuition, Weatherson offers the following
case.
Bob has thought a bit about philosophical views on welfare. In particu-
79
lar, he has spent a lot of time arguing with a colleague who has the G.E.
Moore-inspired view that all that matters to welfare is the appreciation
of beauty, and personal love. Bob is pretty sure this isn’t right, but he
isn’t certain, since he has a lot of respect for his colleague and for Moore.
Bob also doesn’t care much for visual arts...Now Bob has to decide
whether to spend some time at an art gallery on his way home. He
knows the art there will be beautiful, and he knows it will leave him
cold. There isn’t any cost to going, but there isn’t anything else he’ll
gain by going either. Still, Bob decides it isn’t worth the trouble, and
stays out. (Weatherson, 2014, pp. 148-9)
Bob, then, is insensitive to what we might call a “prudential risk,” namely,
the risk that by not stopping at the art gallery he fails to satisfy an important
welfare interest. But, Weatherson thinks—assuming that the pseudo-Moorean view
is wrong and that sterile “appreciation” of great artworks will not in itself make
Bob’s life go better—Bob has not acted irrationally or criticizably. And if Bob’s
case is relevantly similar to the case of the agent with a low but positive credence
that meat consumption is morally wrong, then this agent would likewise not be
acting irrationally or criticizably by continuing to consume meat.
I confess once again that I don’t share Weatherson’s intuitions. Assuming Bob
has a justified conviction that he ought to pursue all the constituents of a good life,
and is justified in assigning non-marginal credence to the view that the appreciation
of beauty is an intrinsic constituent of the good life, then it does not seem utterly
80
unreasonable that he should hedge his bets and make the occasional unrewarding
visit to the gallery.5
Moreover, Weatherson’s intuition can be explained away, revealing in the pro-
cess an important methodological principle for debates over normative uncertainty.
Note that Weatherson has chosen, as his example of a view about prudential inter-
ests in which one might have low but positive credence, a view that nearly everyone
will find extremely uncompelling. It is quite implausible, to most of us, that staring
joylessly at great works of art makes an indispensable positive contribution to the
goodness of a life. This first-order intuition significantly biases us against the second-
order conclusion that Bob should hedge his bets. Moreover, the first-order intuition
makes it hard to genuinely internalize the required stipulation of the thought exper-
iment, that Bob is justified in assigning non-trivial credence to the pseudo-Moorean
view. Where a thought experiment requires us to accept implausible stipulations,
our intuitions will often fail to reflect those stipulations, and that may be what is
happening in the case of Bob. The methodological principle, then, is that thought
experiments intended to pump intuitions for a particular view about how to re-
spond (or not to respond) to normative uncertainty should not involve first-order
views about which we have intuitions that mimic, or illicitly reinforce, the intended
thrust of the second-order intuition pump.6
Consider, by contrast to Weatherson’s case, a case of more reasonable pruden-
5Intuitions are confused here, I think, by the fact that we don’t typically view prudential
concerns as having quite the same sort of binding rational force as do moral concerns. It is unclear
what reason Bob has to pursue all the constituents of a “good life” if, fully informed and having
deliberated carefully, he still finds that some of those constituents leave him cold.
6I should now confess to having violated this principle myself in the Jackson case analogue from
§2.3. The reader may discount the force of that thought experiment accordingly.
81
tial uncertainty. Psychologists who study happiness commonly draw a distinction
between two aspects of subjective well-being, affect balance and life satisfaction (Di-
ener et al., 1999). Affect balance reflects the average of moment-to-moment affective
tone, including physical and emotional pleasures and pains. Life satisfaction reflects
an agent’s assessment of her life as a whole. While affect balance and life satis-
faction are, unsurprisingly, positively correlated, the correlation is far from perfect
and measures of happiness that focus on affect balance can reveal importantly dif-
ferent results than those that focus on life satisfaction. For instance, in one widely
discussed study, Kahneman and Deaton (2010) find that being wealthy rather than
merely middle class improves people’s sense of life satisfaction without improving
the balance of positive and negative emotion they experience. On a more immedi-
ate scale, many psychological studies find evidence for a “peak-end rule” by which
a person’s after-the-fact assessment of the goodness or badness of an experience is
predicted, not by the net balance of positive and negative affect over the course of
that experience, but primarily by her affective state during the most intense mo-
ments of that experience, and at the end of the experience. Thus, for instance,
a painful experience that tapers off slowly can be remembered as less unpleasant,
overall, than an experience of the same kind that was both shorter and less intense,
but involved no tapering period (Kahneman et al., 1993).
Now, it strikes me as perfectly reasonable for someone to wonder whether they
should aim more for experiences that maximize positive affect balance, or for experi-
ences that they will retrospectively regard as satisfying. Further, it seems perfectly
reasonable for an agent who believes (correctly, let us assume) that affect balance is
82
the true essence of subjective wellbeing, but assigns some non-trivial credence to the
possibility that self-assessed life satisfaction has intrinsic significance, should make
some concessions to the latter possibility, by sometimes choosing experiences that
involve slightly less positive affect but that she knows will seem more satisfying, or
be remembered more positively, in retrospect. And if this is right, then Weather-
son’s case has led us astray with respect to the rationality of hedging for prudential
uncertainty.
3.4 Arpaly and Schroeder on Inverse Akrasia
Though not explicitly presented as such, Nomy Arpaly and Timothy Schroeder’s
influential discussion of the case of Huckleberry Finn, and the more general phe-
nomenon that they dub “inverse akrasia,” presents an apparent argument for the
irrelevance of moral belief and uncertainty to subjective normativity. In Twain’s
novel (Twain, 1884), Huck runs away from home with the escaped slave Jim. Be-
lieving that he has wronged Jim’s owner by aiding his escape, Huck is wracked by
guilt and nearly turns Jim in to slave hunters. But his friendship for Jim overcomes
his moral convictions, and while he continues to believe that he is acting wrongly
by helping Jim, he resolves to ignore the purported demands of morality in favor of
helping his friend.7
Arpaly and Schroeder (1999) argue that Huck’s choice to help Jim escape
against his own moral convictions is fully praiseworthy, despite the fact that it is
7The case of Huck Finn originally entered the philosophical literature by way of Jonathan
Bennett (1974).
83
akratic, and that this phenomenon of “inverse” or praiseworthy akrasia indicates
that a morally good agent should not be predominantly motivated by her moral
convictions but rather by more particular, less cognitive and more affective concerns,
like sympathy and friendship.
Does this case refute WSR or SSR? First, note the gap between the claim
that a morally good agent should be exclusively motivated by her moral beliefs and
degrees of belief and the much weaker claims expressed by WSR and SSR, that a
morally good agent should be at least somewhat motivated (because she has at least
some reason) by her moral beliefs and degrees of belief. One can accept Arpaly
and Schroeder’s intuition about the Huck Finn case, and deny the former strong
claim, without denying either WSR or SSR. Indeed, this seems intuitively the right
response to the Huck Finn case: While we are glad that Huck ultimately decides to
help Jim escape, it is also to his credit that he feels bothered by the tension between
his sympathetic motivations and his conscience. If Huck were a simple amoralist
who said, straightaway, “I know that it’s wrong to help Jim escape, but so what,
that’s nothing to me!”, he would seem a less worthy character than he does. So at
minimum, we should say of the Huck Finn case that his de re concern for the welfare
of a friend properly overrides a weaker, but not undesirable, de dicto concern for
the dictates of conscience.
But actually, I think we can go a bit farther than this and undercut the force
of the Huck Finn intuition, in several ways. First, recall the methodological claim
from the previous section that intuition pumps for second-order views concerning
how agents should rationally respond to their moral beliefs and uncertainties ought
84
not take as their subject matter first-order moral views that arouse strong first-
order intuitions. This stricture is now even more salient. To block the easy response
that Huck’s moral beliefs about slavery lack epistemic justification—and that what
he ought rationally to do is revise his moral beliefs on the basis of evidence and
rational reflection, then help Jim escape because that is the right thing to do—it
must simply be stipulated that Huck’s moral beliefs are rationally justified. But this
is a hard stipulation to internalize, if ever there was one. Is it really plausible that
Huck has adequate grounds for believing that slaves are property and that justice
requires runaway slaves to be returned to their owners? And even if this stipulation
feels plausible, are our intuitions about how Huck should respond to this allegedly
justified moral conviction not still colored by our own conviction that Huck’s belief
is, in fact, badly mistaken?
To see that these forces are at work on our intuitions in the Huck Finn case,
consider instead another case of an agent deciding whether to act on justified but
false moral beliefs, where the belief in question does not arouse the same negative
first-order intuitions. Samuel Kerstein (2002) offers a case structurally analogous to
but substantively very different from the Huck Finn case, involving a one Colonel
Mikavitch.
A morally reflective person since she was a child, [Colonel Mikavitch]
has embraced the Formula of Universal Law as the supreme principle of
morality...Unfortunately and unforeseeably, a foreign power has attacked
the colonel’s country, bent on exterminating one of its ethnic minorities.
85
With the enemy nearly on her doorstep and no hope of escape, she
comes to the painful conviction that if she is captured, she will, under
the weight of torture, reveal a secret known only to her: the location of
several minority families. After careful consideration of the alternatives,
she has decided that the only way to save the families is to kill herself.
The colonel finds in herself no inclination to do so and, indeed, believes
that suicide would require her last ounce of courage. Although she thinks
she has a moral duty to save the families, she wonders whether it is
morally permissible for her to take her own life...After careful thought
she judges that this maxim passes the Categorical Imperative test...With
the regretful thought that she must heed the call to save innocent lives,
she takes poison. (Kerstein, 2002, pp. 121-2)
Suppose, however, that Colonel Mikavitch has committed some subtle error
in reasoning when testing the universalizability of her maxim, and that in fact the
Formula of Universal Law, the true supreme principle of morality, does not permit
suicide even in these extreme circumstances. Kerstein asks whether, under the cir-
cumstances, the colonel’s decision might still have moral worth, and concludes that
it does. For our purposes, we should ask a slightly different question: Did the colonel
act rationally, or as she subjectively ought to have acted? Given her prior mistake
in reasoning when she derived the obligation to commit suicide from the Formula
of Universal Law, it seems clear that the colonel’s subsequent decision to commit
suicide was fully rational, and indeed, it is hard to feel much hint of the intuition
86
from the Huck Finn case that it would have been better or more praiseworthy for the
colonel to disregard her moral convictions and avoid suicide. Where the first-order
terrain is relatively neutral, any intuition that “inverse akrasia” is deeply laudable,
let alone rational, seems to dissolve.
There is another way of testing whether the Huck Finn intuition is veridical.
Consider a variant of the case, which we will call Empirical Huck. Empirical Huck’s
beliefs are quite a bit different than those of Twain’s Huck. First, Empirical Huck
is a classical hedonistic utilitarian—and, let us stipulate, this moral belief is both
justified and true. But second, in this version of the antebellum South, everyone is
a convinced utilitarian, and the moral justification for slavery rests not on mistaken
moral doctrines of racial superiority or confused notions of property, but rather on
the empirical belief that the institution of slavery maximizes the general happiness
(say, by increasing economic growth rates). Huck is in no position to evaluate these
empirical claims, but justifiably defers to his teachers and other elders who assure
him that slavery is optimific. Nevertheless, in this world as in our own, slavery is
not optimific.
From here, the story of Empirical Huck unfolds like the original. Huck runs
away with Jim, finds himself pricked by the conscientious conviction that he should
turn Jim in for the sake of the general happiness, but can’t bring himself to do it
and, moved by friendship and sympathy, helps Jim escape instead. Now, I think our
intuitive reaction to Empirical Huck is not substantially different from our reaction
to the original case: It is good, right, praiseworthy, etc, that Empirical Huck helps
Jim escape. Presumably, though, no one is likely to mistake the Empirical Huck case
87
for knock-down proof that what an agent subjectively ought to do does not depend
on her empirical beliefs but only on the empirical facts. Empirical Huck does not,
for instance, refute expected utility theory. Rather, it indicates either that our
intuitions in both versions of the Huck Finn case are subject to interference from
powerful first-order intuitions, or else it simply indicates that the judgments of
goodness, rightness, praiseworthiness and so forth that we readily bestow on both
Huck and Empirical Huck should not be confused with subjective ought judgments
or judgments of practical rationality (a point to which we will return in the next
section).8
3.5 Harman on Moral Ignorance and Blame
One of the most explicit advocates of the irrelevantist position is Elizabeth
Harman. In “The Irrelevance of Moral Uncertainty,” she puts her view as follows.
8One possible explanation for our intuitions about both the original Huck Finn case and Em-
pirical Huck is the psychological phenomenon known as the “curse of knowledge”: Better-informed
agents often find it difficult to put themselves in the shoes of worse-informed agents and disregard
their own supererior information in predicting what those worse-informed agents will do. In an
experimental setting, for instance, participants in a simulated market who must predict the valua-
tion of a firm by other participants whom they know to lack certain information about the firm are
nevertheless biased in their predictions by their own possession of that information (Camerer et al.,
1989). Plausibly, this bias in our predictive judgments begets a bias in our normative judgments:
Because we know that slavery is wrong, we expect others to act as if they know this as well. Even
given the stipulation that Huck lacks our knowledge of the wrongness of slavery, then, the idea
of his turning Jim in seems not like expected behavior in light of his epistemic state, but rather
like an unexpected norm violation that could only be attributed to hidden malevolence or some
other defect of character. Or, on a slightly different hypothesis, perhaps the underlying cause of the
“curse of knowledge” phenomenon is a general inability to fully internalize information about other
agents’ epistemic/doxastic states insofar as they differ from our own, in which case our intuitions
about the case simply suffer from our failure to internalize the stipulation that Huck really be-
lieves (let alone that his evidence justifies him in believing) that he is morally obligated to turn in
fugitive slaves. This is also suggested by the phenomenon of hindsight bias, our apparent inability
to accurate internalize known limitations on our own past evidence, or that of others, in assessing
what ought to have been known or expected given those limitations (Roese and Vohs, 2012). This
is speculation, of course, but it should be matched against the arguably paltry evidence that our
intuitive judgments about cases like Huck’s are veridical to begin with.
88
A person’s moral beliefs and moral credences are usually irrelevant to
how she (subjectively) should act. How a person (subjectively) should
act usually depends solely on her non-moral beliefs and credences; her
moral beliefs and credences are relevant only insofar as they provide
warrant for beliefs and credences about what her non-moral situation
may be. (Harman, 2015, p. 58)
Harman’s argument for this position is straightforward: If an agent’s moral
belief state were relevant to how she subjectively ought to act, then agents who act
on (justified) false moral beliefs would be exculpated from blame for their wrongful
actions. But false moral beliefs don’t exculpate. So moral belief states are irrelevant
to subjective normativity.
Both premises of Harman’s argument are wrong, I think. First, it is not clear
why there must be the tight connection between subjective oughts/rationality and
blameworthiness that Harman supposes. Harman’s defense of this premise is fairly
thin: She asserts without further argument that “A person is blameworthy for some
behavior only if she should not have behaved in that way,” in a subjective sense
of “should” that appears to align with our notion of subjective oughts/rational
requirements (Harman, 2015, p. 75). But the norms that govern praise and blame
might be quite different from the norms of subjective rationality. There is nothing
obviously paradoxical in the idea that an agent caught in the grip of a false moral
view, for which she has adequate evidence and on which she acts enkratically and
as she rationally ought to act, might nevertheless be a legitimate target of blame,
89
and a victim in that respect of bad moral luck. Indeed, certain normative views
like utilitarianism will imply that this is quite often the case: If norms of praise and
blame are ultimately answerable to the principle of utility, then it may often turn
out that subjectively rational agents can incur legitimate blame. That Harman’s
argument relies on normatively substantive assumptions about what makes an agent
deserving of blame, and hence is inconsistent with at least some first-order moral
views, is a point to which we will return in the final section.
Absent the link to subjective rationality, Harman can have her claims that
false moral belief does not exculpate, without this constituting a refutation of WSR
or SSR. As I said in §2.1, my interest is only in the connection between an agent’s
moral belief states and the basic action-guiding normative system that we have been
speaking of in terms of subjective oughts/rational requirements. (Likewise, then,
even if Weatherson is correct that de dicto motivation is a defect of character or
Apaly and Schroeder are right that Huck is more virtuous and praiseworthy for
ignoring his moral beliefs, these claims need not be seen as inconsistent with the
theses I am concerned to defend, nor need they undercut the project of searching
for principles of rationality that govern choice under moral uncertainty.)
Still, I think Harman’s second premise—that false moral beliefs are never
exculpatory—is probably wrong as well, and the support she offers for it insufficient.
She argues for this premise by means of a few thought experiments, of which the
following is a representative example.
Max works for a Mafia “family” and believes he has a moral obligation
90
of loyalty to the family that requires him to kill innocents when it is
necessary to protect the financial interests of the family. This is his
genuine moral conviction, of which he is deeply convinced. If Max failed
to “take care of his own” he would think of himself as disloyal and he
would be ashamed. (Harman, 2015, p. 65)
Harman claim that, if Max commits murder on behalf of his crime family, he is a
“paradigm case” of an agent blameworthy for a wrongful action.
Two remarks: First, note once again the misleading force of first-order intu-
itions in this case. We should all feel, forcefully and without hesitation, that the
particular false moral belief that Max has accepted is badly and obviously mis-
taken, and this prejudices us against thinking that his belief might be exculpatory.
Likewise, it is difficult to appropriately internalize the stipulation (which Harman
recognizes must be made, to counter responses from the proponent of theses like
WSR and SSR) that Max’s moral belief is fully justified, that he has made no mis-
takes of theoretical rationality in reaching this belief that he ought to have corrected
before acting. Thus, our intuitions about the Max case are likely to be misleading.
But further, a single case in which false moral beliefs fail to exculpate does not
Harman’s case make for the much stronger premise, essential to her argument, that
false moral beliefs are never exculpatory. If one shares Harman’s judgment about
Max, then, consider by contrast Kerstein’s case of Colonel Mikavitch from §3.4. It
seems perfectly intuitive that the colonel’s justified false belief that the Categorical
Imperative requires her to commit suicide at least goes some way toward mitigating
91
her blameworthiness for violating the moral law. One can easily imagine other such
cases where an agent acts on moral views that are false but neither unreasonable or
horribly pernicious, out of a sincere sense of conscientious obligation, and in such
cases it generally seems intuitive that the agent’s moral belief state should affect
her liability for blame.9
3.6 Concluding Thoughts
I will close with three general observations about the irrelevantist position and
the arguments for it that we have just surveyed.
9At this point I can’t resist quoting a few lines from Gilbert and Sullivan’s The Pirates of
Penzance, in which the plot revolves around just such a case of sincere though mistaken moral
belief. As a young boy, the protagonist Frederic is mistakenly indentured to a pirate by his nurse
(who mishears the word “pilot”), with his apprenticeship to last until his twenty-first birthday.
Frederic “abhors [the] infamous calling” of piracy, but nevertheless does his best as a member of
the pirate band because, as he says, “[it] was my duty under my indentures, and I am the slave of
duty.”
As the story opens, Frederic believes his apprenticeship has expired and that he can return to
civilized society to atone for his piratical crimes. In the following exchange between Frederic and
the Pirate King, the question of blame and moral belief is raised, and the Pirate King eloquently
expresses a view of the subject that is sharply at odds with Harman’s.
Frederic. “Individually, I love you all with affection unspeakable, but, collectively, I look upon
you with a disgust that amounts to absolute detestation. Oh! pity me, my beloved friends, for
such is my sense of duty that, once out of my indentures, I shall feel myself bound to devote myself
heart and soul to your extermination!”
All. “Poor lad — poor lad!” (All weep.)
Pirate King. “Well, Frederic, if you conscientiously feel that it is your duty to destroy us, we
cannot blame you for acting on that conviction. Always act in accordance with the dictates of
your conscience, my boy, and chance the consequences.”
Before he can put his plan into effect, however, Frederic learns to his shock that he was born
on Leap Day, and that as a result he will not reach his twenty-first birthday and be free of
his indentures until age 84. Horrified but obedient to his sense of duty, Frederic rejoins his pirate
companions, rather than leading the local police force to arrest them. Informed of this development,
the police are nonplussed:
Sergeant. “This is perplexing.”
Police. “We cannot understand it at all.”
Sergeant. “Still, as he is actuated by a sense of duty?”
Police. “That makes a difference, of course. At the same time we repeat, we cannot understand
it at all.”
It seems to me that the police, like the Pirate King, have the right intuition: Frederic’s consci-
entious belief that he is duty-bound to rejoin his pirate comrades does make a difference to our
moral assessment of his actions, of course.
92
First: The irrelevantist view involves a curious setting-aside of the first-person
perspective on rational decision-making. How is someone who accepts a view like
Weatheron’s or Harman’s to actually go about making moral choices in the face of
moral uncertainty (for I assume it cannot be denied that we should all be at least a
little bit uncertain about the big questions of morality)? She knows that she ought
to do whatever the true moral theory implies she ought to do, but she does not
know—at least, does not know with certainty—what the true moral theory is. So
how will she proceed, when her view of subjective rationality gives her instructions
with which she does not know how to comply?
I can see only two alternatives: Either the irrelevantist position collapses into
something like My Favorite Theory, in that the agent who accepts irrelevantism
acts most rationally by taking her best guess and acting on the recommendations of
the theory she judges most likely to be correct — or it collapses into near-complete
arationalism about moral decision-making, since in the absence of any norms the
agent knows how to apply, she has no rational means of choosing among possible
actions. The latter option seems wildly implausible: There must be some way for
an agent to reason, non-arbitrarily, from her beliefs about morality to conclusions
about how she should respond to moral dilemmas. The former option is more
defensible, but to adopt it is to abandon irrelevantism and concede at least WSR.
The advocate of My Favorite Theory is “in the business” of looking for norms of
rationality that take account of an agent’s moral beliefs, just as much as the advocate
of a hedging view. If there is some other way, besides these two, of understanding
how an irrelevantist decides what to do in the face of moral quandaries, I don’t know
93
what it could be.
Second: Even if one finds oneself moved by intuition pumps like Williams’
“one thought too many,” Weatherson’s Bob, Arpaly and Schroeder’s Huck Finn, or
Harman’s Max, the intuition behind the enkratic principle (or some suitably revised
variant thereof) should seem like a far better bet. Apart from the specific reasons
I have given for taking the former group of intuitions with a grain of salt, there
are good reasons to be skeptical of reliance on case intuition in normative ethics
and metaethics generally (see, for instance, Singer (2005)). On the other hand, the
intuitions behind the enkratic principle (which arguably also lend their support to
stronger principles, like MEC, that vindicate SSR as well as WSR) are of a very
different and seemingly more reliable kind. If rationality is, conceptually, a matter
of internal coherence, then it is seems like a conceptual truth (or even a tautology)
that rationality requires an agent’s motivations to align in some fashion with her
beliefs about the good or about objective choiceworthiness. So if it comes down to
a weighing of competing intuitions, the intuitions behind the prima facie case at
least for WSR, and possibly for SSR, are strong enough to overwhelm their putative
defeators.
Third and finally: It strikes me that there’s something suspect, in the context
of asking how agents ought to act under moral uncertainty, about appealing to
substantive, first-order moral claims about things like virtue, praiseworthiness, or
blameworthiness. This move is suspect just as it would be suspect, in developing a
theory of how an agent should act when she is uncertain about the consequences of
her actions, to appeal to claims about those consequences—the very sort of claim
94
about which she is uncertain.10
The particular claims about virtue and blame on which Weatherson/Arpaly
& Schroder/Harman depend are not trivial, but significantly constrain the space
of possible moral theories. On a thoroughgoing consequentialist moral theory, for
instance, none of their claims about virtue and vice (de re vs. de dicto motivation),
praiseworthiness (inverse akrasia), or blameworthiness (Harman’s Max) are even
candidates for general truths about morality. Whether an agent is blameworthy for
an act, to a strict consequentialist, must reduce to the question whether it is opti-
mific to blame her for that act.11 And whether de dicto, “fetishistic” motivations
are a virtue or a vice, likewise, reduces to the question whether they lead an agent
to take better actions, compared to alternate motives by which she might otherwise
be moved. On these normative views, such assessment claims become fairly unin-
teresting from a metaethical standpoint, and even if contingently true, cannot be
called on to do the work that the advocates of irrelevantism require.
Of course, not everyone (perhaps not even most consequentialists) will accept
an equivalence between “blameworthy” and “ought to be blamed,” and hence one
10As MacAskill points out in a similar context, “it would clearly be incorrect to argue against
[maximizing expected choiceworthiness under normative uncertainty] because, in some cases, it
claims that it is appropriate for one to refrain from eating meat, even though (so the objection
goes) there’s nothing wrong with eating meat. That would be double-counting the arguments
against the view that it is impermissible to eat meat. In general it seems illegitimate to move from
claims about first-order normative matters to conclusions about which metanormative theory is
true.” (MacAskill, 2014, p. 40)
11Sidgwick writes: “From a Utilitarian point of view, as has been before said, we must mean by
calling a quality ‘deserving of praise,’ that it is expedient to praise it, with a view to its future
production: accordingly, in distributing our praise of human qualities, on utilitarian principles,
we have to consider primarily not the usefulness of the quality, but the usefulness of the praise...”
(Sidgwick, 1874, p. 428). This passage is cited by Norcross, who adds that “The utilitarian will,
of course, say the same about censure as Sidgwick says about praise: we should assess whether it
is good to punish or blame someone by assessing the utility of doing so. Punishing and blaming
are actions just like promisekeeping and killing and, like those actions, their value is determined
by their consequences, their power to produce utility” (Norcross, 2006, p. 225).
95
might reject the claim that praiseworthiness and blameworthiness are essentially
first-order normative notions.12 Still, the existence of candidate first-order views
that imply this sort of equivalence means that the arguments for irrelevantism we
have surveyed are not neutral with respect to first-order normative theories.
It could turn out, I suppose, that the second-order question, what agents ought
to do under moral uncertainty, depends on what first-order moral theory is true. But
this in itself would not constitute a victory for irrelevantism. It would imply only
that irrelevantism is true if some first-order theory is true that supports it (e.g., the
theories of virtue and blameworthiness the the advocates for irrelevantism presup-
pose). Another kind of first-order theory, like consequentialism, might support My
Favorite Theory, My Favorite Option, or a hedging view.
If the sorts of arguments for irrelevantism that we have been engaged with are
to be accepted as methodologically admissible, despite their reliance on first-order
normative claims that are potential objects of moral uncertainty, then we must
accept the dispiriting conclusion that we cannot know what norms of rationality
govern choice under moral uncertainty until we have resolved the very uncertainties
that make this question seem so pressing. On the other hand, if this sort of “upward
dependence” relationship between first-order moral theories and second-order norms
of rationality seems improbable, then there are methodological grounds to reject the
extant arguments for irrelevantism out of hand.
12An alternative analysis would equate “blameworthy” with “deserves to be blamed.” But like
blameworthiness, the notion of desert is either simply absent from classical utilitarianism or must
be reduced to the principle of utility in such a way that either the fundamental principles of desert
or the normative significance of those principles becomes a contingent, empirical matter.
96
Chapter 4: Dominance-Based Approaches
Having defended the claim that principles of rational decision-making must be
sensitive to an agent’s moral beliefs and degrees of belief, we can now ask what princi-
ples govern a rational agent’s choices under moral uncertainty. Among philosophers
who have sought to give a systematic treatment of the moral uncertainty problem
(e.g. Lockhart (2000), Sepielli (2009)) the goal has generally been to give an expected
value theory that assigns quantitative weights to the considerations put forward by
rival moral perspectives and then discounts those weights according to the degree
of belief the agent assigns to each perspective.
As we will see in Chapters 5-6, however, such expected value approaches en-
counter a very formidable obstacle: the apparent incomparability of moral consider-
ations put forward by competing moral theories, especially theories like Kantianism
and utilitarianism that differ radically in their basic structure. I will eventually
argue that this problem can be at least partially overcome. But before we engage
with the difficulties this task presents, it is worth asking what progress can be made
by more modest means.
In the moral uncertainty literature, there is a widely discussed class of cases
that have the following form: (i) An agent is faced with exactly two options, O and P.
97
(ii) It looks pretty clear that O is not morally required, and hence that P is morally
permissible. But (iii) there is a substantial possibility that O is morally prohibited,
and hence that P is morally required. Typically O is a morally risky action while
P is a morally safe omission, at least by our intuitive sense of the act/omission
distinction. Two widely discussed cases that fit the above schema are abortion and
meat consumption: while it’s pretty clearly morally permissible to avoid having an
abortion or eating meat, there is at least a substantial possibility that having an
abortion or eating meat are morally impermissible. As Moller (2011) puts it, these
cases present an “asymmetry” of moral risk. As such they seem like low-hanging
fruit for any theory of choice under moral uncertainty: the first interesting conclusion
that such theories are likely to yield is that we have reason to avoid actions that
carry such asymmetric or uncompensated moral risk.
A natural way of expressing this idea is by appeal to dominance reasoning. In
the context of reasoning under empirical uncertainty, a weak dominance principle
states that, if you are uncertain about the state of the world, but believe that
conditional on any possible state of the world option O is better than option P,
then you are rationally required to prefer O to P. What we want, however, is an
analogue of the stronger dominance principle stating that, if you are uncertain about
the state of the world, but believe that (i) conditional on any possible state of the
world option O has at least the same value as option P and (ii) conditional on some
possible state of the world option O is better than option P, then you are rationally
required to prefer O to P.1
1Somewhat confusingly, the weaker dominance principle is typically referred to as “Strong
98
MacAskill (2013) offers the most careful analysis to date of dominance-based
approaches to moral uncertainty. His initial statement of a strong dominance prin-
ciple ranging over moral theories goes as follows:
Dominance over Theories (DoT) — If some theories in which you have credence
give you subjective reason to choose x over y, and no theories in which you
have credence give you subjective reason to choose y over x, then, rationally,
you should choose x over y. (MacAskill, 2013, p. 511)
As MacAskill argues, and as we will see, this principle requires at least one important
revision. But it will provide a useful jumping-off point for our discussion.
It’s not clear, sadly, that a principle like DoT can get much traction on the
abortion and vegetarianism cases. For while these cases present a stark asymmetry
of moral risk, there may be substantial non-moral reasons that, for particular agents,
favor having an abortion or eating meat. Hence, in these cases, agents will be likely
to have credence in theories according to which they have overall subjective reason
to choose the morally risky option (abortion or meat), and so the antecedent of DoT
will not be satisfied.
If one is inclined to think that moral reasons always override non-moral rea-
sons, one might propose a modification of DoT that replaces “subjective reason”
with “subjective moral reason.” Such a principle would imply that asymmetric
Dominance” since it prescribes only the elimination of strongly or strictly dominated options, while
the stronger principle is typically referred to as “Weak Dominance” since it also prescribes the
elimination of weakly dominated options. It seems more natural to me, however, to call the weak
principle “weak” and the strong principle “strong,” so this is what I’ll do, where the distinction
needs to be made. In general, though, the following discussion will focus exclusively on the stronger
principle and its analogues for moral uncertainty.
99
moral risks should always be avoided, even when there are strong non-moral rea-
sons in favor of running these risks. But such a principle is controversial, to say the
least—and, perhaps more importantly for our purposes, it is a principle about which
agents should be substantially uncertain, i.e., reasonable agents will have positive
credence in normative theories that reject this principle.
Instead of the abortion and vegetarianism cases, then, let’s focus on a different
pair of cases that seem like even lower-hanging fruit for a dominance-based approach.
The first, which has received a certain amount of attention in the recent literature,
concerns “deflationary” moral views like skepticism or nihilism, and the possibility
that dominance reasoning can render such views practically inert. The second,
which has not yet received much attention, is supererogation. In both these cases
(I will argue), agents may be all-but-certain that some practical option is at least
permissible and potentially obligatory, and that another, incompatible option is at
best permissible and potentially forbidden. It is in cases like these that dominance
reasoning holds the most promise for making progress on the problem of moral
uncertainty.
The objective of this chapter, then, is to assess the potential for dominance-
based approaches to moral uncertainty through an examination of these two cases.
As we will see, the cases are anything but straightforward, and I will ultimately
conclude that dominance reasoning alone is insufficient even in these apparently
most promising contexts. But there will be an optimistic conclusion as well: In-
sofar as dominance reasoning can be conceived not as a freestanding principle but
rather as a limiting case on which various expected value theories and other “hedg-
100
ing” approaches to moral uncertainty converge, such reasoning can be coherently
leveraged to bolster the strength of moral reasons in the face of all but the most
extreme nihilist or supererogationist challenges—implying that very often at least,
moral considerations are still sufficient to generate rational requirements even when
agents have reason to be skeptical of the objective requiring force of those consid-
erations. So, while as a theoretical matter we cannot circumvent the more difficult
task of developing a stronger theory of rational choice under moral uncertainty, we
can as a practical matter have a great deal of confidence that whatever the cor-
rect comprehensive theory of moral uncertainty turns out to be, it will yield certain
non-obvious conclusions: for instance, that the possibility that all actions might be
morally indifferent or that many morally good actions might be merely supereroga-
tory does not negate the reason-giving force of moral considerations or, in most
cases, the rational requirement to act morally.
4.1 Moral Nihilism
Let’s begin with moral nihilism. Ross (2006) argues that we may almost always
disregard what he calls “absolutely deflationary” ethical theories according to which
it’s never the case that one course of action is more choiceworthy than another,
whatever our credence in such theories. He distinguishes, within this class, between
“uniform” theories according to which any pair of options is equally choiceworthy
and “nihilistic” theories according to which “the notions of good and bad and of
right and wrong are illusions and...objectively speaking, no option or state of affairs
101
is better than any other, nor are any two options or states of affairs equally good”
(Ross, 2006, p. 748).
Nihilism presents the more difficult and interesting case for Ross’s dominance
argument, so it is here that I will focus. Ross imagines a case in which he is
nearly certain that moral nihilism is true and finds himself faced with a trolley
dilemma. In such circumstances, he claims, he may simply disregard his credence
in nihilism for purposes of practical deliberation. He supports this claim by the
following dominance argument.
[S]uppose...that I have a degree of credence of .01 in TL [a non-nihilistic
moral theory that prescribes turning the trolley to the left], but...I have
a degree of credence of .99 in a nihilistic theory TN . And again suppose
that I must decide between sending the trolley to the right and sending
it to the left. In this case we could reason as follows. According to TL,
it would be better for me to send the trolley to the left than to send it
to the right. And so my credence in TL gives me pro tanto subjective
reason to send the trolley to the left. The only way this could fail to
be the most rational option would be if my credence in TN gave me
a sufficiently strong reason to send the trolley to the right. But TN
implies that there would be nothing valuable or disvaluable about either
alternative. And so my credence in TN gives me no subjective reason to
favor either alternative. Hence the pro tanto subjective reason to send
the trolley to the left is unopposed, and so this is the rational option.
102
(Ross, 2006, p. 748)
This line of reasoning seems straightforward, but it faces at least two significant
objections.
4.1.1 The Cyclicity Problem
The first objection, raised by MacAskill (2013), suggests that the dominance
principles on which Ross’s anti-skeptical argument depends generate a number of
intolerable results. In particular, attributing to Ross the principle DoT given above,
MacAskill shows that for an agent who has positive credence in one or more theories
that posit value incomparability, DoT has the unwelcome consequence of generating
preference cycles. The demonstration of this result depends on a pair of complicated
thought experiments. Stated schematically, and leaving aside the details MacAskill
offers to illustrate this schema, the simpler of the two thought experiments goes
like this: Agent A has three options O, P, and Q, and divides her moral beliefs
among three theories, T1, T2, and T3. T1 implies that O is better than P and
that Q is incomparable to both. T2 implies that P is better than Q and that O
is incomparable to both. And T3 implies that Q is better than O and that P is
incomparable to both. Hence, by DoT, O  P  Q  O (MacAskill, 2013, pp.
513-514).
To avoid cyclicity, MacAskill claims, we must limit ourselves to dominance
principles whose conditions of application exclude incomparability. In particular,
MacAskill replaces DoT with what he calls
103
Genuine Dominance over Theories (GDoT) If some theories in which you
have credence give you subjective reason to choose x over y, and all other
theories in which you have credence give you equal subjective reason to choose
x as to choose y, then, rationally, you should choose x over y. (MacAskill, 2013,
p. 518)
This weaker principle is bad news for Ross’s argument since nihilistic theories, as
Ross has characterized them, do not claim that we have equal subjective reason
to choose either option from any pair, but deny that notions like value, choicewor-
thiness, and subjective reasons are meaningful and hence that any comparisons in
terms of these notions, even comparisons of equality, can meaningfully be made.
Thus MacAskill concludes that dominance reasoning cannot justify rejecting ni-
hilism.
It seems to me, however, that there is a relatively straightforward escape from
the cyclicity problem, and that Ross has simply made a very inconvenient mistake
by treating equality as a positive value relation, such that nihilistic theories deny
that any two things are equally good. Consider by contrast John Broome’s proposed
definition of equality with respect to a property F : “‘x is equally as F as y ’ means
that [i] x is not F er than y, and [ii] y is not F er than x, and [iii] anything that
is F er than y is also F er than x, and [iv] y is F er than anything x is F er than”
(Broome, 1997, p. 72). If nihilism is true, then all four clauses in Broome’s definition
are trivially satisfied for any x and y and any evaluative property F (e.g. “good,”
“right,” “choiceworthy,” “supported by objective/subjective reasons”): If nothing
104
is better than anything else, then x is not better than y, y is not better than x, and
since neither x nor y is better than anything, it is vacuously true that for anything
either x or y is better than, the other is better as well. Furthermore, by virtue of
these last two clauses, Broome’s definition distinguishes (as Broome intends it to)
between equality and other relations like parity and incomparability in the context
of non-nihilistic theories.
Thus, we can strengthen MacAskill’s GDoT by simply replacing the positive
relation of equality, which nihilistic theories decline to attribute to any pair of op-
tions, with Broome’s negative relation. (Let’s call the latter “equality*,” for ease of
reference.)
Strengthened Genuine Dominance over Theories (GDoT*) If some theories
in which you have credence give you subjective reason to choose x over y, and
all other theories in which you have credence give you equal* subjective reason
to choose x as to choose y, then, rationally, you should choose x over y.
GDoT* allows Ross’s anti-nihilistic argument to go through, while remaining im-
mune to the preference cycles brought on by the incautious principle DoT: In
MacAskill’s case described above, T1 does not imply that there is equal* subjec-
tive reason to choose O as to choose Q, since there is greater subjective reason to
choose O than to choose P but not greater subjective reason to choose Q than
to choose P, nor is there equal* subjective reason to choose P as to choose Q, by
precisely the same reasoning (and likewise, mutatis mutandis, for T2 and T3).
2
2For purposes of this argument, it makes no difference whether we frame GDoT* in terms of
support by subjective reasons or any other normative property like rightness or choiceworthiness.
105
Thus, the dominance argument for rejecting nihilism can escape the threat
of cyclicity. However, as we will see in §4.3.1, MacAskill’s problem will reappear
as a challenge to the dominance argument for rejecting supererogationism, and will
somewhat constrain the scope of that argument.
4.1.2 Nihilistic Undermining
There is another difficulty that strikes me as more worrisome for Ross’s argu-
ment and which has yet to receive attention in the literature. This is the worry that
principles like DoT, GDoT*, and any other dominance principle to which Ross’s
argument might appeal seem to be undermined by the very same nihilistic hypothe-
ses that the argument is meant to justify rejecting—or at least, by some of those
hypotheses.
Consider a form of nihilism I will call Democritean normative error theory
(DNET). According to this view, only fundamental entities exist (flatworldism3)
and all fundamental entities are physical (physicalism). “Reasons,” “obligations,”
“goods,” “values,” etc, are not fundamental physical entities, so no reasons, obliga-
tions, goods, values, etc exist. For any normative/evaluative proposition to be true,
at least one such entity must exist to serve as its truthmaker. Therefore, DNET
holds, no normative/evaluative propositions are true.4
Is such an extreme view plausible? Though I don’t assign it the greater part
3I borrow this term from Karen Bennett (2011).
4The name I have given this view is not meant to imply an attribution to Democritus. But the
spirit of the view is neatly captured by his famous adage: “By convention sweet and by convention
bitter, by convention hot, by convention cold, by convention color; but in reality atoms and void”
(Taylor, 1999). The DNETist just adds to this list: “by convention right, by convention wrong,
by convention rational, by convention irrational...but still, just atoms and void.”
106
of my own credence, it seems to me that DNET is a more reasonable contender
for truth than many popular normative views (I leave it to the reader to fill in
examples), and my impression is that a significant number of philosophers accept
something like it (though, by self-selection, few of these are normative ethicists).
It combines two widely held views in physicalism and flatworldism, and draws a
plausible implication from the conjunction of these views.5 I will assume, then, that
DNET is plausible enough, as nihilistic theories go, to merit some investigation.6
Now, Ross’s argument suggests that no matter how high my confidence in a
nihilistic theory like DNET, I should nevertheless accept a second-order dominance
principle like GDoT*.7 The problem is this: Since GDoT* expresses an affirmative
5Of course, even if one accepts physicalism and flatworldism, one might still try to make nor-
mative propositions true by letting atoms and void alone serve as their truthmakers. This could
be done by paraphrasing away existential claims concerning reasons and the like (“atoms arranged
reason-wise,” perhaps), by eschewing such talk entirely, or by defending it as ineliminable but
ontologically non-committing. Since my goal is not to defend DNET, I won’t attempt to assess
these strategies individually. I do, however, find them at least mildly implausible. Suffice it to say
that normativity in the most interesting sense (features of the world that count in favor of actions
and make our actions matter) must be something that carves the world at a fairly fundamental
joint, and analyses of normative concepts that make them logically or conceptually (rather than
metaphysically) supervenient on atoms and void appear to carve no such joint.
6Ross attempts to avoid the undermining worry by fiat, restricting the scope of his argument
to “a kind of nihilism that denies that there are any objective reasons for action but that concedes
that if one held that there were such reasons, then it would be subjectively rational to act in
accordance with them” (Ross, 2006, p. 749n). In fact, Ross’s stipulation must go a bit further
than this, since his dominance argument is meant to apply not only to agents who hold (i.e., fully
believe) that there are objective reasons for action but to any agent who has positive degree of belief
that there are such reasons. Thus, the version of nihilism to which Ross restricts his attention
must deny that there are any objective reasons for action but also affirm that there are subjective
reasons for action, and that any agent with non-zero credence in the existence of objective reasons
has subjective reasons. There is a significant prima facie tension, however, between denying the
existence of all objective reasons and asserting the existence of subjective reasons, and Ross gives
no argument that plausible nihilistic theories should be understood in this way. So it seems that
this self-imposed restriction would limit the scope of Ross’s argument to a relatively implausible
subset of nihilistic theories.
7A variety of other dominance-like principles could be offered to facilitate Ross’s argument.
MacAskill considers a second principle called “Modified Dominance over Theories” (MDoT) that
achieves the same result as DoT (but shows that this principle still generates diachronic prefer-
ences cycles over certain sequences of pairwise choices) (MacAskill, 2013, pp. 515-6). Weatherson
attributes a different principle to Ross which he calls “ProbWrong,” then suggests a schematic
version of this principle called “GeneralPrinciple” that could be filled in by various evaluative
107
normative proposition (a “should” claim), DNET ` ¬GDoT*. So if I accept DNET
with subjective probability p, I must assign GDoT a degree of belief no greater than
(1 − p). If p > .5, then Cr(GDoT*) < .5. And, assuming that my beliefs ought to
be consistent, if I believe DNET, I ought to disbelieve GDoT*.8 But how can it be
rational, let alone rationally required, to act on a principle that I regard as more
likely false than true, or that I believe to be false?
It seems to me that the solution to this undermining worry, if there is one,
must involve taking the relevant principle of rational requirement—a dominance
principle like GDoT* or some stronger principle from which it can be derived—
to have “external,” belief-independent force in Weatherson’s sense, such that the
agent is subject to its requirements even if she justifiably disbelieves it. Whether we
should accept normative externalism is a question I will not attempt to resolve in
this chapter. (In Chapter 7, however, I will argue that the enkratic principle, rightly
formulated, has external rational requiring force, and propose several candidate
formulations each of which plausibly implies GDoT*.) But we will shortly return
to the challenge of undermining, which arises in a closely analogous form in the
context of the dominance argument for rejecting supererogationism. There I will
argue that the versions of supererogationism that threaten to undermine the relevant
predicates to facilitate the same dominance arguments (Weatherson, 2014, p. 146). But the “un-
dermining” objection described in this section applies equally and straightforwardly to all these
principles.
8If GDoT* is read as a material conditional, this argument becomes slightly more complicated—
GDoT* then is not itself an affirmative normative proposition, but implies such a proposition when
conjoined with empirical propositions about my credence over theories that I presumably accept
(and that I must accept, for GDoT* to apply to me and for Ross’s argument to gain traction).
But this does not affect the substance of the undermining problem: GDoT* is still inconsistent
with the conjunction of empirical facts about my credal state (which make its antecedent true)
and DNET (which makes its consequent false).
108
dominance principles are less plausible than the corresponding versions of nihilism,
and hence that the issue of normative externalism is at least less urgent in the
supererogation context.
4.2 Supererogation
Let’s turn now to our second test case for dominance reasoning, the case of
supererogation. Suppose agent A must choose between options O and P. Given these
options, she is uncertain whether O is morally required or merely supererogatory
and correspondingly, she unsure whether P is forbidden or permissible. In this
circumstance, dominance reasoning seems to suggest that A is rationally required
to choose O, despite having some (perhaps quite high) degree of belief that P is
objectively permissible.
Paralleling (or rather, parroting) Ross, the argument can be put like this: Call
the moral theory according to which O is objectively morally required TR and the
theory according to which it is merely supererogatory TS. According to TR, A is
morally required to choose O and forbidden to choose P, so A’s credence in TR
certainly seems to give her pro tanto subjective reason to choose O. The only way
this could fail to be the most rational option would be if A’s credence in TS gave her
sufficiently strong reason to choose P instead of O. But although TS implies that
it is permissible to choose P, it certainly does not imply that it would be better to
choose P than to choose O. At most, it implies that the two options are equally
choiceworthy (equally “eligible,” one might say). Hence the pro tanto subjective
109
reason to choose O is left unopposed, and so this is the rational option.
4.2.1 “All-Things-Considered” Supererogation
This argument depends crucially on the claim that, according to plausible
supererogationist theories, the strength of one’s all-things-considered reasons for
choosing a supererogatory option OS is (often or always) equal to or greater than
the strength of one’s all-things-considered reasons for choosing any of its merely-
permissible alternatives. Let’s call option OS morally supererogatory iff for some
alternative option OP , OS is morally better than OP , but OP is morally permissible;
and let’s call OS all-things-considered supererogatory iff for some alternative option
OP , OS is better than OP all-things-considered (that is, supported by stronger all-
things-considered reasons) but OP is still all-things-considered permissible (that is,
permissible as a matter of objective rationality). In these terms, then, the claim
on which the dominance argument for rejecting supererogationism depends is that
plausible supererogationist theories should regard most or all supererogatory options
not just as morally supererogatory, but as supererogatory all-things-considered.9
This claim is prima facie plausible when we reflect on paradigm cases of su-
pererogation, like donating a moderately substantial portion of one’s income to
charity. Even if we regard such actions as merely supererogatory, we do not re-
9This restatement sets aside the possibility of exact equality in the strength of reasons for
performing OS and OP . Equality of reasons is, in general, extremely fragile, being disturbed by any
increase or decrease in the strength of reasons favoring either option. Since it is fragile, equality
is also extremely rare. Thus, while the dominance argument for rejecting supererogationism is
compatible with the possibility that there is equal all-things-considered reason to perform OS as
to perform OP , it will far more often be the case that there is at least slightly greater all-things-
considered reason to perform OS .
110
gard an agent who performs them as acting irrationality, contrary to the all-things-
considered balance of reasons. We can of course describe cases where we might
be inclined to say that, although an agent has acted well from the moral point
of view, she has acted on comparatively weak moral reasons and against stronger
prudential reasons, and therefore has acted irrationally overall: think of Ned and
Maude Flanders sending a once-disheveled stranger off in Ned’s best suit, with the
promise to sleep on card tables if he ever wants the use of their master bedroom.10
But such cases go well beyond ordinary acts of supererogatory altruism, and our
attitude toward them is quite different from our attitude toward, say, someone who
donates five percent of her weekly paycheck to GiveDirectly. Unlike the absurd
self-abnegation of the Flanders family, common sense regards these more reason-
able forms of altruism as admirable and praiseworthy, not just from a moral point
of view, but simpliciter.11 Though many of us judge that even moderate altruism
(e.g. at the 5% level) is supererogatory rather than obligatory, it is hard to imagine
that any plausible theory of all-things-considered practical rationality (so long as it
incorporates non-egoistic moral considerations at all) will judge that we have less
overall reason for making such sacrifices, when they will greatly improve the lives of
the very badly off, than we have for spending the money on ourselves.
One might think, however, that this must be a mischaracterization of our
10The Simpsons, S3 E24, “Brother, Can You Spare Two Dimes?”
11Likewise, it is the Flanders-like extremes of altruism that give intuitive plausibility to Susan
Wolf’s skepticism about “moral saints” (Wolf, 1982). As Wolf argues, a life of moral perfection
may be genuinely better from the moral point of view than a more ordinary life, but either worse or
at least incomparable all-things-considered because of the nonmoral goods that must be sacrificed
in the pursuit of moral perfection. But one can grant this conclusion without denying that some
voluntary altruism in excess of the bare moral minimum may be good and praiseworthy from an
all-things-considered (and not merely a moral) point of view.
111
commonsense judgments, since the all-things-considered understanding makes the
notion of supererogation self-evidently and intolerably paradoxical. If rationality
always requires us to choose the option that is supported by the strongest all-things-
considered reasons, then the definition of all-things-considered supererogatory op-
tions as more strongly supported by reasons and yet rationally optional is a simple
contradiction in terms, and so cannot be what is posited by plausible supereroga-
tionist moral theories.
There are a great many proposals in the literature that aim to resolve this
“paradox of supererogation” and render the notion of the supererogatory coherent
with a general theory of rationality. Fortunately, most of these proposals are con-
sistent with if not actively supportive of the commonsense judgment that ordinary
supererogatory acts better and more admirable than their merely-permissible alter-
natives, not just from the moral point of view, but simpliciter. Portmore (2008), for
instance, considers five possible ways of understanding supererogation as part of a
general theory of rational options, several of which allow us to make sense of the idea
of all-things-considered supererogation. Specifically, we may understand all-things-
considered supererogation in terms of a satisficing conception of rationality (that
treats the merely-permissible alternative to a supererogatory act as rationally per-
missible because it meets some satisficing threshold, even though there is stronger
overall reason to perform the supererogatory act); in terms of a distinction between
the justifying strength and requiring strength of reasons (according to which the
reasons that make the supererogatory option better supported by reasons overall
nevertheless lack the requiring strength of produce a rational (or moral) require-
112
ment)12; or in terms of imperfect reasons (which, in supporting the supererogatory
act, make it at least as rational all-things-considered as any alternative, but do not
generate a rational requirement insofar as they can be satisfied by the agent per-
forming other morally good acts in other circumstances).13 In a similar vein, Raz
(1975) proposes to understand supererogation in terms of what he calls “exclusion-
ary permissions,” which empower an agent to set aside or exclude the reasons that,
on balance, favor a particular course of action.14
Finally, Urmson’s classic article in defense of supererogation (Urmson, 1958)
proposes that supererogatory acts be understood as differing from acts of moral
obligation simply in that the difficulty of complying with the more demanding prin-
ciples of supererogatory moral behavior means that demanding such behavior of one
another, treating supererogatory acts as obligatory under our shared moral scheme,
is likely to undermine the normative force of moral requirements in general to an
extent that outweighs any potential social gains from compliance. On this sort of
broadly pragmatic understanding of supererogation there is, likewise, no difficulty
in the idea that many if not all supererogatory acts are more strongly supported by
12Horgan and Timmons, among others, have defended this understanding of supererogation:
“What the case of Olivia suggests, then, is the idea that not all good moral reasons for an agent to
perform some action, even reasons that are plausibly considered ‘best,’ are such as to require that
she perform that action, even prima facie. Some moral considerations clearly do have a requiring
force, but (we submit) others need not.” (Horgan and Timmons, 2010, p. 50)
13The other two possibilities Portmore considers, drawing on the ideas of parity and incompa-
rability/rough comparability between reasons respectively, are incompatible with the argument
I have given, as we will see below. But, as I will argue, they are also incompatible with our
commonsense judgments of the merits of supererogatory acts.
14“A person may have an exclusionary permission to perform an act even though there are
conclusive reasons for him not to perform it, provided that he is entitled not to act for those
reasons, to exclude them from his considerations. In other words: To say that a person is permitted
to perform an act is to say that he may perform it, i.e., that he does nothing wrong in performing
it. He is permitted to perform the act because there are no conclusive reasons against doing so or
because he may exclude such reasons from his considerations.” (Raz, 1975, p. 163)
113
reasons as their merely-permissible alternatives.
To come at the point from the opposite direction, there are just three possible
understandings of supererogation with which the argument in the preceding section
is incompatible. For ordinary cases of a putatively supererogatory option OS and a
merely permissible alternative OP , in claiming that the all-things-considered reasons
to perform OS are equal to or stronger than all-things-considered reasons to perform
OP , I have thereby denied that (a) there is greater reason to perform OP than to
perform OS, that (b) the reasons favoring the two options are incomparable, or that
(c) the reasons favoring the two options are on a par (a fourth relation of evaluative
comparison, supplementing the traditional trichotomy of better/worse/equal, pro-
posed and defended by Ruth Chang, e.g. in Chang (1997b)). (That I must deny
(b) and (c) as well as (a) will be seen in the following section.)
The first of these possibilities, that supererogatory acts are generally opposed
by the balance of all-things-considered reasons, is strongly contradicted by our com-
monsense evaluative judgments, and has not found advocates in the philosophical
literature.15 The second possibility, that of incomparability, is likewise highly im-
plausible, since an enormous number of our everyday moral judgments presuppose
comparability between the sort of altruistic reasons that commonly favor supereroga-
tory acts and the sort of prudential reasons that oppose them. For instance, the
judgments on the one hand that one has stronger reason to save the drowning child
in Singer’s pond that to preserve one’s new pair of shoes (Singer, 1972), and on the
15Douglas Portmore has come nearer than others to defending this view, but disclaims it in
Portmore (2008, p. 382n).
114
other hand that Ned Flanders acts unreasonably in sacrificing his every personal
interest for the sake of whatever needy stranger lands on his doorstep, both indicate
that moral and nonmoral reasons are far from incomparable.
The most plausible understanding of supererogation that my argument must
exclude, then, is one that draws on the notion of parity. On such an understanding,
the strength of all-things-considered reasons for choosing some supererogatory op-
tion OS is neither greater than, nor less than, nor equal to the strength of all-things-
considered reasons for choosing some alternative OP , but a sufficient strengthening
or weakening of the reasons on either side (for instance, an increase in the good
that might be done for others by, or in the personal costs of, performing OS) would
generate a definite inequality, making one option more rational overall.
It seems to me, however, that this parity-based account of supererogation is
still inconsistent with our ordinary attitudes toward supererogatory acts. Parity im-
plies a kind of symmetry between options: Neither option is better than the other,
and either option could come to be better by the addition of reasons in its favor, or
the subtraction of reasons for its alternative. But our normal attitudes toward the
agent who gives more of her income to charity than morality strictly requires do not
reflect such a symmetry: our admiration for her willingness to forgo creature com-
forts of the sake of others in greater need is not mirrored by a deprecating judgment
that she is somewhat soft-headed or imprudent, nor by an equal admiration for the
steely-eyed prudence of her colleague who only ever gives the morally obligatory
minimum. Again, we view the agent who acts supererogatorily as having acted well
and admirably, not simply from one point of view among many, but simpliciter.
115
It seems to me, then, that the most plausible understandings of supererogation
will not treat supererogatory options as on a par with, incomparable with, or less
overall choiceworthy than their merely permissible alternatives. Rather, they will
endorse all-things-considered supererogation, and regard supererogatory options as
more overall choiceworthy than their alternatives, though not all-things-considered
required or obligatory. This gives the dominance argument for rejecting supereroga-
tionism at least prima facie plausibility.
4.3 Rejecting Supererogationism: Cyclicity and Undermining
Revealingly, however, the anti-supererogationist application of dominance rea-
soning faces a pair of objections that closely parallel the objections to Ross’s anti-
nihilistic argument discussed in §4.1.
4.3.1 Cyclicity and Supererogation
MacAskill’s cyclicity worry, remember, arises insofar as nihilistic theories are
construed as treating any two options as incomparable in terms of objective choice-
worthiness. An analogous worry arises for the anti-supererogationist argument in-
sofar as supererogation is construed as a matter of either incomparability or parity
between supererogatory and merely permissible options. Suppose, for instance, I
have some credence in a theory according to which the option of donating $500 to
GiveDirectly is neither more, nor less, nor equally as choiceworthy as the option of
buying a new TV (i.e., the two options are either incomparable or on a par) and
116
some credence in a theory according to which I am obligated to donate to GiveDi-
rectly. Then the principles GDoT and GDoT* will be inapplicable and will not
generate a dominance argument for donating to GiveDirectly. The principle DoT
will generate such an argument, but at the cost of preference cycles, as we saw in
§4.1.
How much does a restriction to acyclic principles like GDoT* limit the anti-
supererogationist argument? One the one hand, we have just seen reasons why the
problematic theories of supererogation as incomparability or parity are implausible.
On the other hand, to whatever extent these versions of supererogationism do seem
plausible to a particular agent, there is no easy way of saving the dominance ar-
gument against supererogation as there is with nihilism. A normative theory that
regards the options of buying a new TV and donating to GiveDirectly as all-things-
considered incomparable does not treat these options as equally choiceworthy even
in the “negative” sense of equality described in §4.1.1, the sense in which nihilistic
theories do necessarily regard any two options as equally choiceworthy. Thus, if an
agent does have positive credence that these options are overall incomparable or on
a par, dominance reasoning alone cannot generate a rational requirement to perform
the morally safe option of giving to charity, on pain of inviting cyclicity.
4.3.2 Undermining and Supererogation
But just as the greatest worry for Ross’s anti-nihilistic argument is that certain
forms of nihilism will undermine the very dominance principles on which it depends,
117
so the greatest worry about the anti-supererogationist argument is that certain forms
of supererogationism will undermine those same principles—albeit more subtly than
radical skeptical theories like DNET.
One way this might happen is if the same arguments that support the existence
of first-order supererogation—in particular, arguments from demanding-ness—can
be simply re-run as objections to dominance principles like GDoT*: that is, GDoT*
should be rejected, just like maximizing utilitarianism, because the demands it places
on moral agents are excessive. But this line of argument is unconvincing for at least
two reasons. First, even if one thinks it plausible that principles of morality must
conform to a requirement of undemandingness, it is much harder to see why prin-
ciples of rationality like GDoT* should conform to such a requirement. Morality,
perhaps, is a human practice that must be designed to accommodate human weak-
ness and imperfection. But the canons of rationality derive from more abstract
standards of consistency or coherence that are not obviously sensitive to such prag-
matic concerns. Second and more importantly, appeals to arguments for first-order
supererogationism are question-begging, or something very much like it, when our
goal is to formulate principles for choice under conditions of first-order normative
uncertainty, where one is inter alia uncertain about the soundness of those very
arguments. That is to say, the conditional claim that a “second-order” theory of
rational choice under moral uncertainty should be rejected if it is highly demanding
is much less plausible than the conditional claim that a “first-order” moral theory
should be rejected if it is highly demanding, because the former conditional would
generate illicit “upward dependence” of principles for choice under moral uncertainty
118
on the very first-order moral questions we are uncertain about (cf. §3.6).
But there is another, more difficult way in which the undermining worry can
arise. To see it, we must introduce a further distinction among conceptions of
supererogation, at a somewhat higher level of generality than the distinctions dis-
cussed in the last section. Suppose we agree, as I have argued, that supereroga-
tion in most plausibly conceived in such a way that there is in some sense greater
all-things-considered objective reason to choose a supererogatory option O than a
merely permissible alternative P. It seems to me that there are then two very gen-
eral ways of making sense of supererogation, such that P can be understood as
all-things-considered rationally permissible at least by the lights of the supereroga-
tionist theory (that is, for an agent whose credence in the supererogationist theory
is 1).
First, suppose one accepts the following two intuitively appealing principles:
(i) an agent is always rationally required to do what she has most subjective reason to
do and (ii) an agent’s subjective reasons derive from her beliefs about her objective
reasons—more specifically, that belief in an objective reason of a given strength
gives rise to a subjective reason of proportionate strength. Then the most natural
way of understanding supererogation is to suppose that, although there is more
objective reason available to support the supererogatory option O, some of the
reasons that favor O are of a sort that the agent has the power to set aside or
eliminate: “justifying” or “enticing” reasons, imperfect reasons, or reasons that
come along with exclusionary permissions. That is, the agent is permitted to perform
option P, according to the supererogatory theory, because she has the power to alter
119
the initial balance of objective reasons by setting aside certain optional reasons,
thereby altering the balance of subjective reasons as well, in favor of P.16 Since
this approach is exemplified by the idea of exclusionary permissions, let’s call it the
“exclusionary conception of supererogation” (ECS).
The alternative, it seems to me, is to deny (i) and hold that supererogation
arises because, in a certain class of cases, agents are rationally permitted to act
against the balance of subjective reasons. That is, one is rationally permitted to
choose P even though one has more subjective reason to choose O. Since this ap-
proach is exemplified by understandings of supererogation in terms of a satisficing
conception of subjective rationality, let’s call it the “satisficing conception of su-
pererogation” (SCS).
Just as DNET posed an undermining threat to the dominance argument for
rejecting nihilism, so SCS poses an undermining threat to the dominance argument
for rejecting supererogationism.17 To make this threat concrete, suppose that I have
.6 credence in a version of SCS that implies the following principle.
Egoistic Permissions Principle (EPP) An agent is always rationally permitted
to choose the option that is prudentially best in expectation, out of those op-
tions in a given choice situation that violate no moral constraint (e.g. against
16Alternatively, one might deny the objective-subjective linkage principle (ii), and hold instead
that agents have the power, not to eliminate certain objective reasons, but rather simply to prevent
them from generating subjective reasons. As far as I can see, everything to be said below about
the view described in the main text will apply to this variant as well.
The idea that we can sometimes alter our reasons by an act of will is plausibilified by other sorts
of cases where this seems to happen: for instance, when an agent forms or abandons a plan or
life project, she may be thought of as creating or eliminating reasons for herself. Supererogatory
moral reasons, on this conception, might be thought of as reasons “out there” in the environment
that an agent can, but need not, adopt as her own.
17Thanks to Shelly Kagan for bringing this problem to my attention.
120
killing the innocent, breaking a promise, etc), even when there is greater all-
things-considered subjective reason for her to choose some other option.
Just as in §4.1.2 we saw that DNET ` ¬GDoT*, so here EPP ` ¬GDoT*.
For consider a situation in which I can decide to spend some money harmlessly on
myself or donate that money to charity, and in which I divide my beliefs between a
moral theory on which donating the money is obligatory and one on which donating
the money is supererogatory. In this situation, GDoT* implies that I am rationally
required to donate the money, while EPP implies that I am not. Just as with
DNET, then, it would seem that if I accept EPP with subjective probability p, I
must assign GDoT* a degree of belief less than or equal to (1 − p); that if p > .5,
then Cr(GDoT*) < .5, and that if I believe EPP, I must disbelieve GDoT*. So
again we may ask: How can I be subject to the rational requirements of a principle,
GDoT*, that I (justifiably, we may stipulate) regard as more likely false than true?
This undermining worry can be mitigated in several ways. First, of course,
there remains the option of taking a principle like GDoT*, or some stronger principle
from which it derives, to have belief-independent rational requiring force. Thus we
might say that, even if I justifiably believe that I am not rationally required to
perform the potentially-obligatory option O because I have greater than .5 credence
in a principle like EPP, I can nevertheless be so required.
But moreover, there is reason to think that the versions of SCS that undermine
dominance principles like GDoT* are uniquely implausible, indeed more implausible
than undermining forms of nihilism like DNET. It strikes me as a basic conceptual
121
truth that I am rationally required to do what I have most subjective reason to
do. To think otherwise is to multiply basic normative concepts beyond necessity,
allowing the concept of rational requirement to float free of the concept of subjec-
tive reasons. And as I suggested above, the sort of demandingness arguments that
might support a satisficing conception of rationality in the context of potentially-
supererogatory acts seems much more plausible in the domain of first-order norma-
tive ethics than in the domain of rationality.18 It is much more plausible, then, to
posit the existence of certain objective reasons that I have the power to set aside or
exclude from consideration such that they do not generate subjective reasons in the
first place than to hold that I can simply ignore the balance of subjective reasons I
do have.
Nevertheless, if we deny that a principle like GDoT* has belief-independent
normative force, then SCS does pose an undermining threat to the dominance ar-
gument for rejecting supererogationism, for an agent who assigns the majority of
her credence to this version of supererogationism. But what about ECS, the ri-
val conception of supererogation according to which an agent can alter the initial
balance of reasons so that she no longer has all-things-considered subjective reason
to choose the supererogatory option? This seems to me the most plausible way of
understanding supererogation, but if so, it raises an interesting complication for the
dominance argument against supererogationism. It seems plausible to suppose, on
this conception, that a fully rational agent exercises her power to alter the balance
18Another sort of argument for a satisficing conception of rationality invokes choice situations
in which there are infinitely many options, none of which is better than all the rest (Landesman,
1995). But it is hard to see how a satisficing view motivated by cases of this kind could support a
theory of supererogation.
122
of reasons in favor of the merely permissible option iff she actually chooses the
merely permissible option: that is to say, it seems at least irrational if not simply
inconceivable for an agent who has the power to alter the balance of reasons in favor
of a merely-permissible option to choose that option but fail to alter the balance
of reasons in its favor, or to so alter the balance of reasons (setting certain moral
reasons aside) but then choose the supererogatory option anyway. Indeed, by what
means could an agent alter the balance of reasons to favor the merely-permissible
option, other than by making a rational choice of that option?
If this is so, then ECS treats cases of supererogation as exhibiting a kind of
extreme act-state dependence: An agent A who is certain of ECS, and certain as
she chooses between options O and P that O is supererogatory and P is merely-
permissible, will be certain that whichever option she chooses will be the one favored
by the balance of reasons. If she is exactly .6 confident that she will choose P, then
she is exactly .6 confident that the balance of reasons will, in the end, turn out to
favor P.19 So, if A is not certain that she will choose O, then she is not certain that
the balance of reasons at the time she makes her choice will not favor P, and so it
looks as if dominance reasoning cannot serve to eliminate P from consideration.
But there is something intuitively wrong with this line of reasoning: For al-
though A is not certain that her reasons to choose O will turn out to be equal to or
greater than her reasons to choose P, she is certain that if she chooses O, then there
will be equal or greater reason to choose O—indeed, she is specifically certain that
19I assume that deliberation does not “crowd out prediction,” i.e. that an agent can have cre-
dences concerning what her own choice will be even as she deliberates about that choice. For
defense of this view, see for instance Joyce (2002).
123
there will be greater reason. If she is certain that her supererogationist theory is
right, then she has the same certainty about P, i.e. that conditional on her choosing
P, P is the option favored by the balance of reasons. But if she has some degree
of belief in a more rigorous moral theory according to which O is obligatory, i.e. on
which she lacks the power to alter the balance of reasons in favor of P, then she
can no longer have this certainty that if she chooses P, P will be the option most
supported by her reasons. This creates an asymmetry in favor of O that seems to
make O the most rational option.
And indeed, a slight reframing of the dominance principle GDoT* that ac-
counts for the problem of act-state dependence allows us to accommodate this in-
tuition. The fully precise way of expressing the idea of dominance reasoning over
moral theories, it seems to me, is as follows.
Final Dominance over Theories (FDoT) If (i) every theory in which an agent
A has positive credence implies that, conditional on her choosing option O, she
has at least as much reason to choose option O as to choose option P, (ii) one
or more theories in which she has positive credence imply that, conditional on
A’s choosing option O, she has greater reason to choose option O than option
P, and (iii) one or more theories imply that, conditional on A’s choosing option
P, she has greater reason to choose option O than option P, then A is rationally
prohibited from choosing P.
If a principle like FDoT that allows dominance reasoning to account for act-
state dependence is correct, then dominance reasoning will succeed against su-
124
pererogationist theories of the ECS variety. This is, of course, a heavily conditioned
conclusion. Nonetheless, FDoT is intuitively plausible (or, at least, formalizes an
intuitively plausible form of reasoning) and there is reason to think that ECS is
more plausible than rival versions of supererogation—both those that deny that
the balance of reasons, conditional on A’s choosing O, is favorable to O and those
according to which it is rationally permissible for an agent to disregard the bal-
ance of subjective reasons. Thus, even this limited conclusion may have substantial
practical scope.20
4.4 Practical Upshots
To the extent that the dominance argument for rejecting supererogationism
is effective, it sheds new light on some contested practical questions concerning the
ethics of philanthropy. Suppose you are a typical member of the middle class in the
developed world, and that on careful examination of the available evidence you judge
that donating to well-chosen charities can save a life in expectation at a cost on the
order of a few thousand dollars.21 The upshot of a dominance principle like FDoT
seems to be that even if you judge such charitable giving to be most likely morally
supererogatory, you are still rationally required to donate a substantial portion of
20Note that to avoid the undermining problem posed by SCS it is only required either that a
dominance principle like GDoT* or FDoT have external normative force or that the agent give the
SCS version of supererogationism less than half her credence. The versions of supererogationism
on which non-moral reasons are stronger than or incomparable with moral reasons pose a more
substantial threat to the dominance strategy, insofar as any positive credence in these theories will
block the application of GDoT* or FDoT. We will address this issue in the concluding section of
the chapter.
21Figures in this range are well-supported by recent research. See for example GiveWell’s esti-
mate of the cost-effectiveness of anti-malarial bed net distributions (GiveWell, 2013).
125
your income to charity. More specifically, as per FDoT, you are rationally required
to give at least to the point at which you judge that, by giving more, you might no
longer be acting for the best, all things considered. In contrast to Peter Singer’s
famous injunction to “give to the point of marginal utility” (Singer, 1972), we might
say that you are required to give (at least) “to the point of all-things-considered
uncertainty.” If you are not yet at this point, then you do not risk acting wrongly
(contrary to the overall balance of reasons) by giving another dollar to charity. But
there is a risk of acting wrongly if you keep the dollar, insofar as you judge that you
might be objectively obligated to donate it. You may believe, with great confidence,
that your donation is objectively supererogatory, but there is no (overall) harm done
by performing a supererogatory action on the grounds that it might be obligatory,
while there is (overall) harm done by refraining from an objectively obligatory action
on the grounds that it might be supererogatory. The latter is a risk to be avoided,
the former is not.
Giving to the point of all-things-considered uncertainty will generally be quite
a bit less demanding than giving to the point of marginal utility. But just where
is the point of overall uncertainty, when it comes to philanthropy? Here intuitions
will no doubt diverge. As I have already suggested, that point can be reached more
quickly when competing moral considerations are in play: moving your family into
a hovel in order to squeeze more deworming money out of your Wall Street paycheck
might well be an overall mistake. Of course, this risk of wrongdoing against one’s
family might be outweighed by the risks of not giving to the point of marginal utility,
but it might not—dominance reasoning alone cannot speak to this question. But
126
it should be readily conceded that there is a point, short of the point of marginal
utility, at which potentially countervailing obligations or non-moral reasons come
into play and the dominance argument can go no further. Still, common sense
suggests that few First Worlders are presently in any danger of stepping past this
point. Is there a significant risk that a typical American would violate a moral
obligation, or otherwise act against the overall balance of reasons, by deciding that
she and her family should live on $50,000 rather than $60,000 per year, if by this
sacrifice they could save two or three lives every year in expectation? Even here,
perhaps, intuitions may diverge. But I imagine no one will think that an agent
in this position who donated, say, $1000 a year to charity would be acting for the
worse all things considered. And so by the above reasoning, this much at least is
rationally required.
More clearly, perhaps, the dominance argument holds implications for where
one’s charitable efforts should be directed. Suppose that a billionaire philanthropist
every year makes a $1,000,000 donation to the local art museum. One year, an
admirable friend suggests to her that this money would be better spent supporting
some life-saving health initiative in the developing world. Here, I don’t think rea-
sonable intuitions can diverge. It seems simply obvious that, between the options of
donating a million dollars to the art museum and donating a million dollars to the
health initiative, there is more all-things-considered reason to choose the latter, and
the former option can be rationally permissible only if the latter is supererogatory.
Put another way, it is highly implausible that our imagined agent would violate a
moral obligation or otherwise act for the worse by choosing to support the health
127
initiative instead of the museum. In so doing, she might deprive the local museum-
going public of the chance to feign enjoyment of one more Jackson Pollock canvas,
but it is quite hard to believe that her obligations in this respect outweigh her
obligations to the global poor.
Here, then, is a straightforward practical conclusion: When an agent is confi-
dent that it is all-things-considered permissible for her to devote certain resources
(time, money, etc) to philanthropic endeavors, and has no reason to think herself
under any other obligation to support some specific philanthropic endeavor (e.g. by
having promised the money to the art museum), the possible obligation to support
efficient endeavors—those that do the most good per dollar spent—is enough to
generate a rational requirement that she choose these efficient endeavors over their
less efficient alternatives.
Rejecting supererogationism may have practical consequences in other do-
mains as well. For instance, an agent might be fully convinced that non-human
animals have enough moral standing that vegetarianism is at least as choiceworthy,
all things considered, as meat consumption, and yet be unsure whether vegetari-
anism is morally required or supererogatory. Likewise, a citizen of a democratic
society might be certain that she has at least as much overall reason to vote as to
stay home on election day, and yet be unsure whether voting is morally required
or supererogatory. For such agents, the argument I have been defending implies
that they are rationally required to choose the morally safe option (vegetarianism
or voting, respectively).
A common theme among the consequences of rejecting supererogationism—
128
favoring altruism and philanthropy in general, efficient philanthropy in particular,
and increased regard for potential moral subjects like non-human animals—is their
broadly utilitarian character. The effect of rejecting supererogationism is largely
pro-utilitarian, I suspect, because utilitarian theories are the only plausible moral
theories that are totalizing, in the sense of rendering every or nearly every choice a
matter of moral obligation. Thus, wherever a potentially obligatory action fails to
maximize utility, it is potentially forbidden, i.e., there is a possible contrary objective
obligation to counterbalance it (even if this possible obligation is outweighed in the
last analysis)—so for any agent who has positive credence in classical utilitarianism,
dominance reasoning cannot imply a rational requirement to choose any non-utility-
maximizing course of action. By contrast, we often face choices where the only
potential moral obligation one way or the other is that implied by a maximizing
utilitarian theory, and in these cases dominance reasoning suggests that we are
often rationally required to meet the utilitarian demand.
4.5 Can Dominance Reasoning Stand on Its Own?
There is a final, very general problem for dominance approaches that we have
so far put off, and that serves to motivate the more ambitious project of the next
two chapters. The problem is this: Any form of dominance reasoning requires a
kind of certainty, namely, certainty that some dominant option is at least as good
as a dominated option. But on the Bayesian understanding of probabilistic belief
we have so far assumed, which accepts regularity, such certainties are disallowed
129
except for some very narrow subclass of a priori truths (e.g. logical or analytic). No
matter how confident we may be of the judgment that it is at least as good to give
to charity as to spend money on oneself, or that it is at least as good not to torture
babies for fun as to torture babies for fun, it is not generally believed that these are
logical, analytic, or indeed even a priori truths.22
Of course, one could defend dominance reasoning by disputing the skeptical
Bayesian attitude toward certainty. Many, no doubt, will be inclined to say it is
certain that it’s at least as good not to torture babies for fun as to torture babies
for fun. But it seems to me that the Bayesian view is correct in this regard, though
it is well beyond the scope of my current project to defend it. To put the point as
intuitively as possible, we should remain open to the possibility that our ordinary
beliefs are radically wrong, even diametrically opposed to the truth, in the moral
domain as in the empirical domain. Even if, as some philosophers have held, it is not
conceptually possible that causing pain to innocent people is good for its own sake
or that knowingly punishing the innocent is just, the difficulties and uncertainties of
conceptual analysis mean that these things remain both epistemically and logically
22This draws attention to another point we have passed over, though I believe justifiably. The
mundane reason why it cannot be true a priori that, say, giving to GiveDirectly is better than
buying yourself a new TV is that the moral comparison of these options depends crucially on their
consequences, which are of course not an a priori matter. We have been assuming thus far that, in
the sort of dominance reasoning we are interested in, these empirical uncertainties are suppressed.
That is, the dominance principles we have considered say roughly that if each moral theory to
which you give positive credence implies that your empirical-belief-relative reasons either favor
option O over option P or are neutral between the two, then you should prefer option O. It seems
permissible to reason in this way, rather than attempt the fruitless task of applying dominance
principles over all maximal conjunctions of moral and non-moral possibilities, because particular
theories will have their own internal methods for handling empirical uncertainty which are often
much stronger than dominance reasoning. Thus, from the standpoint of an agent deliberating
about what to do given her moral uncertainties, she can simply take it as input, for instance,
that utilitarianism assigns greater expected utility to option O than to option P, suppressing any
further consideration of the empirical uncertainties that go into the utilitarian assessment. (I will
defend this sort of “multi-stage” reasoning under moral uncertainty at greater length in §6.2.)
130
possible. Therefore I think we should accept, regretfully, that it is never strictly
certain that one option is at least as good as another.
Does this mean that dominance reasoning is useless (indeed, not just with re-
spect to moral uncertainty, but in its more common applications to cases of empirical
uncertainty as well)? No, but it does mean that the practical force of dominance
arguments is derivative rather than fundamental. That is to say, in cases where we
should act on dominance reasoning, it is because it approximates some other form of
reasoning, e.g. expected value reasoning, in the limiting case where any possibility
according to which option P is better than option O is seen as highly improbable.
The more unlikely it is that spending money on luxuries for oneself is better overall
than donating the money to charity, the more unlikely it is that the correct compre-
hensive theory of decision-making under uncertainty will regard the former option as
rationally required or permissible. Dominance reasoning is useful, therefore, in that
it identifies robust practical conclusions, conclusions in which we can have a great
deal of confidence even while we remain deeply uncertain about the fundamental
principles of rational choice and about what they will imply for more difficult cases.
More specifically, for our purposes, we can say this: So long as some hedging view
of moral uncertainty turns out to be correct (or more broadly, some view consis-
tent with SSR), it will almost certainly imply that we are rationally required to do
things like donate a substantial portion of our incomes to highly effective charities,
and that having high credence in at least the ordinary forms of moral skepticism
and nihilism do not make it rationally permissible to act like an amoralist.23
23In conversation, Samuel Kerstein has suggested that many agents have substantial credence
131
In the next chapter we will begin to develop such a comprehensive theory
of rational choice under moral uncertainty, that attempts to specify precisely the
contributions that an agent’s moral beliefs and credences make toward determining
what she subjectively ought to do. In so doing, we will allow ourselves to say some-
thing about the more difficult cases where plausible moral views conflict, prescribing
incompatible courses of action as morally required.
in views like ethical egoism according to which any philanthropic self-sacrifice is worse overall
that self-interested alternatives, and that for this reason such agents will never be close enough to
certainty that an option O is either obligatory or supererogatory for dominance reasoning to serve
as a plausible stand-in for more general principles of rational choice. I accept this possibility in
principle, though it seems to me that many or most agents have very little credence in any form
of ethical egoism and that such a credal state is at least rationally justifiable. But it is also worth
noting that an agent’s credence in such views should not problematize the conclusions drawn in the
preceding section about where she should direct any philanthropic giving she chooses to undertake.
For an agent choosing between similarly altruistic options (i.e., options that involve similar degrees
of self-sacrifice) like donating $1 million to the art museum or donating $1 million to global health
charities, egoistic theories will agree that the latter option is at least as choiceworthy as the former.
Still, whether and how often we are rationally required to choose potentially supererogatory
options will depend on the more general theory of rational choice under moral uncertainty. Even
if dominance reasoning is often not an acceptably close approximation of this theory, limiting the
scope of the argument I have made in this chapter to a small number of cases, it may still turn
out that we are rationally required to choose a great many potentially-supererogatory options in
the last analysis, since the potential all-things-considered downside of such actions (e.g., if ethical
egoism is true) is outweighed by the potential upside (if a non-egoistic theory is true).
132
Chapter 5: Intertheoretic Value Aggregation
As we have seen, drawing interesting normative conclusions about choice under
moral uncertainty requires that we move beyond dominance principles and offer a
general theory that accounts for cases of conflict between epistemically possible
moral principles. In this chapter, we will begin the task of describing such a general
theory. In the process, we will be able to address two more objections to moral
hedging that we have thus far put off: the problem of intertheoretic value aggregation
(PIVA) and the grounding problem. I will argue that both these objections can
be overcome if we can locate the basis for a particular aggregation of the value
scales of rival theories, not in some universal, freestanding second-order principle like
Lockhart’s PEMT, but rather in the shared structure and content of limited classes
of epistemically possible moral theories, what I will call comparability classes.1
5.1 Two More Challenges to Hedging
We have already encountered the problem of intertheoretic value aggregation
in earlier chapters, but let’s now take the time to describe it in a bit more detail, and
to make clear its intuitive force. In the process, we will see that it in fact represents
1Much of this chapter, particularly §§5.1, 5.3, and 5.4, is based on Tarsney (forthcoming a).
133
two problems, one epistemic and the other metaphysical, each of which must be
addressed by any plausible account of intertheoretic aggregation.2
Suppose, then, that an agent who divides her beliefs evenly between Kantian-
ism and utilitarianism (Cr(Kantianism) = .5, Cr(utilitarianism) = .5) is faced with
a moral dilemma about which these theories disagree: Should she push a large man
off a footbridge so that he falls in the path of an out-of-control trolley, stopping it
before it can hit and kill some number n > 1 of innocent people? Or should she do
nothing, letting the large man live and the n others die (Thomson, 1976)?
If one accepts the core principle of moral hedging, that an agent’s subjective
reasons for action under moral uncertainty are sensitive both to her degrees of belief
in rival moral theories and to the strength of the objective reasons those theories
posit, then it seems one will judge that what the agent in this case subjectively
ought to do depends on the value of n. Perhaps if killing the large man will save
only two others from the trolley (or only one other whose life happens to carry
slightly more expected utility, because he is younger, or healthier, or of a more
cheerful disposition, or valued by a slightly larger circle of friends and family), the
epistemically possible Kantian prohibition on killing innocents as a means to an end
2What I will call the “problem of intertheoretic value aggregation,” the literature to date has
typically called the “problem of intertheoretic value comparisons.” I choose the non-standard
terminology because, as we will see in the next chapter (specifically §6.4), even when we are able
to compare the value scales of two or more moral theories, a further question remains about how
to aggregate those value scales into a single scale for purposes of rational choice. Thus, in my
terminology, the problem of intertheoretic value comparisons is a part, albeit a very large part, of
the broader problem of intertheoretic value aggregation.
The next chapter will also introduce another division between three kinds of barriers to in-
tertheoretic value aggregation (two of which have to do with intertheoretic comparisons while the
third has to do with the further problem of aggregation). This division is orthogonal to the epis-
temic/metaphysical distinction introduced in the current section, i.e., each of the three challenges
to be described in the next chapter will have both an epistemic and a metaphysical dimension.
134
ought to win out over the equally possible but relatively weak utilitarian reasons for
killing. On the other hand, if the trolley is about to crash headlong into a reactor
of the local nuclear power plant, triggering a catastrophic meltdown that will cause
10,000 nearby residents to die agonizing deaths from radiation sickness, then perhaps
the now-much-strengthened utilitarian reasons should take precedence.
So far so plausible. But what about the intermediate cases between these
extremes? A commitment to hedging seems to entail that there is some value of n
that is the tipping point, or the point of indifference, between Kantian and utilitarian
considerations in this case as we have described it. Perhaps if pushing the large
man will result in 27.38 or more net lives saved in expectation, then our agent
subjectively ought to choose the utilitarian option, but if the expected number of
lives saved is any less than this, she ought to choose the Kantian option.3 This value
of n, whatever it might be, represents the ratio between the wrongness of killing an
innocent person as a means to an end, according to Kantianism, and the rightness
of saving a (net) life, according to utilitarianism.
There is something unpalatable about the idea that such a ratio exists. In fact,
there are at least two things unpalatable about it: (1) What could possibly make
it the case that any particular value of n is the true tipping point between Kantian
3Of course, we have to fill in details before such a conclusion could be entailed, in particular
the precise quantities of utility and disutility that would follow from each death caused or averted.
As may already be clear, the general principle of moral hedging as I have described it does not
fully commit us to the existence of such a sharp cutoff; e.g., we might believe in hedging while also
believing that Kantian and utilitarian reasons are only roughly comparable and hence that there is
no definite point of indifference between them in the sort of case we have described. Nevertheless,
it seems that even such a view will commit us to sharp boundaries of some sort, e.g., between
the values of n for which the Kantian or utilitarian reasons are determinately stronger and the
intermediate interval of incommensurability (Broome, 1997, pp. 85-86).
135
and utilitarian considerations, in the above case as we have described it? What facts
about the (empirical or moral) universe could serve to ground the existence of such
a ratio? (2) Assuming that there is such a ratio, how could we possibly discover
it? How is a rational agent to figure out, for instance, whether the right tipping
point when she is faced with such a choice is 3, or 5, or 27.38, or 12pi, or ...? And if
there is no way of figuring it out other than to take a plausible-sounding guess, then
what becomes of the point that we need norms of rational requirement responsive
to an agent’s moral uncertainties precisely because we as agents need action-guiding
norms whose prescriptions are epistemically accessible to us? If a commitment to
hedging forces us to simply take a shot in the dark with respect to the exchange
rate between Kantian and utilitarian values, what has been gained over the simpler
option of taking a shot in the dark at which is the true moral theory and being
guided by whichever theory we regard as more probable?
It is useful, I think, to consider these as separate problems. I will call the
first, metaphysical worry the grounding problem, reserving problem of intertheoretic
value aggregation for the second, epistemological problem. The former problem has
been suggested by several authors in the literature, but to the best of my knowledge
has not been explicitly distinguished from the epistemological problem.4 But it is,
4I take this worry about grounding to be, for example, at least one of the lines of argument
suggested in the following passage from Hudson (1989): “Even mere axiological uncertainty within
an unquestioned subjective consequentialist framework is unhedgeable. Suppose the agent assigns
probability 0.6 to the view that pleasure-minus-pain is the only intrinsic good, and 0.4 to the view
that the good is self-realization. And suppose she must choose between an act that produces ten
hedons and two reals and one that produces nine hedons and thirty reals. (‘Reals’ are the units in
which self-realization is measured.) Which act should she do? The two axiological theories lead
to different answers. Since the hedonic theory is more probable, perhaps she should accept its
answer. But the self-realization theory seems to find more of a difference between the two actions,
and perhaps this should outweigh its slightly lesser probability. But wait—is a difference of twenty-
eight reals really greater than a difference of one hedon? What is the common measure between
136
in my view, at least as urgent as the epistemological problem, and it cannot be
assumed that a solution to the latter will also be a solution to the former. Ross
(2006), for example, suggests that we can find out how to compare the value scales
of rival theories by reasoning backwards from our considered judgments about the
subjective rationality of particular choices given particular credal states in imagined
cases like the footbridge dilemma described above. But even if one regards our
judgments about such cases as prima facie trustworthy, such an approach leaves
it mysterious how or in virtue of what the underlying value comparisons that we
suss out of those case judgments could be correct. Of course, we always have the
option of asserting it as a brute fact that, say, the correct ratio between the Kantian
wrongness of killing and the utilitarian rightness of saving lives is 27.38 (or, more
plausibly, that there is some such precise ratio that is somewhere in such-and-such
range). But it would be strange if facts like these were simply brute, part of the
fabric of the universe on a par with fundamental physical constants. Why should the
hedons and reals? Note that the agent, for all her uncertainty, believes with complete confidence
that there is no common measure: she is sure that one or the other—pleasure or self-realization—is
intrinsically worthless. Under the circumstances, the two units must be incomparable by the agent,
and so there can be no way for her uncertainty to be taken into account in a reasonable decision
procedure.” (Hudson, 1989, p. 224; emphasis added)
Likewise, Gracely writes: “Consider again the extra-marital sex example. It certaintly seems
possible to compare the fact that one system decalres the act to be a terrible wrong, while the other
merely regards it as tolerable, or, at most, slightly more right than wrong. I don’t agree. ‘Terrible
wrong’ is only meaningful within the overall framework of that system. There is no abstract scale
of ‘wrongness’ outside of the rank provided within a theory.
“Moreover, if indeed there is only one true moral theory, the noncomparabilty argument is
further strengthened. In that case, a metatheory that attempted to design a weighting scheme
applicable across systems would necessarily be comparing many false, or partly false, systems to
each other and/or to the one that is true. Why would one expect there to be a valid method of
comparing magnitudes other than that provided by the one true theory?” (Gracely, 1996, p. 331)
Finally, Nissan-Rozen (2015) raises a more specific version of the grounding objection to hedging,
premised on the claim that a theory’s “moral value function” is simply a representation of its
ranking of lotteries (i.e., choices made under various states of empirical uncertainty), and that
since these representations are unique only up to affine transformation (meaning that the choice of
scale and zero point are arbitrary), there is no non-arbitrary or meaningful basis to compare the
value function of one theory to that of another. (I address this argument in Appendix B, §B.2.)
137
universe have been so obliging as to provide us with the presumably thousands of
such constants needed to compare the plethora of possible moral theories to which a
reasonable agent might subscribe, especially given that all but one of those theories
are in fact false?
Having described PIVA and the grounding problem somewhat impressionisti-
cally, let’s make our terminology a bit more precise.
• An intertheoretic value aggregation is a mapping from probability distributions
over value assignments to value assignments. The output of this mapping is
meant to represent, in one way or another, a probability-weighted aggregate
of theories in which an agent has positive credence.
• The grounding problem is the problem of what could make a particular in-
tertheoretic value aggregation objectively correct, i.e., correct in a way not
relative to any agent’s beliefs/evidence.
• The problem of intertheoretic value aggregation is the problem of how to dis-
cover the objectively correct intertheoretic value aggregation and, relatedly,
how an agent should decide what intertheoretic value aggregation to adopt,
given her beliefs/evidence.5
Both the grounding problem and the problem of intertheoretic value aggrega-
tion have been put forward repeatedly in the literature as powerful objections to
5I group these problems together since I assume that an agent’s decision about how to aggregate
the value assignments of the various moral theories in which she has positive credence will be guided
by a concern for truth, i.e., she will want as far as possible to adopt the correct intertheoretic value
aggregation. In Chapter 7 we will consider how an agent ought to proceed when she cannot be
sure what aggregation procedure is correct.
138
moral hedging, with particular emphasis being placed on one or both of these prob-
lems by Hudson (1989), Gracely (1996), Gustafsson and Torpman (2014), Nissan-
Rozen (2015), and Hedden (2016).
The aim of this chapter is to begin addressing both of these problems. The
strategy is as follows: I start by exploring two simple case of intertheoretic moral
uncertainty that appear to offer rationally perspicuous, non-arbitrary methods for
intertheoretic value aggregation. In one of these cases, the natural aggregation
procedure involves both quantitative intertheoretic value comparisons and moral
hedging. In the other case, it involves neither. Reflection on these cases suggests
the more general idea that there are clusters of moral theories, which I will call
comparability classes, that share common principles strong enough to establish clear
internal procedures for value aggregation and choice under moral uncertainty, which
sometimes will and sometimes will not take the form of hedging procedures like
expected value maximization.
If this is so, I argue, then neither the grounding problem nor PIVA can be
taken as general objections to moral hedging as such. That is, even if it ultimately
turns out that some moral uncertainties (like that between classical utilitarianism
and absolutist deontology) are too deep to permit the sort of comparison and ag-
gregation necessary for hedging, or indeed any form of aggregation at all, this fact
cannot plausibly be taken as a reason to forgo hedging with respect to those un-
certainties that do permit non-arbitrary intertheoretic comparisons. There is no
reason, in other words, to think that “incomparability anywhere is a threat to com-
parability everywhere.” Instead, I will argue, if there is a line to be drawn between
139
uncertainties for which a hedging procedure is appropriate and those for which it
is not, that line should be drawn not between moral and empirical uncertainty but
rather between those uncertainties that establish barriers to the necessary interthe-
oretic comparisons and those that do not. In the last section of the chapter, I will
contrast this “content-based” approach to intertheoretic value aggregation with the
“top-down” approaches that have dominated the literature to date, and offer some
general arguments in favor of a content-based approach. Chapter 6 will then develop
this approach in considerably more detail.
5.2 Moral and Empirical Uncertainty, Revisited
A theme of the examples of intertheoretic comparability that we will explore in
the next section is that the decisions principles offered by a given class of moral the-
ories for dealing with morally relevant empirical uncertainty can often be extended,
in a natural and rationally perspicuous way, to structurally similar cases of moral
uncertainty. We have already discussed some of the parallels between empirical and
moral uncertainty in Chapter 2. But to prepare the ground for the arguments of
the next section, it will be useful to bring out a slightly different point, namely
the continuity between empirical and moral uncertainty—that is, the difficulty of
drawing a clear and non-arbitrary line between the two, of the sort that is needed
if we are to judge, with the opponents of hedging, that the principles of rationality
treat the two sorts of uncertainty in fundamentally different ways (demanding, e.g.,
that a utilitarian hedge for her uncertainties about the state of the empirical world,
140
but simply take a shot in the dark with respect to her uncertainties about the moral
world when she divides her beliefs between several versions of utilitarianism).
The best way to illustrate this continuity is to simply enumerate a few of
the cases in which the boundary between empirical and moral questions, and hence
between empirical and moral uncertainty, is blurry, and which therefore make it
implausible that entirely different sorts of principles should govern the two domains.
Here, then, are several such liminal cases.
• Certain cases of uncertainty about moral considerability (or moral status more
generally) turn on metaphysical uncertainties that resist easy classification as
empirical or moral.
– In the abortion debate, uncertainty about when in the course of develop-
ment the fetus/infant comes to count as a person is neither straightfor-
wardly empirical nor straightforwardly moral. Likewise for uncertainty
in Catholic moral theology about the time of ensoulment, the moment
between conception and birth at which God endows the fetus with a
human soul (though this is perhaps closer to empirical than to moral
uncertainty—after all, the soul is concrete even though it is not phys-
ical). Nevertheless, it seems strange to regard these uncertainties as
fundamentally different from more clearly empirical uncertainties about
the moral status of the developing fetus (e.g., uncertainty about where
in the gestation process complex mental activity, self-awareness, or the
capacity to experience pain first emerge), or from more clearly moral un-
141
certainties (e.g., uncertainty, given a certainty that the fetus is a person,
whether it is permissible to cause the death of such a person when doing
so will result in more total happiness and less total suffering).
– Many ethical questions (e.g. about the weight of promises or of indi-
vidual moral desert) turn in part of questions about the metaphysics of
personal identity (Parfit, 1971, 1973). As Moller (2011) points out, one
influential argument for the impermissibility of abortion (the “future like
ours” argument given in Marquis (1989)) seems much more likely to suc-
ceed if one accepts an animalist rather than a psychological theory of
personal identity. Uncertainty about the permissibility of abortion, or
about whether one is still morally required to adhere to some long-ago
childhood promise, that traces back to uncertainty about the criteria for
personal identity is hard to describe as “empirical” uncertainty (after all,
what empirical observation could resolve the debates between animalists
and psychological theorists?), but neither is it uncertainty about basic
moral principles or moral theory.
– Finally, on the theme of metaphysical uncertainty, a consequentialist (or
anyone who recognizes the ethical relevance of consequences) who is also a
physicalist about mental states may be uncertain about the moral status
of various kinds of entities because she is uncertain what sorts of physical
states subvene on morally significant mental states like pleasure/pain,
preference, etc. Does a certain artificially intelligent computer program
142
have preferences? Does a Martian whose pain-like functional states are
realized by D-valves rather than C-fibers feel pain? Are various collec-
tivities like corporations, nations, bee colonies, or mobs of human beings
with walkie talkies acting out philosophical thought experiments capable
of forming desires, pursuing projects, willing ends/maxims, etc? All of
these cases likewise blur the line between empirical and moral uncertainty.
• I have already alluded to the possibility of uncertainty about “thick” moral
properties, e.g., uncertainty whether some course of action would be cruel or
cowardly. Weatherson, as we saw, is ready to admit that hedging for such thick
moral uncertainty may sometimes be appropriate. If so, this cannot be because
thick moral uncertainty is just empirical uncertainty is disguise: I might know
everything there is to know about my beliefs and motivations in taking a
certain action and still wonder whether I acted cruelly. If we admit this sort
of uncertainty as hedge-able, then it will be even more difficult to draw a line
around the “purely moral” uncertainties we regard as un-hedge-able, since
the thickness of moral properties varies more-or-less continuously from cruel
to rights-violating to non-optimific (each of which combines evaluative with
empirical content).
• Certain views imply that first-order moral principles depend either epistem-
ically or metaphysically on empirical or quasi-empirical facts. For instance,
suppose one is convinced that divine command theory is the right account of
the source of moral principles, but uncertain what God has commanded, e.g.,
143
because one is uncertain whether the Bible or the Quran contains the accurate
reporting of God’s commands. Here again it is hard to say on which side of a
proposed empirical/moral divide one’s uncertainty should fall.
• Some ethical views, most obviously rule consequentialism, imply that first-
order moral principles depend on empirical counterfactuals: to a rule conse-
quentialist, the true set of moral principles is that which would yield the most
utility if generally practiced. If I am certain that rule consequentialism is
true, but uncertain whether a rule that permits authorities to “punish” the
innocent for the sake of deterrence is part of the optimific system of rules, is
my uncertainty empirical or moral? This seems to depend on whether one
understands rule consequentialism as itself constituting a first-order moral
theory (that says “follow the set of principles that would, if generally prac-
ticed, maximize utility”), in which case one’s uncertainty is empirical, or as
constituting a metaethical theory of what determines or how we are to iden-
tify the true first-order moral theory (the content of which is, de re, the set
of optimific-if-generally-practiced moral principles). But should anything of
practical significance turn on such a question of taxonomy?
• Most generally, naturalistic metaethical views that treat normative ethical the-
orizing as continuous with natural science will see first-order moral principles
as at least epistemically if not metaphysically dependent on features of the
empirical world. For instance, on Railton’s (1986) view, moral value attaches
(roughly) to social conditions that are stable with respect to certain kinds of
144
feedback mechanisms (like the protest of those who object to their treatment
under existing social conditions). What sort(s) of social conditions exhibit this
stability, given the relevant background facts about human psychology, is an
empirical question. For instance, is a social arrangement in which parents can
pass down large advantages to their offspring through inheritance, education,
etc, more stable or less stable than one in which the state intervenes extensively
to prevent such intergenerational perpetuation of advantage? Someone who
accepts a Railtonian metaethic and is therefore uncertain about the first-order
normative principles that govern such problems of distributive justice, though
on essentially empirical grounds, seems to occupy another sort of liminal space
between empirical and moral uncertainty.
In all these cases, it is implausible that important practical questions should
turn on whether we classify these uncertainties as “empirical” (and hence, very
likely, subject to some form of hedging if not to outright expectational reasoning) or
as “moral” (and hence immune from such reasoning). Conversely, it seems plausible
that the background principles one accepts for decision-making under empirical
uncertainty should have no great difficulty guiding one’s actions under these forms
of uncertainty, if one allows them to do so. Attention to these intermediate cases
highlights the fact that the moral principles we accept will themselves often tell us
how to proceed under at least certain limited forms of moral uncertainty.6
6Contra Hudson, inter alia, who writes: “Any moral theory, in telling the agent what to do, will
ignore the agent’s possible commitment to other moral theories. And while it should be considered
a defect in a theory that it issues instructions that the agent does not know how to follow, it is
not similarly a defect if it issues instructions that the agent decides not to follow because she does
not believe in the theory.” (Hudson, 1989, p. 224)
145
For instance, suppose you are a hedonistic consequentialist and a physicalist
about mental states, and that because you divide your belief among several versions
of physicalism, you are uncertain whether some class C of beings (fish, insects,
AIs, Martians, bee hives, corporations...) experience pain, even though you have no
relevant uncertainties about the chemical composition or functional organization of
those beings or any other relevant “narrowly empirical” uncertainty (uncertainty
about how exactly the atoms are moving around in the void). You have, say, a
credence of .6 in a type-identity theory according to which beings of class C do not
experience pain, and a credence of .4 in a functionalist theory according to which
they do. Though your uncertainty is not narrowly empirical, nevertheless it seems
natural to treat it just as you would treat empirical uncertainty (i.e., just as you
would treat a similar uncertainty about the moral considerability that was narrowly
empirical, e.g., uncertainty whether rabbits experience pain that reduces entirely to
uncertainty whether rabbits have C-fibers): You should weight the possible suffering
of beings of type C by a factor of .4, reflecting the probability that they experience
pain, when calculating the expected utility of your actions.
5.3 Case Studies in Comparability
In this section, we will consider two cases in which the same sort of thing
can be said of clearly moral uncertainties as of the liminal uncertainties considered
above: namely, that decision procedures for empirical uncertainty can and should
be straightforwardly extended to deal with these non-empirical uncertainties.
146
5.3.1 First Case Study: Pluralistic Consequentialism
Consider, then, an agent Alice whose moral beliefs are as follows: (i) She is
certain that consequentialism is true. (ii) She is certain that pleasure and pain
have non-derivative value and disvalue, respectively. (iii) She is uncertain whether
aesthetic goods (beauty etc) have non-derivative value as well. (iv) She is certain
that nothing besides hedonic and aesthetic goods has non-derivative value. (v)
She is certain that, when faced with purely empirical uncertainties, she ought to do
what maximizes expected value. She therefore divides her beliefs between two moral
theories: a monistic consequentialist theory T1 that values only hedonic goods and a
pluralistic theory T2 that values both hedonic and aesthetic goods. The latter theory,
we will stipulate, treats hedonic goods and aesthetic goods as comparable—that is,
it establishes some “exchange rate” between these two categories of value-bearer.
Since we will have reason to refer to Alice’s case several times in this chapter and
the next, let’s give it a name, and call it the Easy Case.
It seems to me that, in the Easy Case, Alice should experience no great dif-
ficulty in hedging for her moral uncertainties. We may stipulate an arbitrary unit
of hedonic value, the hedon, and may then define a unit of aesthetic value, the aes-
theton, as the quantity of aesthetic value that according to T2 is equal in value to a
single hedon. Now suppose Alice assigns T1 probability p and T2 probability (1−p).
In this state, I propose, she can quite naturally calculate the expected value of any
practical option O as follows: EV(O) = hedons(O) + (1− p)(aesthetons(O)).
Prima facie, it seems quite plausible that an agent in Alice’s doxastic situation
147
rationally ought to choose the option that maximizes this formula. The two theories
in which Alice has positive credence agree on the value of hedonic goods—take it
as a stipulation, to which we will return shortly, that T1 and T2 differ not at all in
the grounds they offer for, or anything else they have to say about, the intrinsic
value of a hedon. As a result T2 contains within itself, in the exchange rate it posits
between hedonic and aesthetic goods, a procedure for comparing its own evaluation
of practical options to those offered by T1, which are exclusively hedonic.
7,8
The suggestion from this simple case can be easily generalized. Let an agent
distribute her beliefs over any number of pluralistic theories, along with her monistic
theory, so long as all those theories agree on the value of hedonic goods. Then we
can simply create a unit of value for each theory equal to the quantity of value
that it assigns to a single hedon (an empirically measurable unit of pleasure/pain,
whose value, we have by stipulation every reason to believe, remains constant across
theories), thereby normalizing the value scales of the various competing theories.
Then we can simply compute the expected value of an option O across theories
T1, T2, ..., Tn by something very much like the standard formula for expected value:
7For now, I leave these claims at the level of plausible intuition. Appendix B gives an extended
explication and defense of the sort of reasoning employed in the last two paragraphs, making use
of certain ideas to be introduced in Chapter 6.
8The thought that moral certainties, e.g. of the value of hedonic goods, can be used to con-
struct a scale by which uncertain moral values and disvalues can be quantified for purposes of
moral hedging is suggested by Sepielli (2009), who refers to this scale as a “background ranking”
and shows that background rankings can be used to construct reasonable hedging procedures for
limited forms of moral uncertainty, e.g. concerning the permissibility of abortion. Sepielli, however,
attempts to infer a shared background ranking between a pair of theories simply from similarities
in their value assignments (specifically, identical ratios between the differences in choiceworthiness
assigned to some triple of options), which can generate multiple, inconsistent normalizations of
the same pair of theories (MacAskill, 2014, pp. 142-5). Ross (2006) has suggested that some the-
ories can be compared in the way that the hedonism/pluralism case illustrates, based on deeper
agreement with respect to certain forms of value or kinds of reasons, applying this observation to
a case where two theories disagree about the correct theory of rights but are in perfect agreement
concerning all other moral considerations (Ross, 2006, pp. 764-5).
148
EV(O) =
n∑
i=1
(EV(O)|Ti)(Cr(Ti))
This formula will be applicable even to an agent who distributes her belief over
infinitely many competing theories, corresponding for instance to all possible real-
valued exchange rates between hedonic and other forms of value over some bounded
interval. In fact, we can generalize even further, dropping the assumption that
there is a category of value-bearer common to every theory in which the agent has
positive credence. Suppose, for instance, that an agent regards hedonic experience,
beauty, and knowledge all as potential non-derivative value-bearers, and that her
credences regarding the value of each of these goods are mutually independent, so
that she has some credence in theories according to which all, none, or any other
combination of the three are non-derivative goods. Comparisons between these
goods can nevertheless be made by appealing to the exchange rates given by the
various pluralistic theories that value more than one of the goods in question, so
long as these exchange rates are consistent across the relevant pluralistic theories.
Thus, for instance, the monistic theory according to which only pleasure is a non-
derivative good and the monistic theory according to which only knowledge is a
non-derivative good are made comparable by the agent’s credence in one or more
theories according to which both are non-derivative goods.
What I have said so far involves an important presupposition, namely that
when two theories agree that some feature of the world is a non-derivative bearer
of moral value, those two theories attribute the same kind and same degree of
149
value to that phenomenon—e.g., that the value of a hedon according to Alice’s
hedonistic theory is equal to the value of a hedon according to the pluralistic theory
that also values aesthetic goods. Clearly, this need not always be the case. For
instance, consider an agent Alex who, like Alice, divides his moral belief between
two theories, a hedonistic and a pluralistic version of consequentialism. But suppose
that Alex also divides his metaethical beliefs between a robust moral realism and a
fairly anemic anti-realism, and that his credence in hedonistic consequentialism is
mostly or entirely conditioned on his credence in robust realism while his credence in
pluralism is mostly or entirely conditioned on his credence in anti-realism. (Suppose
he inclines toward a hedonistic view on which certain qualia have intrinsic value or
disvalue entirely independent of our beliefs, attitudes, etc, which we are morally
required to maximize. But if this view turns out to be wrong, he believes, then
morality can only consist in the pursuit of whatever we contingently happen to
value in some distinctively moral way, which includes pleasure but also knowledge,
aesthetic goods, friendship, etc.) In this case, the sort of procedure that I have
been describing no longer seems rationally compelling: Although the various moral
theories to which Alex assigns positive credence all treat pleasure as a bearer of
non-derivative moral value, there is no obvious reason to think that they assign the
same degree of value to a given unit of pleasure.
But there is no reason to assume that such problems will always arise. An agent
may be uncertain about first-order moral questions, like whether anything besides
pleasure and pain has non-derivative value, without this reflecting any underlying
metaethical uncertainties. An agent who divides her beliefs between various monistic
150
and pluralistic theories might nevertheless be in no doubt as to the nature, basis,
or degree of value possessed by some category of goods, like hedonic goods, that
all the theories she entertains recognize as non-derivatively valuable.9 The lack of
any uncertainty concerning hedonic value makes it a constant feature of the various
theories in which she has positive credence, and allows it to serve as a basis for
normalization.10
To sum up what I have said so far, then, the suggestion is this: An agent who is
certain that some form of (maximizing, agent-neutral) consequentialism is correct,
but uncertain—even quite radically uncertain—about what sorts of things have
value or disvalue nevertheless can find in the shared (or overlapping) assumptions of
the various first-order theories to which she assigns positive credence sufficient basis
for rationally perspicuous, non-arbitrary intertheoretic value aggregation. Nothing
I have said establishes incontrovertibly that such agent is rationally required to
maximize expected value, computed in the sort of way I have described. So far, I
have only tried to establish that, contra pessimists about intertheoretic comparisons,
there is a reasonable and non-arbitrary way of aggregating the value assignments of
rival moral theories in the face of certain kinds of moral uncertainty.
Importantly, in the Easy Case and its close cousins, it seems that both PIVA
and the grounding problem have easy and natural solutions. Alice, in the Easy
9Alternatively, if she does have metaethical uncertainties about the nature of value, these may be
probabilistically independent of her first-order moral beliefs. Just as metaethical uncertainties need
not pose any problem for an agent maximizing expected utility in the face of empirical uncertainties,
so long as the metaethical and empirical uncertainties are probabilistically independent, so those
metaethical uncertainties need pose no obstacle to hedging for first-order moral uncertainties, so
long as the two are independent.
10See Appendix B, particular §B.2 and §B.5 for an extended defense of the claim that this sort of
intertheoretic agreement or overlap can provide a basis for normalization between moral theories.
151
Case, can say quite easily what grounds the procedure she uses to aggregate her
hedonistic and pluralistic theories: namely, the structure and content of the theories
themselves. As both Hudson (1989) and Gracely (1996) point out, it cannot be the
truth of these theories (i.e., the moral facts that these theories correctly describe)
that grounds the correctness of the aggregation procedure, since at least one and
possibly both of these theories will turn out to be false. Rather, what seems to
make it rational for Alice to normalize her two theories at the value of a hedon
and maximize expected value is her credal state, i.e., the fact that she distributes
her moral credence over a pair of moral theories with particular contents, plus the
truth of some principle of rationality (like the enkratic principle) that translates her
beliefs about objective reasons into subjective reasons and rational requirements.
5.3.2 Second Case Study: Absolutist Deontology
Is the Easy Case an isolated exception with respect to intertheoretic aggre-
gation? There is no use denying that it is an especially favorable case—a case of
intertheoretic moral uncertainty that “behaves” helpfully like empirical uncertainty
and seems to present only one salient option for normalizing the moral theories in
which the agent has positive credence. Nevertheless, I will claim, there is an un-
derlying idea that has the potential to generalize quite widely: namely, that sets of
moral theories united by the right common content can be aggregated by procedures
grounded in that shared content.
To illustrate this general idea (and more especially, to show that we were not
152
in the last section simply taking advantage of an idiosyncratic feature of consequen-
tialist moral theories), let’s consider another case far afield from the first: namely,
absolutist deontological theories that categorically prohibit certain kinds of actions
like telling a lie, breaking a promise, or killing the innocent. No less than conse-
quentialists, defenders of such theories must find something to say about how agents
ought to act under morally relevant empirical uncertainty, e.g. uncertainty whether
some act would violate a deontological constraint. If I am absolutely prohibited from
breaking my promises, what should I do when I am unsure whether I made some
particular commitment as part of a long-ago act of promising? If I am absolutely
prohibited from killing the innocent, how am I to assess courses of action that carry
some (perhaps moderate to vanishingly small) risk of killing the innocent, or when
I have good reason to kill (in defense of myself or others) someone who might be
either an innocent or a malevolent threat?11
Let’s imagine, then, a deontologist David who has the following simple view
of how to deal with uncertainty: An absolute moral requirement is incumbent on
an agent whenever his degree of belief that the conditions of that requirement are
11This last possibility is suggested by Jackson and Smith (2006). They describe a case in which
a skier is headed down a mountain slope in a direction that will trigger an avalanche, killing ten
people, unless you shoot and kill him. You are uncertain whether the skier is ignorant of the
danger he poses to the ten and therefore morally innocent, or intentionally trying to bring about
their deaths. The deontological theory to which you subscribe permits (or requires) you to shoot
the skier in the latter case but prohibits it absolutely in the former.
As Jackson and Smith point out, formulating the antecedents of deontological principles in terms
of intentions—e.g. that we are prohibited not from breaking a promise or killing the innocent but
from forming an intention to break a promise or kill the innocent—does not avoid the need to deal
with uncertainty, both because (as even Kant admits) we can be quite radically uncertain about
the contents of our own intentions and because it is far from obvious whether and in what cases
forming an intention to perform some action that I believe I may have promised not to do, or to
kill someone who may be innocent, should count as impermissibly forming an intention to break
my promise or to kill the innocent.
153
satisfied is greater than .5. So, for instance, if it is in his power to perform some
action that he might or might not have promised to perform, he is absolutely required
to perform it so long as he believes it more likely than not that he did so promise;
he is absolutely prohibited from uttering a particular sentence if he believes it more
likely than not that his intention in doing so would be to deceive; and so on.
This sort of view is open to serious objections, but it seems to me as plausible
as anything else deontological absolutists can say about the problem of uncertainty,
so let us suppose that this is how some plausible class of deontological theories deals
with the problem.12 Now, just as the injunction to maximize expected value held
in common by the consequentialist theories considered in the last section extends
easily to encompass intertheoretic moral uncertainty (among theories of the relevant
class) as well as empirical uncertainty, so too the probability threshold view of
deontological obligations provides ready guidance to an agent like David who divides
his beliefs between various deontological theories that have this threshold view in
common.
Suppose, for instance, that David must decide whether to kill Thomas in or-
der to prevent Thomas from killing ten innocent people. He is empirically uncertain
whether Thomas is an innocent threat or a malevolent aggressor, but he is also
12Two prominent responses to Jackson and Smith’s (2006) uncertainty-based objection to ab-
solutist deontology, namely Hawley (2008) and Aboodi et al. (2008), both defend versions of this
“threshold” proposal. In Tarsney (forthcoming b) (the penultimate draft of which is reproduced as
Appendix A of this dissertation), I argue that, given a natural representation of the value assign-
ments of deontological theories, deontologists acting under either empirical or moral uncertainty
are at least rationally required to avoid options that carry a risk greater than or equal to .5 of
violating a deontological moral constraint when there are morally safer options available, since
those safer options will stochastically dominate their riskier alternatives. This leaves open the
possibility that deontologists are subject to more stringent requirements of moral caution, e.g. on
the basis of expectational reasoning.
154
morally uncertain, dividing his beliefs between one absolutist deontological theory
that prohibits the killing of innocent threats in other-defense and another that per-
mits it. David, then, must divide his beliefs among three salient possibilities: (i)
Thomas is a non-innocent threat, and may therefore permissibly be killed to save
a greater number. (ii) Thomas is an innocent threat, and innocent threats may
permissibly be killed. (iii) Thomas is an innocent threat, and innocent threats may
not permissibly be killed.
It seems quite natural to say, given that both moral theories to which David
assigns positive credence agree on a probability threshold of .5 at which deontological
prohibitions subjectively “kick in,” that he is subjectively prohibited from killing
Thomas iff the probability he assigns to possibility (iii)—that is, the conjunction of
the empirical belief that Thomas is an innocent threat and the moral belief that it is
absolutely impermissible to kill an innocent threat—is greater than .5. Absent some
prior reason for David to treat moral and empirical uncertainty differently in his
practical deliberations, the principle he accepts with certainty, that it is subjectively
wrong to choose some practical option O whenever O carries a risk greater than
.5 of violating an absolute moral constraint, seems to address the former kind of
uncertainty as much as the latter.
Again, I have thus far not attempted to give any knockdown positive argument
that this is the only rational way for an agent in David’s position to deliberate. As
in the last section, the point for now is simply that there is a rationally perspicuous,
non-arbitrary way of responding to a particular kind of intertheoretic moral uncer-
tainty that takes advantage of common principles shared by some class of moral
155
theories over which an agent’s credence is distributed.
5.4 Is Incomparability Anywhere a Threat to Comparability Every-
where?
So far I have suggested that an agent who divides her beliefs among a class of
sufficiently similar moral theories can possess a rational basis for intertheoretic value
aggregation and hence for moral hedging. But I have not addressed the (doubtless
more realistic) case of an agent who assigns positive degree of belief to a much
more diverse class of theories that share no obvious features to ground intertheo-
retic aggregation—e.g., an agent who assigns positive credence to both maximizing
consequentialist theories and absolutist deontological theories.
In the next chapter we will discuss some possible approaches to these “hard
cases” of intertheoretic value comparison. Some of these approaches will involve
a retreat from the ambition of full-blooded quantitative comparison between any
two moral theories, no matter how unlike, while the approaches that attempt to
sustain this ambition will remain tentative and underspecified at best. Absent a
solution to the problem of intertheoretic value aggregation in its full generality,
therefore, a pessimist may argue as follows: “The principle of hedging under moral
uncertainty requires that agents be able to compare and aggregate degrees of value
across theories to which they assign positive degrees of belief, in general. Even if
there are some easy cases where it looks like such aggregation is possible, as long as
the hard cases remain unresolved, the problem of intertheoretic value aggregation
156
as a whole is unresolved and so still poses a decisive obstacle to the principle of
intertheoretic moral hedging.”
My response is simply this: Someone who opposes any form of hedging for
moral uncertainty, but shares the ordinary view that agents should be responsive
to empirical risks and uncertainties, must justify drawing a line between the moral
and the empirical. What cases like those of Alice and David show is that PIVA
cannot justify drawing such a line. Rather, absent some other argument that agents
should not be responsive to moral uncertainties (like those considered in Chapter
3), problems of incomparability suggest the need for special principles of rationality
to handle incomparability, rather than special principles (or a special absence of
principles) for responding to moral as opposed to empirical uncertainty.
To put it slightly differently, the existence of plausible procedures for interthe-
oretic value comparison and aggregation in certain cases of moral uncertainty, taken
together with the various forms in incomparability that can arise in the absence
of any moral (or other normative) uncertainty, suggests that the problem of in-
comparability is simply orthogonal to the distinction between moral and empiri-
cal uncertainty. Just as intertheoretic value comparisons do not always generate
(even apparent) incomparability, so (apparent) incomparability of normative con-
siderations does not arise only in the context of intertheoretic comparisons. Many
philosophers, for instance, have held that certain pairs of moral values are gen-
uinely incomparable—that is to say, they have advocated first-order moral theories
that generate intratheoretic incomparability. An agent might believe, say, that pa-
triotic and familial obligations are genuinely incomparable, such that when these
157
values come into conflict (as when she feels called to volunteer for a war of na-
tional self-defense, but cannot do so without abandoning an ailing relative), there
is no uniquely rational resolution. Incomparability may arise in more mundane,
non-moral contexts as well—for instance, if I am shopping for home decor and must
choose between satisfying my own aesthetic sensibilities and pleasing my spouse,
whose tastes I regard as gauche.13 Finally, even on a simple utilitarian theory of
value that does not admit this sort of incomparability, purely empirical uncertainty
can give rise to a different species of incomparability, as it does in the Pasadena
Game (Nover and Ha´jek, 2004), for which the expected values of playing and not
playing appear to be incomparable.
But it would be absurd, on the basis of such cases, to deny the rationality
of quantitative value comparisons in general, to claim that the existence of one
or more of these forms of incomparability shows that the procedure of expected
value maximization is unreasonable in any context and hence to claim that there
is no rational basis, say, for investors in the stock market to choose investments
that maximize their expected financial returns. Likewise, it seems equally absurd
to suggest that the (apparent) impossibility of quantitative comparison between
utilitarian and deontological moral considerations should also preclude comparisons,
e.g., among possible values in a consequentialist scheme under conditions of more
limited moral uncertainty.14
13Thanks to an anonymous reviewer for this example.
14MacAskill (2013) points out that incomparability can be “infectious” in that an agent who
has any positive credence, however slight, that incomparable values are at stake in a given choice
situation and tries to calculate the expected value of her various options will in general find them to
be undefined. But again, this worry is not unique to the problem of moral uncertainty, and is not
solved by ignoring moral uncertainty. The investor who has some vanishingly small credence that
158
If we conclude that intertheoretic value aggregation is sometimes possible and
sometimes not, though, what picture of rational choice under moral uncertainty
does this leave us with? One possibility is that the sorts of views endorsed by pes-
simists about intertheoretic aggregation—like MFT, MFO, or irrelevantism—should
be recast with comparability classes of moral theories in the role that the standard
versions of these views give to individual moral theories. That is, if we conclude
that comparison and aggregation are simply impossible for some pairs of moral the-
ories, we might then accept a second-order view like “Choose an option that the
comparability class of theories in which you have highest credence regards as per-
missible,” or “Choose the option that a credence-weighted plurality of comparability
classes in which you have positive credence regards as permissible,” or “Choose an
option that the true comparability class of moral theories (i.e., the comparability
class that includes the true moral theory) regards as permissible.” Alternatively, it
may turn out that small comparability classes that offer precise, quantitative bases
for intertheoretic aggregation can be grouped together into larger classes that offer
progressively weaker or rougher bases for aggregation, in a kind of nested hierarchy.
We will explore these possibilities at greater length in the next chapter.
her investment decisions implicate values like patriotism and family that she judges, as a matter of
first-order moral theory, to be incomparable, will find herself in a similar predicament. The problem
of infectious incomparability, in this respect, is structurally analogous to more familiar decision-
theoretic problems like Pascal’s Wager (a problem, one might say, of “infectious infinities”), and
must likewise be resolved in a general way by decision theory.
159
5.5 “Top-Down” vs. “Content-Based” Approaches to Intertheoretic
Aggregation
It may seem unparsimonious to hold that different cases of moral uncertainty
should be treated differently—e.g. that some classes of theories, like the consequen-
tialist theories in which Alice has positive credence, should be aggregated expec-
tationally, while others, like the deontological theories in which David has positive
credence, should be aggregated according to a threshold principles, while perhaps
still others cannot and should not be aggregated at all. If one is sufficiently averse
to such a piecemeal approach, then one might conclude on this basis that the exis-
tence of incomparability between any pair of moral theories is enough to render all
intertheoretic comparisons moot, even when those comparisons seem quite natural.
I will close this chapter, therefore, by describing what I take to be some general at-
tractions of this piecemeal approach—what I will call the “content-based” approach
to intertheoretic aggregation—before attempting to develop the approach in greater
detail in Chapter 6.
The standard approach to intertheoretic aggregation, comprising nearly all
the positive theories of rational choice under moral uncertainty that have been
proposed in the literature to date, is what I will call the “top-down” approach.
It is characterized by the following two features.
Top-Down Aggregation (1) There is a single general principle of intertheoretic
aggregation and choice under moral uncertainty that applies to all first-order
160
moral theories and to all agents regardless of their distribution of belief over
first-order moral theories. (2) This general aggregation procedure is sensitive
only to the value assignments of the first-order theories in which the agent has
positive credence (as well as the degree of credence assigned to each theory),
not to any of the other content of those theories.
This sort of top-down approach is exemplified by Lockhart (2000). On Lock-
hart’s view, morally uncertain agents should always choose options that maximize
expected moral rightness15, where the expected moral rightness of options is deter-
mined by normalizing the value scales of all moral theories in which the agent has
positive credence according to the following principle.
Principle of Equity among Moral Theories (PEMT) “The maximum degrees
of moral rightness of all possible actions in a situation according to competing
moral theories should be considered equal. The minimum degrees of moral
rightness of possible actions in a situation according to competing moral the-
ories should be considered equal unless all possible actions are equally right
according to one of the theories (in which case all of the actions should be
considered to be maximally right according to that theory).” (Lockhart, 2000,
p. 84)
Thus on Lockhart’s view both the principle for making intertheoretic com-
15More precisely, Lockhart’s principle PR4 asserts that “In situations in which moral agents are
uncertain of the degrees of moral rightness of some of the alternatives under consideration, a choice
of action is rational if and only if the action’s expected degree of moral rightness is at least as great
as that of any other alternative” (Lockhart, 2000, p. 82). This is later supplemented by another
principle, PR5, that governs situations in which an agent cannot be sure which option satisfies
PR4 (p. 95).
161
parisons (viz., PEMT) and the procedure for aggregating rival theories in light of
those comparisons for purposes of rational decision-making (viz., maximize expected
moral rightness) apply to any agent regardless of her distribution of first-order moral
belief. And PEMT, which treats the range of every theory’s value assignment over
options in a given choice situation as equal, is insensitive to any features of those
first-order moral theories apart from their value assignments.16
The alternative approach suggested by the cases of Alice and David differs
from this top-down approach in both respects.
Content-Cased Aggregation (1) Different aggregation procedures are appropri-
ate for different sets of moral theories, or at least for agents who distribute
positive degree of belief over different sets of moral theories. (2) These aggre-
gation procedures are, at least in some cases, sensitive not just to the value
assignments of the theories they aggregate, but to the metaethical and other
theoretic claims that ground, justify, or explain those value assignments.17
Setting aside arguments for or against any specific top-down or content-based
16Other approaches to decision-making under moral uncertainty that are “top-down” in this
sense include MFT, MFO, the “Conceivability PEMT” approach (proposed but not endorsed
by Sepielli (2013)), and Bostrom’s “parliamentary model” (Bostrom, 2009). The “background
rankings” approach defended in Sepielli (2009) is top-down insofar as it is sensitive only to the
value assignments of the theories it aggregates, but Sepielli does not claim that this approach to
aggregation can be applied to any pair of first-order theories. The approach defended in MacAskill
(2014) is difficult to classify, but includes a large top-down component that involves normalizing
theories at the variance of their value assignments. (MacAskill stresses the distinction between
cardinal and ordinal theories, but I will argue in the next chapter that the variance normalization
approach effectively ignores this distinction.) As noted above, Ross (2006) briefly suggests the
possibility of what I am calling content-based aggregation. To my knowledge, this is the closest
anyone in the literature has come to defending a content-based approach.
17Greaves and Ord (forthcoming) also consider a “content-based” approach to intertheoretic
value comparisons. But they use this term differently than I do, in particular to cover ideas
like Sepielli’s “background rankings” approach that are sensitive to features of a theory’s value
assignment that an approach like Lockharts PEMT ignores, but are not sensitive to any features
of a theory beyond its value assignment.
162
theory of rational choice under moral uncertainty, there are two general reasons to
prefer the content-based approach. The first is the intuitive relevance of the con-
tent underlying a first-order theory’s value assignment, which top-down approaches
by definition ignore. On the one hand, in cases like the Easy Case, top-down ap-
proaches cannot yield the intuitively correct normalization of rival theories. Lock-
hart’s PEMT, for instance, will almost always imply that the value of a hedon
according to Alice’s hedonistic theory is either more or less than the value of a he-
don according to her pluralistic theory, even though there seems to be every reason
to think these values equal.18 On the other hand, top-down approaches miss in-
tuitive inequalities between theories with indistinguishable value assignments. For
instance, consider two versions of hedonistic utilitarianism, one that derives its value
assignment from a divine command theory of ethics plus an unusual reading of scrip-
ture, and another that derives its hedonism from the Humean theory of reasons plus
an unusual theory of human psychology. The value assignments of these two theories
are isomorphic (and therefore, prior to normalization, indistinguishable). Neverthe-
less, it seems quite plausible that an agent who has some credence in hedonistic
utilitarianism should give those utilitarian reasons more weight if her credence is in
the divine command theory rather than the Humean theory. A top-down approach
like Lockhart’s, however, will assign the same weight to both theories.
The second reason to prefer the content-based approach is more subtle. The
content-based approach, I have suggested, offers a plausible response to the ground-
ing problem, i.e. a plausible account of why an agent is rationally required to aggre-
18See §§B.2-B.3 for an extended defense of this claim.
163
gate her various first-order theories in a particular way: namely, that the aggregation
procedure in some sense follows from the content of those theories, so that she ra-
tionally ought to accept it conditional on any of those theories being true. For
instance, the principle that an agent ought never choose any option that carries a
risk greater than .5 of violating an absolute moral constraint is endorsed by both the
deontological theories in which David has positive credence, and implies a particular
procedure for aggregating those theories and making rational choices in light of his
uncertainty between them. Top-down approaches cannot claim this advantage, be-
cause the second-order principles they endorse apply to any agent regardless of her
first-order credences, and therefore apply to agents whose first-order moral beliefs
may be in deep tension with, if not actively contradict, the second-order princi-
ple proposed by the top-down account. That is, any top-down approach runs the
risk of “begging the question,” in the sense we have encountered in previous chap-
ters, against first-order theories in which many agents may have justified positive
credence.
Consider the ways in which Lockhart’s approach might be in tension with
an agent’s first-order moral beliefs. First, as Lockhart himself acknowledges, the
principle that uncertain agents should always maximize expected moral rightness
assumes the lexical priority of moral over non-moral reasons (a principle Lockhart
labels the “finality thesis”). This is a claim, however, that many first-order nor-
mative views reject (e.g. Wolf (1982)), and about which a reasonable agent might
be substantially uncertain. Second, the application of PEMT requires that all the
first-order theories over which an agent distributes her belief be in the business of
164
assigning “degrees of moral rightness” to actions. Many deontological theories, of
course, do not treat rightness and wrongness as exhibiting gradation, and though
one can for instance simply treat such theories as assigning +1 to all morally oblig-
atory actions, −1 to all morally prohibited actions, and 0 to all morally neutral
actions, one arguably does such theories an injustice by cramming them into this
sort of quantitative mold. But even more clearly, virtue-theoretical approaches to
ethics will not fit neatly into Lockhart’s box. Virtue ethics is arguably not in the
business of classifying actions as right or wrong at all, let along right or wrong in
degrees, and it is certainly at least reasonable to have substantial credence in a ver-
sion of virtue ethics that conceives of itself in this way. If you do, then Lockhart’s
second-order view will be in some tension with your first-order beliefs.19 Third, as
Sepielli (2006) notes with respect to Lockhart’s view, as Hedden (2016) notes with
respect to expectational approaches to moral uncertainty generally, and as we saw
in the last chapter, approaches like Lockhart’s are in tension with first-order views
that embrace features like agent-centered option or supererogation, denying in one
way or another that one is rationally required to choose the option for which there
is most reason. As I conceded in the last chapter, first-order theories according
to which one is sometimes permitted to ignore the balance of moral reasons might
conflict with and hence undermine second-order principles like Lockhart’s.
This list of potential tensions is by no means exhaustive, but it will serve to
illustrate the point. Lockhart’s approach will look much more appealing to someone
19For useful discussion of the limited compatibility of deontological and virtue ethical theories
with the sort of decision-theoretic framework Lockhart presupposes, see Colyvan et al. (2010).
165
who divides her first-order moral credence mainly among maximizing consequential-
ist theories than to someone who divides her credence mainly between Kantian de-
ontology, Aristotelian virtue ethics, and “commonsense morality” views that incor-
porate large helpings of moral-prudential conflict and supererogation. The minimal
point is that the latter sort of agent will likely be very uncertain about Lockhart’s
principle, suggesting either that she should adopt some alternative second-order
principle that seems more plausible in light of her evidence, or that she should be in
the market for some third -order principle of manage her uncertainty about second-
order principles. It is less obvious that Lockhart’s principle actually contradicts any
of the first-order views this agent entertains—after all, it is not logically inconsistent
to hold, for instance, that as a first-order matter rightness and wrongness don’t come
in quantitative degrees, but that as a second-order matter it is rational to treat them
as if they come in degrees (and likewise for other incongruous pairings of first- and
second-order principles). But in any case, there is something quite plausible about
the thought that our first agent, who distributes her first-order belief mainly over
forms of maximizing consequentialism, should appeal to an expectational principle
like Lockhart’s to guide her actions, given how well it jibes with her first-order be-
liefs, while there is something quite implausible about the thought that our second
agent should do the same, given how poorly these expectational principles (at least
in their Lockhartian manifestation) jibe with her first-order beliefs.
The advocate of a top-down approach like Lockhart’s, therefore, cannot plau-
sibly claim that any agent, whatever her other beliefs and evidence, is rationally
required to believe the same second-order principle. For any such principle (like
166
“maximize expected rightness” + PEMT), there will be many agents who, in light
of their justified first-order moral credences, have reason to disbelieve that principle.
Rather, top-down theorist seems forced to claim that their preferred second-order
principle has external, belief-independent normative force. As I will argue in Chap-
ter 7, externalism itself is not an unacceptable price to pay—as Weatherson claims,
we all must accept it in at least some weak form to escape the threat of infinite
regress. But, as I will also argue in Chapter 7, the principles that can most plausi-
bly be regarded as having external normative force are those that are constitutive
of rationality as such (namely, I will claim, the enkratic principle). Particular in-
tertheoretic aggregation procedures involving technical normalization principles like
PEMT are not good candidates to play this sort of constitutive role in a theory of
rationality. Thus it is more plausible to hold that at least some features of intertheo-
retic aggregation are relativized to the theories being aggregated. This will leave us
with a piecemeal picture of rational choice under moral uncertainty, but it will also
leave us with a picture that can be largely grounded in the agent’s own normative
beliefs, rather than imposed on those beliefs by force.
So far, however, we have only seen hints of what a content-based approach
to intertheoretic value aggregation might look like, derived from examination of
particular cases like those of Alice and David. The task of the next chapter will be
to develop a more general theory that can go beyond these simple cases and address
more difficult cases in which agents distribute their beliefs over a wide range of
first-order theories.
167
Chapter 6: Outline of a General Theory
In the last chapter I argued that the existence of classes of moral theories
within which precise and perspicuous intertheoretic value comparisons are possible
means that the problem of intertheoretic value comparisons does not constitute a
general objection to moral hedging. And I gestured at a general account of decision-
making under moral uncertainty suggested by the existence of such comparability
classes. The goal of this chapter is to make good on that suggestion and describe
in greater detail a decision procedure based on aggregating the prescriptions of
competing moral theories via a nested hierarchy of comparability classes.
The account given in this chapter will nevertheless be incomplete and tentative.
The problem of identifying an ideally rational, fully general procedure for decision-
making under normative uncertainty is enormously difficult, as demonstrated by the
weaknesses of every extant attempt in the literature (some of which I have already
criticized, and others of which will be addressed in this chapter). My aim is simply
to offer a new candidate approach that avoids many of the drawbacks of extant
proposals, but which will need a great many details filled in before it can be fully
assessed.
To describe this approach, we must first give a more precise account of the
168
problem it aims to solve. So in the next section I will distinguish three aspects of the
problem of intertheoretic value aggregation, i.e., three barriers that an adequate ap-
proach to intertheoretic aggregation and decision-making under moral uncertainty
must overcome. Then, in §§6.2-6.4, we will examine each of these three problems
individually, describing and comparing several candidate approaches to each prob-
lem. Finally, §6.5 combines what seem like the most promising approaches to each
aspect of the problem of intertheoretic value aggregation to arrive at a tentative,
content-based account of rational decision-making under moral uncertainty.
6.1 Three Components of Comparability
The examples of apparent intertheoretic comparability given in the last chapter
possess three advantageous features of which I implicitly or explicitly took advantage
in arguing for the possibility of aggregation in these cases. In this chapter I will
suggest, as a plausible hypothesis, that these three conditions are jointly sufficient
for intertheoretic comparability and aggregation whenever they are fully realized by
a pair of theories T1 and T2. The three features are as follows.
1. Structural Compatibility The value scales of T1 and T2 have identical struc-
ture (e.g., ordinal, interval, or ratio).
2. Scale Normalizability There is some set of practical options {O1, O2, ...On}
such that T1 and T2 are in full agreement concerning all the features of the
world that ground their evaluations of O1−n, so that the value scales of T1 and
T2 can be normalized based on their agreement with respect to those options.
169
3. Decision-Theoretic Compatibility T1 and T2 endorse the same decision
theory, i.e., the same principles for aggregating conflicting value assignments
in the face of uncertainty.
For instance: In the Easy Case, where Alice divides her beliefs between hedo-
nistic utilitarianism and a pluralistic theory that values both hedons and aesthetons,
(i) structural compatibility is given by the fact that both theories have cardinal value
scales; (ii) scale normalizability is given by the fact that both theories are in per-
fect agreement about the choiceworthiness of options that implicate only hedonic
and not aesthetic value; and (iii) decision-theoretic compatibility is given by the
fact that both theories endorse expected value maximization as the correct decision
procedure under uncertainty.
In the case of David, who divides his belief between two deontological theories,
some details must be filled in, but we can tell a similar story. For instance, per-
haps both deontological theories share a lexical value structure with one ordinally-
structured, lexically prioritized dimension representing deontological obligations and
prohibitions and another, cardinally structured, lexically deprioritized dimension
representing all other sorts of reasons. Scale normalizability comes from the pre-
sumed fact that both theories, despite disagreeing for instance about the permissili-
bity of killing innocent threats, nevertheless agree about the strength of other moral
obligations (for instance, against breaking promises or killing innocent non-threats)
as well as the strength of non-moral reasons. And decision-theoretic compatibility is
given by the fact, which we stipulated, that both theories endorse the same thresh-
170
old approach to decision-making under uncertainty about deontological constraint
violations.
Are these three conditions, in all cases, jointly sufficient to establish the pos-
sibility of intertheoretic comparison and aggregation under moral uncertainty? I’m
not sure that any general proof of this claim is possible. But intuitively, structural
compatibility and scale normalizability allow us to put the value assignments of
T1 and T2 on a single, shared scale, and decision-theoretic compatibility gives us
a decision procedure, the correctness of which we are certain of conditional on the
disjunction of T1 and T2, for handling uncertainty about the values of options on
that shared scale. So it seems reasonable to think that whenever all the theories in
a given class satisfy these three conditions with respect to one another, it should
be possible for an agent to aggregate the competing value assignments of those the-
ories. On the tentative assumption that these three conditions are sufficient for
full intertheoretic comparability and aggregation, let’s call a class of moral theories
that fully satisfies these conditions (that is, a class any two members of which have
identically-structured and normalizable value scales and endorse the same decision
theory) a minimal comparability class.
But if we take these three conditions as jointly sufficient for intertheoretic
aggregation, we should next ask to what extent they are individually necessary—
that is, how should an agent decide what to do when she distributes her credence over
a class of theories that satisfies one or more of these conditions only partially, or not
at all? If we start from the idea that the sorts of aggregation procedures described
above, in simple cases where all three conditions are met, are rational when an agent
171
distributes the entirety of her moral credence over a single minimal comparability
class, then a natural way to develop a general theory of decision-making under
moral uncertainty is to ask what happens as we relax each of these conditions.
This is the approach I will adopt in this chapter: We will consider each of the
three conditions in turn, discuss various ways in which that condition might be less
than fully satisfied, and consider various ways of extending the account of decision-
making under moral uncertainty from the simple case of minimal comparability
classes to accommodate these increasing difficulties. My primary aim is to identify
possibilities, i.e., reasonable candidate answers for how to accommodate weakening
of each of the three conditions that can then be combined into reasonable candidate
theories of decision-making under moral uncertainty in general. But in §6.5 I will
identify the combination of responses, and hence the general theory of decision-
making under moral uncertainty, that strikes me as most plausible.
6.2 Structural Compatibility
Perhaps the most striking challenge of intertheoretic value aggregation is the
need to compare theories with differently structured value scales. For instance,
anyone who wishes to assert that maximizing expected choiceworthiness provides a
universal principle for decision-making under moral uncertainty must assume that
all moral theories assign cardinal values to options that, suitably normalized, can
be taken as inputs to compute intertheoretic expectations. But, as MacAskill in
particular has emphasized (MacAskill, 2014, pp. 14-15, 52ff), it is simply a fact to
172
be reckoned with that not all theories have cardinal structure. Some moral theories
have ordinal structure, ranking options as better or worse without giving answers
to (or recognizing as meaningful) the question “How much better/worse?” Since
ordinal values cannot be multiplied by probabilities to yield cardinal values, an
agent who has positive credence in merely ordinal theories cannot assign expected
values to her options.
Faced with such structurally uncooperative theories, it is tempting to simply
demand uniform structure from our moral theories. Lockhart’s PEMT, for example,
effectively forces cardinal structure on theories: (i) Lockhart holds that the maximal
and minimal “degrees of rightness” assigned by each theory to any option in a given
choice situation should be treated as equal on a cardinal scale, e.g., assigning the
maximally right option value 1 according to that theory and the minimally right
option value 0.1 (ii) If a theory treats any option in a given choice situation as
intermediate in rightness, value, or choiceworthiness between the best and the worst
option, then it must be assigned some intermediate value on that cardinal scale.
But if an agent has some credence in a theory with a non-cardinal value scale,
i.e., a theory according to which the choiceworthiness of practical options does not
come in cardinal degrees, then an approach like Lockhart’s is simply refusing to
grapple with the phenomenon, i.e., with the agent’s actual doxastic state. Thinking
of theories, again, as maximal consistent sets of propositions, some theories include
propositions like “The choiceworthiness of practical options can be compared ordi-
1Lockhart doesn’t specify to which cardinal values the maximal and minimal degrees of rightness
for each theory should be normalized, but this choice is of course arbitrary.
173
nally but not cardinally.” To treat a theory T that includes this proposition as if
it must secretly be assigning cardinal values to options that can be factored into a
computation of expected value if we can only get at them is to solve the problem of
moral uncertainty, not for an agent for has positive credence in theory T, but only
for an agent who has credence in some cardinal surrogate of T.
If we want to know how an agent should act when she distributes belief over
an arbitrary range of moral theories, therefore, we must grapple with the problem of
structural diversity. MacAskill (2014) is the first to seriously attempt this, offering
an account of rational choice under moral uncertainty that is bifurcated between
cardinal and ordinal theories: the appropriate decision rule for cardinal theories, he
holds, is to maximize expected choiceworthiness, while the appropriate decision rule
for ordinal theories is a complex, probability-weighted version of a Borda count.
As I will argue at more length shortly, it seems to me that MacAskill’s approach
fails to genuinely respect the phenomenon of merely ordinal theories, since the Borda
counting approach (and MacAskill’s version of it in particular) is non-arbitrary only
on the assumption that there are “hidden” cardinal values underlying the ordinal
rankings of merely-ordinal theories. But a more immediate problem for MacAskill’s
account is that it acknowledges only a very small part of the potential structural
diversity of moral theories—for many more structures are possible than the cardinal
and ordinal structures on which MacAskill’s account in focused.2
It is worth cataloging some of the ways in which moral theories might differ
2MacAskill distinguishes between interval-structured and ratio-structured cardinal theories, but
given his view that we should always maximize expected choiceworthiness over cardinal theories,
this distinction plays no role in his ultimate theory. He also briefly acknowledges the possibility of
lexically-structured theories with two cardinal dimensions (MacAskill, 2014, pp. 47-50).
174
from one another in the structure of their value scales.
1. The cardinal structure of the real number line might be extended in various
ways to accommodate the possibility of infinite or infinitesimal values. For
instance, a theory that recognizes the possibility of infinite value or disvalue
but does not countenance distinctions of degree between infinite values has
a value scale with the structure of the extended real number line (the real
numbers plus limit values +∞,−∞). Alternatively, if we wish to distinguish
between degrees of infinite value and/or to represent the possibility of infinites-
imal value, we might have some credence in a theory whose value scale has
the structure of the hyperreal numbers or the surreal numbers. And of course
many other such extensions of the real numbers as possible as well.3
2. A theory’s value scale might have more than one dimension. MacAskill briefly
considers a theory that represents degrees of choiceworthiness by two-dimensional
vectors, lexicographically ordered (i.e., (x1, y1) ≥ (x2, y2) ⇐⇒ (x1 > x2 ∨
(x1 = x2 ∧ (y1 ≥ y2)). Such a two-dimensional structure could also be used to
represent incomparable values ((x1, y1) ≥ (x2, y2) ⇐⇒ (x1 ≥ x2 ∧ y1 ≥ y2)).
Either of these ideas, of course, could be extended to arbitrarily many dimen-
sions (even infinitely many dimensions if, for instance, one thinks that there
might be infinitely many incomparable forms of value). And the various di-
mensions of a multi-dimensional value scale could themselves display various
3Bostrom (2011) describes, without endorsing, one way of using the hyperreal numbers to
compare the utilitarian value of infinite worlds. Several other approaches to the problem of infinite
utility (e.g. those of Vallentyne and Kagan (1997) and Arntzenius (2014)) extend the structure of
the utilitarian value scale beyond that of the real numbers is less drastic ways.
175
structures, e.g. ordinal, interval, ratio, extended reals, hyperreals, etc.
3. At the other end of the scale of complexity, a theory could simply have what
we might call binary structure: classifying options as permissible or impermis-
sible, but countenancing no ranking of permissible options as better or worse.
This might be the structure, for instance, of an extreme libertarian view ac-
cording to which morality consists entirely of a list of absolute, negative side
constraints, which it is never permissible to violate, and such that all options
that do not violate any side constraint are equally choiceworthy.
Is there any order underlying this chaos of possible structures? Perhaps not,
but the following strikes me as a plausible hypothesis: What is fundamental to
any normative theory is binary structure—more specifically, the identification of
some subset of an agent’s options as permissible or “eligible for choice.” Since the
fundamental job of any normative theory is to guide an agent’s choices, to count
as a normative theory it must “tell the agent what to do,” i.e., it must declare
one or more of her options in any given choice situation choiceworthy simpliciter.4
Therefore, nothing can count as a normative theory (or, a fortiori, as a moral theory)
which does not make the binary distinction between permissible and impermissible
options. But the diversity of possible structures explored above is explained by the
4Of course, pace irrelevantism and MFT, the value assignment given by a particular moral
theory is not fully action-guiding in that it does not tell an agent what to do in light of her moral
uncertainties. If we consider the fully objective value assignments given by a complete (moral and
non-moral) theory of the world, this is all the more true, since an objective value assignment does
not guide an action’s actions even in light of her empirical uncertainties. Nevertheless, the value
assignments given by moral theories (whether fully objective or partly subjective) play an essential
role in action guidance, both as inputs to an action-guiding, fully subjective value assignment, and
insofar as an agent who is certain that a particular moral theory is true must be able to derive
action guidance from its value assignment.
176
many possible extensions of binary structure: any structure that extends the binary
structure of permissible/impermissible (or, equivalently, that can induce a binary
structure) is a possible structure for a moral theory’s value scale.
Since this idea of universal binary structure offers helpful suggestions for the
general theory of decision-making under moral uncertainty, it is worth a brief explo-
ration. First, is it true that all moral theories make a binary classification of options
as permissible or impermissible?5 Consider some examples: Classical utilitarianism
ranks options on a cardinal value scale, and hence does not simply classify options as
permissible or impermissible. But it holds that all and only the options that produce
maximal utility are permissible, so it has a binary structure induced by its cardinal
structure. A satisficing version of utilitarianism has the same cardinal structure as
maximizing utilitarianism, but draws the line between permissible and impermissi-
ble options differently, regarding some options that don’t produce maximal utility
as nevertheless permissible.6 Like satisfying utilitarianism, commonsense morality
(which recognizes the supererogatory) and perhaps also Kantianism (which its idea
of imperfect duties) recognize that some permissible options are more choiceworthy
than others, while nevertheless preserving a distinction between permissible and
impermissible.
One type of theory that might seem to challenge the universality of binary
5Here it is particularly worth noting that the following discussion is meant to generalize to
normative theories, moral or otherwise, but I focus on moral theories for simplicity.
6This suggests that theories like utilitarianism have more than just cardinal structure, i.e., they
have cardinal structure plus a “permissibility” predicate that does not logically supervene on the
assignment of cardinal values to options. Some might prefer to treat the division of options as
permissible or impermissible as something distinct from a theory’s value assignment, but to the
extent that the notion of universal binary structure proves helpful, it gives us reason to think of
that binary structure as part of the value assignment itself.
177
structure comprises those theories that recognize value incomparability and/or moral
dilemmas (i.e., genuine conflicts of obligation). Suppose I must decide whether to
stay home and care for my ailing parent or go join the Resistance, and that this choice
presents either a clash of incomparable values (familial obligation vs. patriotism,
justice, etc) or simply a conflict of obligations. Does a moral theory that issues
these verdicts tell me that I have permissible and impermissible options? I think so:
It may be, especially on a “moral dilemmas” view, that both my options are simply
impermissible—the universality of binary structure does not entail that at least one
option in any given choice situation must be classified as permissible, any more than
it entails that at least one option in any given choice situation must be classified as
impermissible, though of course the former principle strikes many philosophers as
extremely compelling. Perhaps it is more plausible to say that both of my options are
permissible, despite in some sense violating moral obligations, on the grounds that I
am in a situation where I cannot help violating a moral obligation. Incomparabilist
views, though they likewise do not identify either option in this situation as better
than the other or as equal in choiceworthiness (in the sense that any small addition
of reasons one way or the other would make the newly favored option better overall),
nevertheless must give me some guidance on what to do, and will presumably say
that either option is permissible (though again, they have the option of saying that
both options are impermissible).
Another kind of theory that might challenge the universality thesis is scalar
utilitarianism (Norcross, 2006). Scalar utilitarianism explicitly disclaims the notions
of obligation, prohibition, and permissibility, ranking actions only as better or worse
178
(and consequently as supported by cardinally greater or lesser degrees of objective
reason). It strikes me that there is something unstable about the scalar utilitarian
position: An ideally informed agent who accepts scalar utilitarianism and faces a
choice situation in which she knows with certainty that some option O is objectively
best would exhibit a clear rational failure were she to choose any option other than
O—which seems to provide a clear sense in which O, and only O, is a permissible
option in this choice situation, i.e., O is the only option that is objectively rational,
or that is rationally permissible for an ideally informed agent. The one way in which
a scalar utilitarian could disclaim this implication, it seems to me, is by insisting
that scalar utilitarianism offers only “a theory of the good” and not “a theory of
the right”—i.e., that it is an evaluative theory (though evaluating actions as well as
states of affairs) rather than a normative theory that describes or implies reasons
for action. But of course if scalar utilitarianism is not a normative theory, then it is
not a moral theory in the sense with which we are here concerned, and it need not
be accommodated by a general theory of decision-making under moral/normative
uncertainty.7
An interlocutor might complain, at this point, that I am doing exactly what I
criticized Lockhart for doing, namely, imposing structural requirements on theories
that the theories themselves disclaim, and hence generating a false representation of
the credences of some agents (who have positive credence in theories that disclaim
7The version of scalar utilitarianism defended by Norcross (2006) is explicitly normative in the
sense I have described, countenancing reasons for action but not obligation/prohibition, right-
ness/wrongness, or oughts. Norcross is focused on moral notions, so whether his view rules
out the notions of objective rational requirement/prohibition/permission that I have described—
understood in terms of what options an ideally rational and fully informed agent might choose—is
not obvious.
179
any binary structure). But there must be some conceptual constraints on what
counts as a moral or normative theory: Any agent might have positive credence
in any number of sets of propositions (or at least, sets of sentence) that proclaim
themselves to be moral theories but simply are not—e.g., “The only moral truth is
that 1 + 1 = 2” (a theory that says nothing about practical options) or “The only
moral truth is that all choices made on a Tuesday or Friday are blue and all choices
made on any other day of the week are green” (a theory that assigns properties to
practical options, but not the right properties to count as a moral theory). That an
account of choice under normative uncertainty fails to accommodate such theories is
no strike against it—the goal of such an account is to aggregate the verdicts of those
theories that actually meet the conceptual requirements to count as a normative
theory. Cardinal structure is not plausibly such a requirement (though it may well
be plausible as a requirement that the true normative theory must satisfy). But the
binary classification of options as eligible for choice or ineligible for choice in light of
objective reasons plausibly is such a requirement (and one that can be maintained
even if we wish to shed much of the conceptual baggage associated with permissibility
and impermissibility).
Are there other structural constraints besides binary structure that all norma-
tive theories must satisfy? One might think that the action-guiding role of normative
theories requires that they identify not just some subset of an agent’s options in a
given choice situation as eligible for choice but one option as uniquely eligible for
choice, as “the thing to do.” A normative theory that merely identifies some set of
options as choice-eligible leaves agents with a “liberty of indifference” problem: even
180
if it is the case that, say, chocolate and vanilla ice cream are both equally choicewor-
thy, then a full normative theory should tell me how to choose between them, e.g. by
telling me to choose on the basis of the first arbitrary whim that comes into my head,
or to flip a coin. But it seems to me that the liberty of indifference is not so great
a problem as to justify this “uniqueness” requirement as a conceptual constraint on
normative theories. A commonsense theory of reasons, for instance, seems to tell me
that I am permitted to choose chocolate straightaway, choose vanilla straightaway,
or adopt any sort of cost-free mixed strategy (like flipping a coin), and which of those
options I choose is simply up to me, underdetermined by my reasons. The fact that
my normative theory does not direct me to a particular option (whether a simple
action, a deterministic decision procedure, or a stochastic decision procedure) does
not mean that I will fail to make a decision.8
It might also be suggested that binary structure is just a degenerate or im-
poverished limiting case of ordinal structure, and that we should in fact conclude
that ordinal structure is the thing all normative theories have in common. It may
be unproblematic to concede this claim.9 But it strikes me as worthwhile to distin-
guish between theories (like the simple libertarian view given above) that recognize
only two possible degrees of choiceworthiness—permissible and impermissible—and
theories that countenance the possibility of arbitrary ordinal comparisons. It is a
structural feature of the former class of theories that there are never three options
8It is worth remembering that an agent’s options, in any given choice situation, will typically
include various deterministic and stochastic procedures for selecting among other options.
9Though it has substantive implications if one thinks that a voting method like Borda counting
is the right way to aggregate ordinal theories—in particular, it puts a great deal of pressure on
the question of how to extend the simple Borda counting procedure to accommodate ties (i.e.,
normative theories classifying two or more options as equally choiceworthy).
181
O1, O2, O3 such that O1  O2  O3, while fully ordinal-structured theories allow
this.10 Binary structure may be a degenerate case of ordinal structure, but it is in
any case distinct from full ordinal structure.
I conclude, therefore, that the value scales of moral theories can display a much
greater diversity of structures than has previously been recognized: namely, every
possible extension of binary structure, including ordinal, interval, ratio, extended
real, hyperreal, surreal, multidimensional (lexical or incomparabilist), and surely
many more.
6.2.1 Structural Enrichment and Structural Depletion
Having laid a deal of groundwork, we may now ask how a moral agent should
deliberate when she divides her beliefs among moral theories with differently-structured
value scales. It seems to me that there are three general options.
1. Structural Enrichment Give each theory all of the common structure nec-
essary for intertheoretic comparison by enriching, in one way or another, its
own self-proclaimed structure, e.g. mapping ordinal rankings onto a cardinal
value scale in order to achieve comparability with other cardinal theories.
2. Structural Depletion Adopt a decision procedure that relies only on mini-
mal structure, either the least structure had in common by all the theories in
which the agent has positive credence, or the least structure had in common
10And of course, to further emphasize the possibility of structural diversity, we could imagine
theories that recognize exactly three, four, or any finite number of possible degrees of choiceworthi-
ness. More esoterically, ordinal theories might also differ in whether they allow or disallow infinite
ordinals in comparing infinite classes of options, or confine their value scales to the finite ordinals.
182
by all normative theories full stop, namely binary structure.
3. Multi-Stage AggregationAggregate the value assignments of theories within
a comparability class C1−1 in ways that make use, where appropriate, of their
common structure. Then take the aggregate “value assignment” of the com-
parability class an input to a new aggregation procedure for a larger compara-
bility class C2−1 = {C1−1, C1−2, C1−3, ..}, where that new procedure makes use
only of the structure common to all theories in C2−1. Continue in this way un-
til all theories in which the agent has positive credence have been aggregated
into a single ranking of options in terms of subjective reasons.
The structural enrichment approach is typified by Lockhart’s PEMT which (as
described above) simply forces theories into cardinal form. But it is also implicitly
exemplified, I believe, by MacAskill’s approach. MacAskill holds that merely ordinal
theories should be aggregated by means of a Borda count. Borda counting is a voting
method which, in its simplest form, assigns each candidate a score equal to the sum
over every voter’s ballot of the number of candidates ranked below her on that ballot,
and then selects the candidate with the highest score. On MacAskill’s approach,
moral theories play the part of voters and practical options the part of candidates.
Ignoring for the moment the possibility of a theory ranking two options as equally
choiceworthy, the Borda score of an option O is equal to the sum over every moral
theory in which the agent has positive credence of the number of options ranked as
worse than O according to that theory times the agent’s credence in that theory.11
11MacAskill also proposes that Borda scores should be weighted by a measure on the set of
options, so as to avoid problems with option individuation, but we can safely ignore this compli-
183
Borda counting holds some appeal in the social choice context because, in many
cases, we might hope that the number of options/candidates that a given voter ranks
above or below a given candidate offers a rough approximation of the cardinal utility
she assigns to that option/candidate—that is, it is sometimes reasonable to presume
that for most voters, the difference in strength of preference between the first- and
second-ranked candidate is similar to the difference in strength of preference between
the second- and third-ranked candidate, the third- and fourth-ranked candidate, and
so on. Since it is plausible that, at least in many contexts, we would ideally like
to maximize cardinal preference utilities if we could only get at them, the Borda
method might look like an appealing approximation.
But in the context of choice under moral uncertainty, if we accept that ordi-
nal theories are genuinely ordinal, and not simply cardinal theories that stubbornly
refuse to reveal their cardinal value assignments, then this rationale for Borda count-
ing is unavailable. Indeed, absent an assumption of implicit cardinality, the Borda
counting approach appears arbitrary: Why, for instance, should being ranked third-
from-last by a given theory contribute exactly twice as much to an option’s score as
being ranked second-to-last, if the ranking genuinely does not reflect any cardinal
information?
Moreover, MacAskill defends a version of Borda counting that handles the
possibility of ties in a theory’s ordinal ranking of options by normalizing each the-
ory’s Borda score at its variance (MacAskill, 2014, p. 119).12 But this amounts
cation.
12That is, MacAskill’s approach first represents a theory’s ordinal ranking with cardinal values
1, 2, 3, etc. It then calculates the variance of those cardinal values, i.e., the average squared
distance from the mean. Finally, it “stretches” or “contracts” each theory’s ranking in order to
184
to an explicit cardinalization of ordinal theories, for there is no such thing as the
“variance” of an ordinal ranking. This makes it all the more clear that the Borda
method relies on treating ordinal theories as implicitly cardinal.13
Approaches that implicitly or explicitly impose alien structure on a theory, e.g.
by mapping ordinal rankings onto a cardinal value scale, strike me as ad hoc and
as failing to grapple with the problem of what an agent ought to do when she has
positive credence in theories with genuinely sparse structure. So it is worth looking
for alternative approaches that allow us, e.g., to aggregate ordinal theories without
treating them as cardinal.
The second approach, structural depletion, satisfies this demand. This ap-
proach involves appealing to decision procedures that rely only on the structure
actually present in all theories in which an agent has positive credence. If an agent
has positive credence only in a set of theories that display something less than the
full diversity of possible structures, this approach is compatible with a wide vari-
ety of decision procedures, about which more will be said below. But for now let’s
focus on the simplest version of structural depletion, which assumes that an agent
distributes her belief over all possible moral theories and therefore demands a de-
cision procedure that relies only on the structure that all moral theories have in
common—namely, as I have argued, binary structure.
equalize their respective variances. When none of the ordinal theories being aggregated treat any
two options as equally choiceworthy, this amounts to a simple Borda count, since each theory will
assign a number of ranks equal to the number of options, and the variance of every theory will be
the same to begin with. But when some theories posit ties, the variance normalization method has
a substantial impact on the outcome of the Borda method, giving comparatively greater weight to
theories that posit a large number of ties.
13I discuss MacAskill’s Borda method at much greater length in Tarsney (unpublished).
185
If we set aside all the surplus structure of every moral theory apart from its
classification of practical options as permissible or impermissible, then there seems
to be one salient option for intertheoretic aggregation: namely, the “My Favorite
Option” approach that directs an agent to choose an option that has maximal prob-
ability of being objectively permissible.14 In the absence of any information about
rankings of relative choiceworthiness among permissible options or among impermis-
sible options, or about cardinal differences in choiceworthiness, there can never be
any basis for rejecting the option that has the best chance of being objectively per-
missible.15 So the structural depletion view, at least in its simplest form, amounts
to MFO.
I find MFO much more plausible and congenial than either MFT or irrelevan-
tism. Nevertheless, it cannot accommodate any of our intuitions in favor of hedging,
i.e., intuitions that the surplus structure of a theory’s value assignment (and in par-
ticular, its cardinal structure) is often relevant to how we ought to act under moral
uncertainty. We have described these intuitions before, but it is worth adding one
more to drive the point home. MFO (and hence the simple version of structural de-
14Of course, irrelevantism and My Favorite Theory are still available as well, but these approaches
do not engage in any form of intertheoretic aggregation. In principle, structural depletion is also
consistent with a threshold view, a more lax alternative to MFO on which an agent is permitted
to perform any option that has a probability greater than t of being objectively impermissible (or,
if no option exceeds threshold t, the option that comes closest). This might be motivated by the
desire to avoid the intuitive over-demandingness of MFO in cases where several options carry only
a very small risk of being objectively impermissible.
15It seems plausible, moreover, that all plausible decision procedures that might be proposed in
more enriched contexts will collapse to MFO when given only binary information—for instance,
probability-weighted Borda counting or approval voting do so straightforwardly, as does stochastic
dominance, as does expectational reasoning (regardless of risk attitudes) if one represents permis-
sibility and impermissibility as arbitrary cardinal values such that the value of permissibility is
greater than the value of impermissibility. Some relatively weak methods, like statewise domi-
nance reasoning or Condorcet methods that simply issue no verdict when there is no option that a
probability-weighted majority of theories prefer to each alternative, will not converge with MFO,
but only because they are silent in many cases where MFO is not.
186
pletion) disallows hedging even in the cases where intertheoretic comparisons seem
easiest. For instance, in the Easy Case from Chapter 5, where Alice’s certainty with
respect to every proposition concerning hedonic value makes it natural to normalize
her two theories by assuming that they value hedons equally, we might imagine that
Alice assigns credences .51 to hedonism, .49 to pluralism, and is faced with options
O ⇒ 100 hedons, 0 aesthetons and P ⇒ 99 hedons, 1000 aesthetons. MFO tells
us that Alice should choose O, since it has the greater likelihood of being objec-
tively permissible (.51 to .49), contrary to what seems like overwhelming reason for
thinking that she should choose P.
Some versions of the structural depletion view might allow Alice to hedge
so long as she distributes her belief only over theories with the right structure
(presumably, only over theories with cardinal structure of one sort or another), but
structural depletion must deny that Alice ought to hedge if we modify the case to
give her vanishingly small but positive credence in every moral theory (such that her
credences in hedonism and pluralism are reduced to .51− ε and .49− ε respectively,
for some very small ε). I conclude, therefore, that the structural depletion view is
unsatisfactory.
6.2.2 Multi-Stage Aggregation
This leaves us with the intermediate option that I have called multi-stage ag-
gregation.16 This approach is best illustrated by means of an example. Suppose
16This idea is introduced by MacAskill, who calls it a “multi-step procedure” (MacAskill, 2014,
p. 118).
187
that Dorothy divides her credence between four theories: the two consequentialist
theories, hedonism and pluralism, from the Easy Case, and two ordinallly struc-
tured deontological theories that disagree about whether one ought to kill innocent
threats, but are otherwise in complete agreement. Dorothy must decide whether
to kill Eustice: Eustice was innocently driving his truck down the street when his
brakes failed, having been sabotaged by some nefarious wrongdoer, and his truck
is now careening unstoppably toward five innocent people trapped in a narrow al-
leyway. The only way to save the five is for Dorothy, perched on a nearby rooftop,
to destroy Eustice’s truck with a bazooka. In the truck with Eustice, moreover,
are a dozen priceless works of art and the only extant score of a newly rediscov-
ered Beethoven symphony. If Dorothy opens fire, therefore, both Eustice and his
truckload of aesthetons will be destroyed, but five innocent lives will be saved.
The death of an innocent person, let’s assume, amounts to a loss of 20 hedons,
while the loss of the artistic contents of Eustice’s truck would amount to a loss of 10
hedons.17 However, their destruction would also amount to a loss of 200 aesthetons,
so while Dorothy’s hedonistic theory (T1) supports killing Eustice to save the five,
her pluralistic theory (T2) opposes it. Of Dorothy’s deontological theories (T3−4),
meanwhile, one claims that it is permissible (indeed, obligatory) to kill Eustice and
impermissible not to, while the other claims that it is permissible (“”) not to kill
Eustice and impermissible to kill him. (For now, we leave it unspecified whether
the structure of these theories is binary or ordinal.) Table 6.1 represents Dorothy’s
17Suppose that, although people will derive enjoyment from these works, that enjoyment is
highly substitutable, so that if Eustice’s truck is destroyed, the would-be appreciators of the art it
carries would be able to derive nearly as much pleasure from the appreciation of other artworks.
188
Credence O1 (kill) O2 (don’t kill)
T1 (hedonism) .3 −30 −100
T2 (pluralism) .3 −230 −100
T3 (Kill IT’s.) .25 permissible impermissible
T4 (Don’t kill IT’s.) .15 impermissible permissible
Table 6.1: Dorothy’s Dilemma, Stage 1
credences and the value assignments of each theory to the options of killing and not
killing Eustice.
T1 and T2 have the same cardinal structure and both endorse expected value
as the correct aggregation procedure in the face of uncertainty, so on a multi-stage
approach, they should be aggregated expectationally18: The expected value of O1,
relative to the comparability class of theories that includes T1 and T2, is (.3×−30)+
(.3×−230) = −78, and the expected value of O2 is (.3×−100)+(.3×−100) = −60.19
Therefore, as far as Dorothy’s consequentialist theories are concerned, she ought to
choose O2 (i.e., not kill Eustice).
But how do we aggregate these consequentialist theories with the deontolog-
ical theories T3 and T4? If we assume that the deontological theories have merely
binary structure, the multi-stage view gives a straightforward answer: We take the
comparability class C1 = {T1, T2}, along with the theories T3 and T4, as inputs to
18What the multi-stage approach entails, strictly, is that comparability classes of theories should
be aggregated according to whatever aggregation procedure is appropriate given their shared struc-
ture and content. For now I assume that the appropriate aggregation procedure for cardinal theories
is expectational, at least when the theories themselves endorse an expectational decision procedure.
But we will return to this question in §6.4 and Chapter 7.
19It makes no difference whether we use the agent’s unconditional credences or her credences
conditional on the comparability class (that is, conditional on T1 ∨ T2), though the latter would
conveniently allow the credences to sum to 1.
189
Credence O1 (kill) O2 (don’t kill)
C1 (consequentialism) .6 impermissible permissible
T3 (Kill IT’s.) .25 permissible impermissible
T4 (Don’t kill IT’s.) .15 impermissible permissible
Table 6.2: Dorothy’s Dilemma, Stage 2
a binary aggregation procedure. Since, I have claimed, the aggregation procedure
appropriate to binary-structured theories is MFO, that means taking the aggregated
evaluation of C1, reducing it to only its binary substructure, and taking that as an
input to MFO. Table 6.2 represents this second stage of the aggregation procedure.
Since Dorothy’s aggregate credence in views according to which it is permissible to
kill Eustice is .25, while her aggregate credence in views according to which it is
permissible not to kill Eustice is .75, we conclude that she ought not kill Eustice.20
Note that the two-stage procedure we have just described is not equivalent
to MFO simpliciter : According to MFO, Dorothy ought to choose the option that
has the greatest subjective probability of being objectively permissible, and this is
O1, i.e., killing Eustice: The fine-grained theories that prefer this option, T1 and T3,
command a combined credence of .55 while the theories according to which Dorothy
ought not kill Eustice, T2 and T4, command a combined credence of only .45. The
source of the divergence is that the inputs to the final, MFO-like aggregation pro-
cedure are not all maximally fine-grained theories: C1, treated in this second stage
20I assume that T3 and T4 have merely binary structure for the sake of simplicity. However,
given the stipulation that there are only two options, any plausible decision procedure for ordinal-
structured theories (Condorcet methods, Borda counting, approval voting, etc) will work just like
MFO when it is applied at the second stage of the decision procedure. So all that the example
really requires is that T3 and T4 have the same structure, either binary or ordinal.
190
as a single theory, subsumes the already-aggregated value assignments of two fine-
grained theories. From here on, let’s call the pure version of MFO that I claimed
was entailed by structural depletion fine-grained MFO, MFO-like aggregation pro-
cedures like the one just described in the case of Dorothy coarse-grained MFO, and
allow “MFO” without qualification to refer to both fine-grained and coarse-grained
MFO.21
The attraction of the multi-stage approach is that it preserves sensitivity to
the relevant structure of richly-structured (e.g., cardinal) theories without artificially
imposing this structure on less-structured (e.g., ordinal) theories: In Dorothy’s case,
this means that the cardinal values assigned by T1 and T2 are allowed to make a
difference, without implicitly or explicitly cardinalizing the non-cardinal theories T3
and T4. On the other hand, in order to capture these virtues of structural enrichment
and structural depletion respectively, the multi-stage view must open itself up to
21Note that MFO conceived as a decision procedure only for dealing with moral or normative
uncertainty is already coarse-grained. For instance, suppose an agent A has .6 credence in utilitari-
anism and .4 credence in Kantianism, and must choose between two options, O and P. Kantianism
implies that O is morally prohibited and P is morally obligatory. O has greater expected utility
than P, and hence would be the more rational choice if A were certain of utilitarianism. But
A has empirical uncertainties as well as normative uncertainties, and conditional on the truth of
utilitarianism, she judges that there is a .3 probability that P would in fact produce more utility
than O (as compared with a .7 probability that O would produce greater utility than P). Thus
the probability that P is objectively permissible is .58 while the probability that O is objectively
permissible is just .42. But A’s total credence in moral theories that “prefer” O is .6 while her
total credence in moral theories that prefer P is only .4.
Fine-grained MFO, which instructs A simply to choose the option with the highest probability
of being objectively permissible, would thus tell her to choose P. This may seem counterintuitive,
especially given certain specifications of the possible utilities that might result from A’s choice, since
it simply disregards relevant cardinal information (for instance, suppose there’s a .3 probability
that P would produce just one more utile than O, but a .7 probability that O would produce 1000
more utiles than P). But on the other hand, coarse-graining at the level of moral theories sacrifices
the attractive simplicity of fine-grained MFO, and will seem arbitrary absent some motivation for
cutting off cardinal aggregation at the boundary between empirical and normative uncertainty: If
we are willing to make cardinal comparisons between the objective value of acts given different
states of the empirical world, conditional on a particular moral theory, why not also make cardinal
comparisons between the objective value of acts given different moral theories, at least so long as
such comparisons are possible (e.g., within minimal comparability classes)?
191
two corresponding lines of attack.
First, it may seem that the multi-stage view is insufficiently sensitive to the
relatively enriched structure of theories like T1 and T2. MacAskill, who previously
endorsed a version of multi-stage aggregation (Crouch, 2010), repudiates the view for
this reason in his (2014). He describes a case in which an agent has (i) slightly more
than .5 credence in a pair of cardinal theories that disagree over which of options
O1 and O2 has greater value but, when aggregated, assign slightly greater expected
value to option O1, and (ii) slightly less than .5 credence in an ordinal theory that
prefers O2 to O1. A multi-stage aggregation approach yields the conclusion that the
agent should prefer O1. MacAskill finds this implausible. He points out in particular
that the agent might come to prefer O2 by reducing her credence in the cardinal
theory that prefers O2 and redistributing that credence proportionately over her
other moral theories, if this gave her a credence in the ordinal theory greater than
.5. This violates a criterion MacAskill calls “updating consistency,” to the effect
that increasing one’s credence in a theory T that regards option O as maximally
choiceworthy, while proportionately reducing one’s credence in all other theories,
should never make O cease to be the most rational option, if it was so initially
(MacAskill, 2014, pp. 117-9).
There is an intuitive cost here, to be sure, but to my mind not as great a
cost as MacAskill supposes. The fact that, in the initial description of the case,
O1 has only “slightly greater” expected value than O2 according to the comparabil-
ity class of cardinal theories should make no difference when the cardinal theories
are aggregated together with ordinal theories: Since ordinal value differences don’t
192
come in degrees, any sense that the value difference between O1 and O2 is greater
according to the ordinal theory than according to the aggregated cardinal theories
is simply misleading. And the principle of updating consistency, though intuitive, is
not sacrosanct: Multi-stage aggregation does seem to generate threshold effects, e.g.
that whenever an agent has credence greater than .5 in a merely-ordinal theory that
regards option O as most choiceworthy, then she is rationally required to choose
option O, and these threshold effects can create violations of updating consistency,
as MacAskill’s case illustrates. But the multi-stage view explains why these thresh-
old should exist and why we should expect updating consistency to sometimes be
violated. Finally and most importantly, the costs of the multi-stage approach must
be compared with the costs of structural enrichment and structural depletion, which
strike me as considerably greater.
A second objection to multi-stage aggregation is as follows: It seems arbitrary
that the fine-grainedness of theories or classes of theories taken as inputs to an
aggregation method should effect the outcome of that aggregation method. To
avoid such arbitrariness, we should therefore either limit ourselves to aggregation
methods that are insensitive to “theory individuation” (that is, insensitive to how
we group sets of fine-grained theories into coarser-grained theories) or else, whenever
we are forced to rely on such methods, take only maximally fine-grained theories as
input.
Two replies: (1) The coarse-graining of theories into comparability classes is
not arbitrary. T1 and T2 are grouped together by the multi-stage aggregation proce-
dure described above because they share common structure that T3 and T4 do not.
193
Dropping for a moment the technical sense of “theory” as a maximal consistent set of
propositions, one might regard them as variants of a single theory, viz., consequen-
tialism. More particularly, they are aggregated before the second, coarse-grained
MFO stage of the aggregation procedure because they possess relevant structure
(viz., cardinal structural) to which MFO is insensitive, while the binary-structured
T3 and T4 do not.
(2) There is, as far as I can see, no way of satisfy the demands of this response
without falling into the greater difficulties associated with structural enrichment or
structural depletion: If there are theories that genuinely have only binary structure,
then the only procedure available for aggregating such theories is MFO, which is
sensitive to theory individuation.22 But if we take maximally fine-grained theories as
inputs to such a decision procedure, then we are left with fine-grained MFO or some
other method that disregards cardinal information entirely and therefore precludes
moral hedging.23
I have no general proof to offer that structural enrichment, structural deple-
tion, and multi-stage aggregation are the only options for solving the problem of
22Likewise, any ordinary voting method that might be used to aggregate merely-ordinal theories
will be sensitive to theory individuation—for instance, since Borda counting in a two-option case is
equivalent to MFO, a probability-weighted Borda count of T1, T2, T3, and T4 will choose O1 while
a probability-weighted Borda count of C1, T3, and T4 will choose O2. MFO itself might be used
as a method of aggregating ordinal as well as binary theories, since it is the analogue of approval
voting in the social choice context.
23Note again that the “standard” version of MFO is not maximally fine-grained, since it ag-
gregates by non-MFO-like procedures within moral theories (e.g., aggregating different possible
consequences of a course of action expectationally within consequentialist theories) before engag-
ing in MFO-like aggregation across theories. This, I have argued, is at least as arbitrary a form of
coarse-graining as that involved in multi-stage aggregation. But the alternative, for the principled
advocate of fine-grained MFO, is a complete insensitivity to scale of risk: For instance, if an agent
A is almost certain that utilitarianism is true, and must choose between options O and P, where O
has a .51 probability of producing slightly bettter consequences than P but a .49 probability of pro-
ducing catastrophically worse consequences (say, ending all life on Earth), maximally fine-grained
MFO requires A to choose O, which is quite implausible.
194
structural diversity. Nevertheless, in the absence of a better options, and given the
difficulties of structural enrichment and structural depletion, I conclude that multi-
stage aggregation is the right approach to take: Different classes of moral theories
with differently structured value scales should be aggregated separately by methods
appropriate to the shared structure of each class, and then aggregated with one
another in ways that make use of only the (perhaps very minimal) structure that all
the classes being aggregated have in common. I have only described one simple ex-
ample of how such a multi-stage aggregation procedure might go—in more complex
cases, the procedure might involve several stages, and many distinct comparability
classes of theories (individuated, for instance, by cardinal vs. ordinal structure,
perhaps interval vs. ratio structure, transfinite structure, dimensionality, etc) that
are aggregated independently before the final stage.
6.3 Scale Normalizability
When the value scales of two theories share the same structure, or can be jus-
tifiably re-represented as sharing the same structure (by enriching and/or depleting
the structures of one or both theories, either straightaway or in the course of a multi-
stage procedure), the next question is how to normalize those scales—that is, which
points and/or intervals on one scale correspond to which points and/or intervals
on the other. Consider, for instance, hedonistic and preference utilitarianism, two
straightforward maximizing consequentialist theories that agree on every feature of
morality, except that hedonistic utilitarianism regards pleasure and pain as the sole
195
non-derivative bearers of moral value while preference utilitarianism regards satis-
fied and dissatisfied preferences as the sole non-derivative bearers of moral value.
Both theories, we may stipulate, have the same cardinal structure. But this struc-
ture does not answer the crucial question for expectational reasoning, how the value
of a hedon according to hedonic utilitarianism compares to the value of a preference
utile according to preference utilitarianism—that is, for an agent who divides her
beliefs equally between the two theories and wishes to hedge when they conflict, how
much hedonic experience does it take to offset the dissatisfaction of a preference of a
given strength (or vice versa)? Likewise, of course, in trolley problem situations that
pit consequentialist and deontological theories against one another, even if we could
overcome the apparent structural incompatibility of these rival theories, the thorni-
est question seems to be: How many net lives must be saved, according to some
particular version of consequentialism, to offset the wrongness of killing an innocent
person, according to some particular version of deontology? It is this problem of
scale normalization that has attracted most of the attention in the extant literature
on intertheoretic value comparisons.
In fact, there are several distinct challenges involved in scale normalization
that are rarely distinguished in the literature. Though I doubt that this accounting
is exhaustive, three problems of particular importance are unit normalization, level
normalization, and dimensional normalization. Unit normalization consists in a
mapping between intervals on the value scales of two theories that designates pairs
of intervals as equal in magnitude—for instance, designating any interval of one
hedon on the value scale of hedonistic utilitarianism as equal in magnitude to any
196
interval of three preference utiles on the value scale of preference utilitarianism.
Level normalization consists in a mapping between points on the value scales of
two theories that designates pairs of points as equal in absolute value—for instance,
designating simple omissions (e.g., refraining from any bodily motion) as having the
same value according to a consequentialist and a deontological theory. Dimensional
normalization consists in a mapping between dimensions of the value scales of two
multi-dimensional (i.e., lexical or incomparabilist) theories—for instance, mapping
the sole dimension of a hedonistic utilitarian value scale onto one of the dimensions of
a “leximin” egalitarian theory (on which the wellbeing of less well off social groups is
represented by lexically prioritized dimensions on a multi-dimensional value scale).24
These components of scale normalization are not independent, of course. Where
multi-dimensional theories are in play, unit or level normalization requires dimen-
sional normalization. Given two theories with straightforward cardinal structure,
establishing two or more points of level normalization between the theories (i.e.,
two or more pairs of corresponding points on the value scales of the two theories
24MacAskill (2014) discusses another such case: Suppose I divide my beliefs between classical
utilitarianism and a lexically structured deontological theory on which certain moral reasons (e.g.
to keep one’s promises or refrain from telling lies) are lexically stronger than other kinds of reasons
(e.g. prudential reasons or reasons of beneficence). Suppose further that both dimensions of the
deontological theory’s value scale have the same cardinal structure as the single dimension of the
utilitarian theory’s value scale—in order to represent the judgment, for instance, that telling two
lies is twice as bad as telling one lie. In attempting to normalize the utilitarian and deontolog-
ical value scales, then, the first question we must ask is whether utilitarian moral values should
be normalized with the lexically prioritized or the lexically deprioritized dimension of the deon-
tological value scale. The latter approach may seem natural if one thinks of utilitarian reasons
as corresponding to reasons of beneficence in the deontological schema, but this approach also
has the counterintuitive consequence of giving potential deontological reasons, e.g. for promise-
keeping, absolute priority over all utilitarian moral considerations under moral uncertainty, even if
my credence in deontology is quite low. MacAskill claims, and I agree, that it is non-obvious which
dimension of the deontological value scale the utilitarian value scale ought to be normalized with in
this case, and suggests that an agent in such circumstances should be represented as dividing her
beliefs between two versions of utilitarianism that normalize with the upper and lower dimensions
of the deontological value scale respectively (MacAskill, 2014, pp. 47-50).
197
that are equal in absolute value) will suffice to establish unit normalization as well.
But the various aspects of normalization can also come apart in interesting ways.
For instance, in the Easy Case from Chapter 5, the hedonic element common to
Alice’s hedonistic and pluralistic theories allowed us to establish unit normaliza-
tion between the theories (both theories assigning equal value to an increment of
one hedon) without in any obvious way yielding a level normalization of the two
theories. On the other hand, one plausible way of establishing a partial level nor-
malization between (at least) aggregative consequentialist theories is to hold that
all such theories assign the same value (namely, 0) to an “empty universe” that
contains no value-bearers at all. If true, this establishes one point of level normal-
ization between apparently disparate consequentialist theories like hedonistic and
preference utilitarianism, but in the absence of another such point, does not yield
unit normalization.
The significance of these various forms of normalization depends to some extent
on what decision procedures we use to aggregate theories. Risk-neutral expectational
reasoning requires only unit normalization, not level normalization.25 Risk-weighted
expectational reasoning (whether risk-seeking or risk-averse) requires both unit and
level normalization.26 Stochastic dominance reasoning, on the other hand, requires
25This is equivalent to saying that it requires only interval-scale comparability, i.e., comparability
with respect to the interval-scale properties of the theories’ value scales, without regard to any
possible ratio-scale properties.
26Suppose A has equal credence in two theories, T1 and T2, and is faced with two options, O
and P. T1 assigns O a value of x and P a value of x+ 5. T2 assigns O a value of y and P a value
of y − 5. Assume that these values are unit-normalized, so that the difference in value between O
and P is the same according to the two theories, though they disagree about which is better. Now
suppose that A wishes to hedge for her moral uncertainty, not by simply maximizing expected
value, but by maximizing risk-weighted expected value, with a slight aversion to risk. To do this
she will need to make level comparisons as well as unit comparisons between T1 and T2: If x = y,
for instance, then O is risk-free while P is risky, so A should choose O. But if x = y − 10, then O
198
only level normalization—and though in the context of cardinal theories, the full
level normalization required for stochastic dominance implies unit normalization as
well, this point becomes relevant in the context of ordinal theories for which unit
normalization is impossible (pace MacAskill’s Borda counting approach) but level
normalization is possible.27
Beginning in the last chapter, I have argued for the idea of content-based
aggregation, including content-based normalization: e.g., that the hedonistic and
pluralistic theories in the Easy Case can be unit-normalized in virtue of their shared
understanding of hedonic value, or that hedonistic and preference-based utilitarian-
ism can be level-normalized at a shared zero (the empty universe) based on the the
aggregative character of each theory. But even if such an approach is correct in the
cases so far described, it is unclear how far this approach can generalize: e.g., can
anything in the content of classical utilitarianism and Kantianism provide a basis for
an adequate normalization of their respective value scales? Treating this as an open
question, there are several candidates for a general approach to the normalization
problem.
1. Universal Content-Based Normalization (UCBN) First appearances
notwithstanding, there are in fact hidden bases for content-based comparisons
between even such disparate moral theories like Kantianism and utilitarianism—
bases analogous to, but presumably much subtler than, the shared understand-
is risky while P is risk-free, so A should choose P.
27Thus, while in the context of cardinal theories the most salient problem seems to be that of
unit normalization (given a background commitment to risk-neutral expectational reasoning), in
the context of ordinal theories the most salient problem is likely to be level normalization.
199
ing of hedonic value that provided a basis for normalization in the Easy Case.
2. Incomparability In some cases (like, perhaps, Kantianism and utilitarian-
ism), normalization is simply impossible. Rather than attempting to solve
the normalization problem in these cases, we should instead ask what decision
procedures are appropriate in the face of incomparability.
3. Binary Normalization Sufficiently disparate theories can be normalized only
in the most coarse-grained manner: namely, by treating permissibility and
impermissibility as equivalent designations across all theories, and thus as
providing a very weak form of level normalization. Thus, as per the structural
depletion and multi-stage views discussed in the last section, the aggregation
of a maximally diverse set of moral theories must in one way or another involve
some form of MFO (or threshold procedure), since these are the only options
left open by the minimal form of normalization possible among a full range of
moral theories.
4. Top-Down Normalization When nothing in the content of theories T1 and
T2 grounds a particular normalization between them (and perhaps even when
the content of T1 and T2 seems to ground a particular normalization), the
theories should be normalized according to some general, content-independent
normalization principle like Lockhart’s PEMT.
As in the last section, our task is now to spell out and compare these options.
My conclusions with respect to the normalization problem will remain more agnos-
tic, however, than my conclusions with respect to the problem of structural diversity.
200
The picture that I will argue for is this: (i) There is at least some reason to hope
that the UCBN strategy will succeed, i.e., that a sufficiently careful examination of a
rational agent’s credal state will generally, and perhaps always, reveal bases for some
form of normalization between any two normative theories in which she has positive
credence. Even if content-based normalizability proves to be not-quite-universal, its
failures might still be confined to a small enough number of cases that incompara-
bility/binary normalization (and hence, some decision procedure like MFT, MFO,
or a threshold view) would yield intuitive acceptable results if taken as the final
stage of a multi-stage decision procedure, to be invoked only when content-based
normalization gives out. (ii) Given this hope, and our intuitions in favor of moral
hedging even between very disparate theories like Kantianism and utilitarianism,
we should be reluctant to adopt incomparability or binary normalization as general
responses to the normalization problem (i.e., as the default state of affairs once we
move beyond minimal comparability classes). (iii) If our desire to vindicate hedging
even in difficult cases is strong, and our progress in finding plausible content-based
grounds for normalization in those difficult cases turns out to be disappointing, then
we may have reason to adopt a top-down normalization principle. The most plau-
sible such principles can be understood as explications of what MacAskill calls the
“principle of equal say” (PES) (MacAskill, 2014, pp. 101ff), according to which
the weight or influence of a moral theory on an agent’s choices should be in some
sense proportionate to her credence in that theory (at least, where no other basis
for normalization exists). However, though it has some intuitive appeal, PES is less
well motivated than content-based normalization methods like that illustrated by
201
the Easy Case. Additionally, as we will see, there are many possible precisifications
of the idea of “equal say,” varying along at least two dimensions, and the apparent
arbitrarity of any particular choice of top-down normalization method further di-
minishes the attractiveness of this approach. Thus, I will tentatively conclude that
we should prefer to extend the reach of content-based normalization methods inso-
far as this proves possible, but that it is worth continuing to investigate top-down
normalization methods as well in case content-based normalization proves unable to
yield intuitively satisfactory results in too great a number of difficult cases.
6.3.1 Extending Content-Based Normalization
The Easy Case from Chapter 5 illustrates a simple basis for normalization
that we may call content overlap: Alice’s hedonistic T1 and pluralistic T2 include all
the same propositions concerning the nature, explanation, and grounding of hedonic
value, and from this we are able to conclude that they assign the same degree of
value to a given unit of hedonic experience.28
On its face, content overlap does not seem to solve much of the normalization
problem, since most pairs of moral theories do not seem to overlap in the neces-
28See Appendix B for a more complete explanation of the Easy Case and in particular of how
Alice’s doxastic state grounds the intuitive normalization of her two moral theories.
In conversation, William MacAskill has described the approach that this case exemplifies as a
“shared rightmakers” approach. But it is important to remember, as I noted in Chapter 5, that the
correctness of normalizing Alice’s theories at the value of a hedon depends on their agreeing, not
just on the identity of the rightmakers for certain acts (viz., the hedonic consequences of acts that
have no positive or negative aesthetic consequences) but also on the explanation of their status as
rightmakers (e.g., whether the moral significance of hedonic experience is a basic, mind-independent
evaluative fact or a consequence of conventions within some particular moral community). When
two theories both treat hedonic consequences as rightmakers, but give different explanations of
their status as rightmakers, there is no longer any reason to normalize the theories at the value of
a hedon.
202
sary way, sharing a complete explanation of some category of moral considerations
or reasons that can serve as a basis for normalization. But there are a few ways
in which the content overlap approach can be generalized. The first is by finding
pluralistic theories to serve as “bridges” between non-overlapping theories. This is
illustrated by a simple variant of the Easy Case in which Alice has some credence
in a monistic theory that only values aesthetic goods, alongside her hedonistic and
pluralistic theories. Given that the pluralistic theory fully overlaps with the hedo-
nistic theory with respect to hedonic goods, and fully overlaps with the aesthetic
theory regarding aesthetic goods, its value scale can be normalized with each of these
monistic theories, and by extension, the monistic theories can be normalized with
one another.29 By the same sort of reasoning, an agent who has positive credence
in both hedonistic and preference utilitarianism might have content-based grounds
for a particular normalization of the two theories, if she also has positive credence
in a pluralistic theory that attaches moral significance both to hedonic experience
and the satisfaction of preferences and that regards these goods as comparable with
one another.
In principle, the bridging approach seems to have extremely wide applicability.
Let’s introduce a new piece of terminology and call the propositions that identity
and explain a particular feature of the world (like hedonic experience or the non-
universalizability of an agent’s maxims) as a source of moral value or reasons a
value claim. Any non-deflationary moral theory (i.e., anything other than nihilism,
error theory, and the like) will include one or more such value claims. For any
29See Appendix §B.4.1 for a more complete defense of this sort of “indirect comparability.”
203
pair of theories T1 and T2, if T1 includes a value claim C1 and T2 includes a value
claim C2 such that C1 and C2 are logically compatible (i.e., don’t jointly imply
a contradiction), then there is some logically possible pluralistic theory T3 that
includes their conjunction, C1 ∧ C2. In fact, there are almost certain to be many
such theories, characterized by different exchange rates between C1-value and C2-
value. But if an agent can aggregate her credence in such theories (i.e., her credences
conditional on C1∧C2) to yield a single exchange rate between the two sorts of value,
then this may enable comparisons between T1 and T2 by providing a basis, in the
agent’s doxastic state, for normalizing their value scales (at the aggregated exchange
rate between C1-value and C2-value).
30
There are a few reasons to worry about how far this approach can really get
us, however. First, some value claims may simply be logically incompatible, such
that no consistent pluralistic combination of them is possible. For my part, I find
it hard to think of many plausible cases of outright incompatibility between value
claims, but as a toy example, consider the Kantian value of autonomy and the
Buddhist value of recognizing the non-existence of the self (anatta¯, one of the seven
“beneficial perceptions” (Bhikkhu, 2013)). If the Kantian explanation for the value
of autonomy presupposes the existence of the self and the Buddhist explanation for
the value of anatta¯ presupposes the nonexistence of the self (and if we set aside the
possibility that this is a merely-verbal dispute), then no pluralistic combination of
the two value claims is possible.31
30See §B.4.2 for an explanation of how such aggregation over possible normalizations of two
theories might be carried out.
31Nonetheless, indirect normalization between theories that value Kantian autonomy and theo-
ries that value the recognition of anatta¯ might still be possible, e.g. if each of these value claims
204
Second, for some value claims, the most plausible pluralistic theories that com-
bine them may treat them as incomparable. For instance, if one were persuaded
by the positive value claims of both Nietzschean perfectionism and classical utili-
tarianism, one might do best to conclude that these two value schemes represent
equally valid but irreconcilable normative perspectives rather than weighable inputs
to a single, cardinal value function (a conclusion along the lines of Sidgwick’s (1874)
“dualism of practical reason”).
Third and finally, the most plausible combination of certain value claims might
be lexical. Certain value claims involve, in their self-justificatory content, assertions
of lexical priority over other sorts of value claims. The Kantian explanation of the
value of conformity with the categorical imperative, for instance, seems to essentially
involve an assertion of lexical priority over all instrumental reasons (“hypothetical
imperatives”) and hence over the value claim of any moral theory whose value
claims reduce to instrumental reasons (e.g. the contractarianism of Hobbes (1651)
and Gauthier (1986) or the relativism of Harman (1975)). Thus, someone who
was fully convinced by all the positive claims in both the Groundwork and Morals
by Agreement, and therefore accepted a pluralistic hybrid of Kant’s moral theory
and Gauthier’s, would presumably conclude that Kantian reasons, where operative,
take lexical priority over contractarian moral reasons (as well as other instrumental
reasons). The value claims of a divine command theory of morality, plausibly, involve
claims to lexical priority over all other value claims whatsoever, so that if any
figures in a different pluralistic theory along with come third value claim C3 that allows the two
theories and hence the incompatible Kantian and Buddhist values to be normalized.
205
pluralistic hybridization of divine command theory with other ethical perspectives
is logically possible, the result would be a lexically structured theory within which
God’s commands have absolute force.
The first two of these worries (incompatibility and incomparability) suggest
that the bridging approach may have limited reach. The third worry suggests a
broader objection, that content-based normalization approach in general is vulnera-
ble to fanaticism. Suppose we conclude that a pluralistic hybrid of Kantianism and
contractarianism would give lexical priority to Kantianism, and on this basis con-
clude that an agent who has positive credence in Kantianism, contractarianism, and
this pluralistic hybrid ought to give lexical priority to Kantianism as well. This may
seem intuitively implausible—one is tempted to say, it seems “unfair” to contrac-
tarianism to grant it no weight at all when it conflicts with Kantianism. Certainly
it is intuitively implausible that an agent who has .999 credence in contractarianism
and .001 credence in Kantianism, or divine command theory, or some other similarly
self-assured moral theory, should give absolute priority to the theory she regards as
almost certainly false over the theory she regards as almost certainly true.
I am willing to bite the bullet on this objection, up to a point: Some value
claims may simply be more intrinsically weighty than others, and in some cases
absolutely so. In cases where the agent’s credence in the lexically prioritized value
claim approaches zero, however, the situation begins to resemble Pascal’s Wager
(Pascal, 1669), the St. Petersburg Lottery (Bernoulli, 1738), and similar cases of
extreme probabilities and magnitudes that bedevil decision theory in the context
of merely empirical uncertainty. It is reasonable to hope, then, that the correct
206
decision-theoretic solution to these problems (e.g. a dismissal of “rationally negligible
probabilities” (Smith, 2014, 2016) or general rational permission for non-neutral risk
attitudes (Buchak, 2013)) will blunt the force of the fanaticism objection.
Another way of extending the content overlap approach to normalization, apart
from pluralistic bridging, is to consider the “sub-moral” content of moral theo-
ries: that is, the underlying theories of reasons or normativity on which they are
built. One region of logical space where this idea seems particularly promising is
the Humean theory of reasons and the various moral/normative theories built on
top of it. These latter include the aforementioned views of Hobbes, Gauthier, and
Harman, many “commonsense” moral views like that of Bernard Williams, Philippa
Foot’s erstwhile “hypothetical imperatives” view (Foot, 1972), and most anti-realist
or “quasi-realist” approaches to morality. Indeed, if I was right to claim in Chap-
ter 1 that the Humean and enkratic conceptions of rationality exhaust the space
of possibilities, and if (as seems plausible) the Humean view commands at least as
many adherents as the enkratic view, then a majority of the moral theories actually
embraced by philosophers may turn out to be, in the last analysis, Humean theories,
in which an agent’s moral reasons derive in one way or another from the contents
of her motivational set.
There is clear potential for a general content-based normalization of Humean
moral theories, since all such theories share a “common currency,” viz., Humean
reasons and the desires/motivational states from which such reasons arise. Ceteris
paribus, it seems reasonable to normalize such theories by assuming that they all
assign the same degree of normative significance to the direct satisfaction of a given
207
non-moral desire (at least, when it neither conflicts with nor results in the satis-
faction of any other element of the agent’s motivational set). Since Humean moral
theories presumably will not treat moral reasons as lexically or infinitely stronger
than non-moral reasons, normalizing these theories at the strength of the non-moral
reasons about which they have no disagreement should yield a satisfactory normal-
ization of their moral value scales as well. For instance, suppose an agent A divides
her beliefs between Gauthier’s contractarianism and Harman’s relativism, and faces
a situation in which a Gauthier-norm N conflicts with a Harman-norm M. Sup-
pose that Gauthier’s view and Harman’s view are in full agreement, say, about the
strength of A’s entirely non-moral reason to accept a free ice cream cone. And sup-
pose the contractarian view says that her reasons to comply with N in this situation
are 150 times stronger than her reason to accept a free ice cream cone, while the
relativist view says that her reasons to comply with M in this situation are just 50
times stronger than her reasons to accept a free ice cream cone. Then it appears
that the value of complying with norm N according to contractarianism is three
times the value of complying with norm M according to relativism.32
32In some cases of intra-Humean disagreement, it may seem that an apparent disagreement
between “moral theories” is actually an empirical disagreement, namely, a simple disagreement
about what course of action would best satisfy an agent’s desires. In some simple cases this may
be true: A simpleminded Humean moral theory can be constructed from the Humean theory of
reasons plus the observation that a typical human being has some broadly moral motivations,
like empathy, which on the Humean theory of reasons she can some reason to satisfy. “Moral”
disagreements can then arise from rival theories of moral psychology, i.e., of what the exact content
of those moral motivations or concerns amounts to. But many intra-Humean disagreements are
genuinely normative, in that they involve disagreements about exactly how an agent’s reasons
relate to her desires and other motivational states (e.g. what forms of practical deliberation from
an agent’s existing motivational state result in the discovery of preexisting reasons, as opposed
to alterations in an agent’s motivational state that simply give her new reasons she didn’t have
before) or about the relationship between rationality and objective reasons (e.g. the question of
“straightforward maximizing” vs. “constrained maximizing” taken up by Gauthier (1986)).
208
Can we say anything similar about non-Humean views—i.e., is there some
“common currency,” analogous to the currency of desire, in which non-Humean
views must trade, that would enable a general content-based normalization of their
value scales? This is non-obvious. I claimed in Chapter 1 that the Humean view
identifies objective reasons with the satisfaction of desires or other motivational
states while the enkratic view identifies objective reasons with goodness, value, or
choiceworthiness, where these concepts are at least not conceptually linked to desire
satisfaction. To simplify, let’s say that enkratic views identify objective reasons with
desire-independent choiceworthiness. On face, this conception of reasons does not
offer the same potential for a general normalization of moral theories as the Humean
conception. For while the strength of desires is an empirically measurable feature
of the world, and therefore all normative theories that identify reasons with desires
and strength of reasons with strength of desires can be put on a common scale, there
is no similar way of measuring degrees of desire-independent choiceworthiness that
would generate a common scale of reason strength for non-Humean theories.
One strategy that might allow a general normalization of non-Humean theories
is suggested by MacAskill (2014). MacAskill argues that all moral theories can be
understood as having as the range of their value assignments the same “universal
scale” of degrees of value. He explicates this idea of a universal scale by analogy with
the metaphysics of quantity: There is a debate among metaphysicians as to whether
mass properties (5 kg etc) or mass relations (more massive than, less massive than,
equally as massive as) are more fundamental. One reason to prefer the former,
“absolutist” position is its ability to make sense of intuitive trans-world mass rela-
209
tions (e.g., the possibility of a world identical to ours in every respect except that
everything is twice as massive). Similarly, MacAskill suggests, if we suppose that
there is a scale of absolute value properties analogous to the mass-absolutist’s scale
of absolute mass properties, then those value properties (being abstract entities)
exist at all possible worlds, including the possible worlds postulated by rival moral
theories. Thus, it seems, all moral theories should be comparable by virtue of the
comparability of the absolute value properties (all belonging to a single “universal
scale”) instantiated at the corresponding possible worlds.
To my mind, however, MacAskill’s proposal moves too quickly from an on-
tological claim about the existence of certain properties to a semantic claim that
all moral theories can be understood as making reference to those properties. Let’s
suppose that practical options in the actual world instantiate absolute value prop-
erties arranged on, say, an interval scale. Does this mean that all moral theories,
including merely-ordinal theories, multidimensional theories, hyperreal theories, etc,
in fact assign interval-scale degrees of choiceworthiness to options, despite their own
protestations to the contrary? Presumably not. Even given that a particular set of
interval-scale value properties are the ones instantiated in the actual world, it is still
possible for a (mistaken) moral theory to attribute a different set of value proper-
ties to options, properties that are not instantiated in the actual world. Further, as
MacAskill acknowledges (pp. 154-5), some moral theories may be explicitly compar-
ativist, and not attribute any absolute value properties at all. Finally, even a moral
theory that attributes absolute, interval-structured value properties to options may
not attribute properties from the same interval scale as those properties that are in-
210
stantiated in the actual world—that is, mere structural agreement between theories
does not obviously guarantee that the theories make reference to the same scale or
that their scales can be normalized.33 Of course, we might always insist that any
theory that does not attribute absolute value properties from the particular scale
instantiated at the actual world simply doesn’t count as a moral theory, but given
the structural diversity of (apparent) moral theories, if nothing else, this stipulation
would force us to implausibly exclude the large majority of what we take to be moral
theories.34
33The analogy to mass properties is importantly misleading in this respect. Suppose we accept
the absolutist theory of mass properties, and now wish to know whether “intertheoretic mass com-
parisons” are possible: For instance, can we compare the mass attributed to a given object by the
Aristotelian theory of mass with the mass attributed to the same object by the Einsteinian theory
of mass? The answer seems to be “yes,” since we are able to make uncontroversial translations
between Aristotelian and Einsteinian mass ascriptions (e.g., 1 obol = .72 grams). But these trans-
lations assume a particular normalization of the Aristotelian and Einsteinian mass scales: that
both theories agree on the mass of a certain class of objects, e.g. small metal coins. (The Aris-
totelian measures the coin and concludes that it has a mass on 1 obol, the Einsteinian measures the
coin and concludes that it has a mass of .72 grams, and since we assume that they don’t disagree
about how massive the coin is, we translate obols to grams at a rate of 1:.72.) But suppose we
took as our reference point, not small metal coins, but air-inflated party balloons. This would give
us a different normalization of the two mass scales. (The Aristotelian measures the balloon and
concludes that it has a mass of 1 obol, the Einsteinian measure the balloon and concludes that it
has a mass of, say, 2 grams.) If we assume that the Aristotelian and the Einsteinian agreed about
the mass of air-inflated party balloons, then we have to conclude that they disagree about the mass
of small metal coins, and vice versa. Intertheoretic mass comparisons between the Aristotelian and
Einsteinian theories seem possible only because of the contingent fact that there is one particularly
natural way of normalizing the scales: namely, by assuming that they agree on the mass of small
metal coins and other similarly much-denser-than-air objects. Although I have argued that there
is often a similarly natural way of normalizing value scales in a moral context (e.g., assuming that
Alice’s hedonistic and pluralistic theories agree on the value of a hedon in the Easy Case), the
mass analogy should not convince us that such a natural normalization method exists in all cases
of intertheoretic moral uncertainty. Just as the possibility of intertheoretic mass comparisons does
not follow simply from the existence of absolute mass properties, so the possibility of intertheoretic
value comparisons would not follow simply from the existence of absolute value properties.
34The argument for the universal scale view, as MacAskill presents it, relies on the claim that
absolute value properties, qua abstract objects, must exist at all possible worlds. As he subse-
quently acknowledges, the possible worlds corresponding to false moral theories are presumably
epistemically but not metaphysically possible worlds, and it is unclear why we should think that
all actual abstract entities exist at all epistemically possible worlds. (In fact, this seems clearly
wrong, since for any abstract entity, I can rationally have non-zero credence that it does not exist,
i.e., assign some credence to epistemically possible worlds where it does not exist.) MacAskill in-
terprets this point as an objection to absolutism about value properties, but this strikes me as the
wrong lesson. Rather, the point is that even if practical options instantiate absolute (i.e., unary,
211
It is unclear, therefore, why accepting (as a first-order matter) the existence of
a scale of absolute value properties should lead us to conclude that all or even most
moral theories assign values on that scale. Nevertheless, the idea that many (though
not all) moral theories might make reference to the same set of abstract entities
(absolute value properties arranged on a value scale with a particular structure)
does suggest the possibility of content-based normalizations that go beyond the
overlapping-content form of normalization illustrated by the Easy Case. It seems to
me that this approach is more promising with respect to non-Humean theories than
with respect to Humean theories, since it is more natural to understand Humean
theories as comparativist rather than absolutist (especially in light of their agent-
relative character), but it is at least conceivable that the universal scale approach
might allow us to bridge the gap between (some) Humean and non-Humean theories
as well.35
I have so far explored some positive conclusions that might be drawn about
the potential for extending content-based normalization beyond small comparabil-
ity classes like the one illustrated by the Easy Case. Are there any general negative
conclusions we can draw with confidence—i.e., definite barriers to content-based
normalization? The clearest route to such negative conclusions, as far as I can see,
non-relational) value properties at the actual world, those properties need not feature in every
moral theory—both because they may not exist at the epistemically possible worlds corresponding
to certain moral theories, but also because, even if they exist at a given possible world, they may
not be instantiated at that world.
35MacAskill concedes that the universal scale approach will not establish comparability between
all moral theories, but claims that it establishes comparability between all theories that accept an
absolutist rather than comparativist understanding of value (MacAskill, 2014, p. 155). But the
arguments I have given above suggest that this is too optimistic, since absolutist moral theories may
attribute different sets of absolute value properties, from different and potentially incomparable
value scales.
212
is the hybridization test, that is, looking for pairs of value claims that are either
logically incompatible or that seem to yield incomparability when combined to form
a pluralistic theory. As I have said, I don’t think there are many clear examples of
either kind, but I will close this discussion by noting the possibility of one very gen-
eral barrier to normalizability: the boundary between Humean and enkratic moral
theories. In Chapter 1, I postponed the question whether the Humean and enkratic
views offer rival conceptions of shared underlying concepts of reasons, rationality,
etc, or simply attach labels like “reasons” and “rationality” to fundamentally dif-
ferent concepts. Now as then, I won’t attempt to answer this question, but I raise
the question in order to observe that, to the extent one understands these views as
dealing in entirely different concepts, one might expect that content-based normal-
ization between the two categories of theory will prove impossible. For if Humean,
desire-based “reasons” and enkratic, desire-independent “reasons” are not both ar-
ticulations of some underlying normative concept, then what sense could it make to
compare them? To say that “My reason to choose option O according to [enkratic
moral theory] T1 is twice as strong as my reason to choose option P according to
[Humean moral theory] T2” would be a simple category mistake, on a par with say-
ing “The gravity of the situation on the Korean peninsula is twice as great as the
surface gravity of Ganymede.” That is, quantitative comparisons only make sense
when there are two quantities (or possible quantities) of the same something to
be compared, and unless there is some underlying concept of reasons, normativity,
or the like that unites anything labeled a “moral theory,” then there will be some
213
insuperable barriers to meaningful intertheoretic comparison.36
As the preceding discussion has illustrated, content-based normalization and
the comparability class approach to decision-making under moral uncertainty seems
to create a patchwork theory in which aggregation procedures for particular classes
of theories depend, perhaps in non-obvious ways, on the content of those theories—
not just their value assignments, but also the arguments or explanations they offer
for those value assignments, the metaethical and other philosophical premises on
which those arguments rely, and perhaps other sorts of propositions that might
figure in a complete theory of the world, in ways that we have not yet anticipated.
This contrasts to the neat simplicity of a principle like Lockhart’s PEMT. But
such simplicity is purchased at the price of a Procrustean indifference to the actual
content of the moral theories being compared and aggregated, a desire to fit those
theories into an attractive formalism even if this requires misrepresenting some of
their features (e.g., representing ordinal structure as cardinal) and ignoring others
(e.g., ignoring the overlapping content that produces natural normalizations at odds
with the general normalization principle). Thus it seems to me that content-based
normalizations should be adopted wherever possible, even if it turns out that many
sets of moral theories cannot be normalized by reference to their content. What
36This consideration raises a possibility, which I will not explore at any length, that the resolution
of the debate between the Humean and enkratic theories of reasons should be taken as having
“external” normative significance in Weatherson’s sense: If the Humean and enkratic views simply
attach the label of “reasons” to different concepts, then plausibly rationality does not require us
to account for our uncertainty between the Humean and enkratic views, but simply to account for
our uncertainties about our reasons (that is, the normative concept fundamental to whichever of
the Humean or enkratic views turns out to be correct), while ignoring whatever beliefs we might
have about schmeasons (the wrongheaded, perhaps incoherent concept that goes by the name of
“reasons” in whichever of the Humean or enkratic views turns out to be mistaken).
214
we should do in these latter cases—in particular, whether we should simply accept
defeat, or should adopt a top-down normalization procedure like PEMT—is the
question to which we now turn.
6.3.2 Normalization Failure: Incomparability/Binary Normalization
One possibility suggested by the multi-stage aggregation procedure described
in the last section is to aggregate comparability classes of theories up to the point
where content-based comparison gives out, and then perform the last stage of the
aggregation procedure by means of a norm like MFT or MFO that does not rely on
any precise intertheoretic normalization. For simplicity, let’s confine our attention
to these two possibilities: multi-stage aggregation with either MFT or MFO as the
final stage.
The latter procedure was introduced in the last section, where we distinguished
this sort of “coarse-grained” MFO from the more general “fine-grained” view on
which an agent should simply choose the option that has the highest probability of
being objectively permissible. We should now draw a parallel distinction between
“coarse-grained” and “fine-grained” MFT. Fine-grained MFT says that an agent
should always choose the option preferred by the moral theory in which she has
highest credence, where “moral theory” is understood in our familiar sense, as a
maximal consistent set of moral propositions.37 Coarse-grained MFT says that an
37It’s important to clarify, in this context, that a moral theory as opposed to a theory of the
world generally should be understood as a maximal consistent set of “purely moral” propositions,
or of general moral principles, that does not include particular injunctions like “A objectively ought
to choose option O at time t.” Otherwise “moral theories” would embed empirical information,
e.g. about which acts have the best consequences, resulting in a view more objectivist than the
proponents of MFT typically intend. To say that a moral theory “prefers” a particular option O,
215
agent should always choose the option preferred by the maximal comparability class
of moral theories in which she has highest credence, where a maximal comparability
class is a set of theories among which some form of content-based comparison and
aggregation are possible, but such that no form of content-based comparison or ag-
gregation is possible between theories inside the set and any theories outside the set.
This view says, in effect, “Aggregate theories as far as content-based comparisons
will allow, and when no further content-based comparisons are possible, do what
the most plausible comparability class of theories says, in aggregate.”
Coarse-grained MFT is, arguably, an improvement on fine-grained MFT in-
sofar as it offers a more plausible answer to the problem of theory individuation.
Gustafsson and Torpman (2014), in the most thoroughgoing defense of MFT to
date, embrace a view very close to what I have called “fine-grained MFT,” claiming
that we should “[r]egard moral theories T and T ′ as versions of the same moral
theory if and only if you are certain that you will never face a situation where
T and T ′ yield different prescriptions” (Gustafsson and Torpman, 2014, p. 13).
But as MacAskill (2014) has pointed out, this has the implausible result that a
typical agent ought to be guided exclusively by a moral theory in which she may
have almost no credence. For instance, suppose an agent has 99% credence that
some form of commonsense morality is correct and only 1% credence that classical
utilitarianism is correct. Nevertheless, among possible commonsense moral views,
she is uncertain of a great many particular, logically independent moral questions.
then, is to say that it implies that the agent has more subjective reason to choose O than to choose
any alternative, given her non-moral belief state.
216
(Exactly how wrong is it to break a promise? Is it wrong to have an abortion at
5 months? How about at 4 months 28 days? Are my obligations to my children
twice as strong as my obligations to distant strangers, or five times as strong, or
what?) Hence, her credence in any fully specified version of commonsense morality
is less than 1%. Classical utilitarianism, meanwhile, is a fully specified theory to
begin with, i.e., she has no uncertainties about which form of classical utilitarian-
ism is true that might result in conflicting practical guidance. So on Gustafsson
and Torpman’s view, she ought to always do what classical utilitarianism recom-
mends, even when every version of commonsense morality in which she has positive
credence tells her to do the opposite. Coarse-grained MFT has the potential, at
least, to correct this implausible result, for if some or all of the commonsense views
in which she has positive credence admit of content-based comparisons (say, based
on their overlapping content in the many moral cases about which they are in full
agreement), then her “favorite” maximal comparability class will plausibly turn out
to be a less-than-maximally-specified version of commonsense morality in which she
has substantial (up to 99%) credence (assuming for the sake of argument, of course,
that commonsense morality and classical utilitarianism turn out to be genuinely
incomparable).
Coarse-grained MFT seems a natural option if one holds that, between some
pairs of moral theories, no form of normalization whatsoever is possible. But there
is also a more moderate position, holding that a kind of rough normalization is
possible even between the most disparate moral theories: namely, that the shared
concepts of permissibility and impermissibility allow us to treat these designations
217
as equivalent across theories, where other forms of normalization fail. This claim, it
seems to me, is implicit in any version of MFO, fine-grained or coarse-grained, since
the basic principle of MFO is to sum up an agent’s credence in theories (of whatever
coarseness of grain) that treat a given option as permissible, giving equal weight
to the designations of each theory. Absent something like a normalization claim,
we might be left to wonder, for instance, why an option’s permissibility by a more
demanding moral theory (that regards only a few options as permissible) should not
count for more than its permissibility according to an un-demanding moral theory
(that regards many options as permissible). Thus, while this claim is not crucial to
anything else in this chapter and I am not deeply attached to it, it seems to me that
MFO involves at least a kind of normalization, namely, binary normalization of the
designations permissible and impermissible.
So we have at least two relatively pessimistic options for what to do when
content-based normalization proves impossible: (1) accept incomparability and use
coarse-grained MFT as the last stage of a multi-stage aggregation procedure or (2)
take binary normalization as the best we can get and use coarse-grained MFO as
the last stage of a multi-stage aggregation procedure. Both of these views involve
moral hedging—perhaps a great deal of hedging—within comparability classes, but
simply declare that there are some kinds of moral uncertainty for which hedging
is impossible. I have nothing to say in assessment of these views that I have not
already said: Their chief drawback lies in our intuitive judgment that it is rationally
appropriate to hedge even between widely disparate moral theories, e.g., to take
account of the utilitarian stakes when one faces a choice situation like the trolley
218
problem with beliefs divided between utilitarianism and Kantianism. Thus, we
should at least continue to explore the potential for other views—content-based or
top-down normalization approaches—to vindicate these intuitions.38
6.3.3 Top-down Normalization
If we conclude that there are some pairs of theories whose content does not
ground any particular normalization between them, but between which hedging is
nevertheless rationally appropriate, then we must have some method of normaliza-
tion that does not derive from the content of the theories being normalized. As I have
noted, the archetype of this sort of normalization method is Lockhart’s PEMT, ac-
cording to which in any choice situation, “[t]he maximum degrees of moral rightness
of all possible actions...according to competing moral theories should be considered
equal” and “[t]he minimum degrees of moral rightness...should be considered equal
unless all possible actions are equally right according to one of the theories (in which
case all of the actions should be considered to be maximally right according to that
theory)” (Lockhart, 2000, p. 84).
As I have also noted, PEMT is subject to apparently decisive objections
(Sepielli, 2013). Nonetheless, it offers a useful template from which we can gen-
erate other, more plausible normalization principles. As MacAskill (2014) observes,
there are two salient features of PEMT that can be varied to produce alternative
38It is worth noting that the multi-stage aggregation approach that I defended as a response to
the problem of structural diversity does not commit us to accepting one of these defeatist views:
We might hold that some form of hedging, or at least something richer than MFT/MFO, is possible
even at the last stage of the aggregation procedure. For instance, we might hold that the last stage
of aggregation combines cardinal, ordinal, and binary theories using a voting method richer than
MFO, e.g. a Condorcet method or (pace my objections) a Borda count.
219
views.39 First, it considers only the options available to a single agent in a single
choice situation. Second, it normalizes the value assignments of rival moral theories
in that choice situation by treating the ranges of those value assignments (i.e., the
difference in value between the best and worst options) as equal. We can then con-
struct a wide variety of alternative principles for normalizing rival value assignments
by varying either the set of options relative to which the normalization method is
defined (i.e., the domain of the value assignments being normalized) or the feature
of the value assignments to be equalized. Let’s call this first feature the scope of
a normalization method and the second feature its normalization point. Scope and
normalization point, then, are two dimensions along which top-down normalization
methods can vary.
Potential values of the scope variable include the following:
1. The options that make up a single choice situation (as in Lockhart’s PEMT).
2. All conceivable options over which a theory’s value assignment is defined (as
in Sepielli’s “Conceivability PEMT”).
3. The options that make up all (past/future/lifetime) choice situations faced by
a particular agent.
4. The options that make up all (past/future/past-present-and-future) choice
situations faced by all actual agents.
Possible normalization points include:
39The following discussion is heavily indebted to MacAskill’s analysis.
220
1. The range of each theory’s value assignment (as in Lockhart’s PEMT and
Sepielli’s Conceivability PEMT).
2. The difference between the mean value of a theory’s value assignment over the
specified domain and either the maximum or minimum value of the assignment
over that domain (approaches MacAskill calls Max-Mean and Mean-Min).
3. The variance of each theory’s value assignment (the approach endorsed by
MacAskill).
4. The standard deviation of each theory’s value assignment.
5. The mean absolute deviation of each theory’s value assignment, relative to
some specified measure of central tendency (mode, median, arithmetic mean,
geometric mean...).
Neither of these lists is exhaustive, of course. With respect to normalization
points in particular, any possible measure of dispersion, including any measure of
typical or maximal deviation from any measure of central tendency, constitutes a
feature of value assignments that might be treated as equal to achieve a normaliza-
tion of rival theories. Some combinations of scope and normalization point are open
to fairly decisive objections, and can be easily set aside: For instance, any view with
the extremely narrow scope of Lockhart’s PEMT will have the consequence that nor-
malizations between theories change from choice situation to choice situation and
hence that an agent may be rationally required to choose a course of action that
she knows in advance is strictly worse than some available alternative according to
221
every moral theory in which she has positive credence (Sepielli, 2013). On the other
hand, the extremely wide scope of Sepielli’s Conceivability PEMT coupled with its
use of range normalization yields disastrous consequences since for many theories
(e.g. nearly any form of aggregative consequentialism), the range of their value as-
signments over all conceivable options is either undefined (if the theory regards all
and only finite degrees of value as possible) or infinite (if the theory regards infinite
degrees of value as possible). Nevertheless, there is no obvious reason to think that
such decisive objections will narrow the set of plausible options down to one.40
This surfeit of possibilities is, paradoxically, an important reason for skepticism
concerning top-down normalization approaches in general. The principal motivating
idea behind any of these approaches is what MacAskill calls the “principle of equal
say” (PES), the idea that an agent who assigns equal credence to two moral theories
should give them equal weight or influence over her choices, or more generally that
the weight or influence of a moral theory on an agent’s choices should be in some
sense proportionate to her credence in that theory (MacAskill, 2014, pp. 101ff).
But as the preceding discussion illustrates, this principle is subject to a great many
40MacAskill argues for variance voting by a process of elimination, showing that range normal-
ization, Max-Mean, and Mean-Max have highly counterintuitive consequences in certain cases.
But variance normalization is not immune from such counterexamples: For instance, suppose an
agent A has equal credence in two theories T1 and T2 and faces a set of ten options, O1−10, of
which both theories agree that only O1 and O2 are reasonable options. T1 (on its pre-normalized
value scale) assigns O1 a value of 100, O2 a value of 95, O3−9 values of 0, and O10 a value of
−10, 000. T2 (on its prenormalized value scale) assigns O1 a value of 95, O2 a value of 100, and
O3−10 values of 0. The two theories, then, are quite similar except that they disagree about which
of O1 and O2 is more choiceworthy, and T1 has some particular, extremely strong moral objection
to O10 that leads it to assign an extreme negative value to this option. Because T1’s extreme
assignment to O10 makes its variance, on the prenormalized value scales, much greater than that
of T2, MacAskill’s variance normalization approach tells us that A should give greater weight to
T2’s preference for O2 over O1 than to T1’s preference for O1 over O2, and is therefore rationally
required to choose O2. But given that T1 and T2 both regard O3−10 as obviously bad options, it is
at least strange that the availability of O10, and the disagreement between T1 and T2 about how
bad this particular bad option is, should be the decisive factor in A’s choice between O1 and O2.
222
interpretations of which none is uniquely salient or obviously preferable to the rest.
Just as Bertrand-style paradoxes weaken the intuitive appeal of the principle of
indifference in probability theory (Van Fraassen, 1989, pp. 303ff), so the multiplicity
of “equalization” approaches to intertheoretic normalization weakens the intuitive
appeal of the principle of equal say.
It is also unclear to what extent PES is well-motivated in the first place. As
both Sepielli and MacAskill have pointed out, it is not as if I would treat a theory
“unfairly” by giving it less weight than another theory in which I have equal credence
(Sepielli, 2006, p. 602; MacAskill, 2014, p. 101). Since it seems at least intuitively
clear that some first-order moral theories do in fact propose weightier moral reasons
than others (say, divine command theory vs. Harman-style relativism), why should
there be even a presumption of equal weight or equal say among theories? And in
the absence of such a presumption, what principle could we appeal to in place of
PES to motivate any particular top-down normalization principle?41
Finally, there are the various technical problems that afflict all extant interpre-
tations of PES, including Lockhart’s PEMT and MacAskill’s variance normalization
approach: sensitivity to irrelevant alternatives, an apparent inability to accommo-
date theories whose range/variance is infinite or undefined, and the difficulty of
generalizing beyond simple cardinal structures (e.g. to ordinal or hyperreal struc-
41One option I haven’t discussed, but that deserves mention here, is a subjectivist approach
to intertheoretic normalization. On such an approach, at least when there are no content-based
grounds for any particular normalization between a pair of theories, the agent should simply
normalize the two theories in whatever consistent manner she chooses, or in whatever consistent
manner seems most appropriate to her. On this view, what grounds the appropriateness of a
particular normalization is a preference, a “seeming,” or some other such mental state of the agent.
This view may appeal to some, but of course will not satisfy those who think there is a more robust
fact of the matter about how, say, utilitarianism and Kantianism ought to be normalized.
223
tures).42
For these reasons, I regard top-down normalization as less attractive than
content-based normalization. Nevertheless, these worries are far from decisive, and
given the difficulty of foreseeing how far content-based normalization can be taken
along with the intuitive appeal of moral hedging, it is worth continuing to investi-
gate top-down normalization approaches in the hope of finding one that is uniquely
well-motivated and yields intuitively acceptable results. My own tentative view,
then, is that there is reasonable hope of finding universal or near-universal content-
based grounds for intertheoretic comparisons that reduce incomparability (or merely
binary normalization) to an acceptable minimum, but that if this hope cannot be
realized, then some top-down normalization principle will likely turn out to be the
next most attractive option.43
42Lockhart’s PEMT approach, as I have already said, simply assumes that all theories have or
can be appropriately represented as having cardinal structure, and so seems incapable of genuinely
accommodating ordinal structure. MacAskill attempts to make his Borda counting approach to
ordinal aggregation continuous with his variance normalization approach to the aggregation of (not-
otherwise-comparable) cardinal theories by handling ties in ordinal rankings in a way that amounts
to normalizing ordinal theories “at the variance of their Borda scores” (MacAskill, 2014, p. 247),
which I have argued amounts to tacit cardinalization as well. Any attempt at unit normalization
of ordinal theories will, I think, be subject to the same objection. However, level normalization of
ordinal theories—i.e., a normalization stipulating that a certain ordinal rank on the value scale of
T1 represents the same degree of absolute choiceworthiness as a certain ordinal rank on the value
scale of T2—does not involve any such implicit cardinalization, and is necessary if we wish to use
statewise or stochastic dominance as principles of decision-making under uncertainty over merely
ordinal theories. (An alternative is to rely on voting methods, e.g. Condorcet methods, that do
not in any obvious way require either unit or level normalization of ordinal theories.)
43One approach to top-down normalization that I regard as particularly promising, though it has
yet to be substantially developed or explored in the literature, eschews normalization points like
range or variance in favor of a broadly game-theoretic approach on which moral theories are repre-
sented as agents with utilities corresponding to their value assignments and “bargaining power” of
one kind or another corresponding to their subjective probabilities. One version of this approach
is Bostrom’s “parliamentary model,” described in a brief online article (Bostrom, 2009), on which
rational deliberation under moral uncertainty is modeled as a legislative body, with theories repre-
sented by a number of delegates proportionate to the agent’s credence, and in which decisions are
made by a deterministic voting method. (To give greater room for moral hedging in this model,
Bostrom adds the somewhat ungraceful epicycle that the delegates in the moral parliament believe
the voting method to be stochastic, so that even when a majority coalition favors a particular op-
224
6.4 Decision-Theoretic Compatibility
Combining the value assignments of different moral theories requires not only
a means of reconciling their potentially diverse structures and normalizing their
value scales but also an aggregation procedure or decision procedure: a procedure,
like dominance reasoning or expected value maximization, that tells us how to deal
with uncertainty over several different representations of the value of options, given
that we know how to compare those representations.44
To show that disagreement over aggregation procedures (just like differences
tion, they have incentives to bargain for additional support, e.g. by offering their support in other
choice situations to delegates who regard those choices as more urgent.) Other possibilities in a
similar spirit include the use of axiomatic bargaining solutions, taking moral theories as bargaining
agents with some appropriate, probability-sensitive outcome (say, MFO) as the threat point, or
modeling a set of choice situations as a multi-item auction in which theories are endowed with capi-
tal proportionate to their subjective probabilities with which they can “bid” for control of particular
choice situations. Such approaches have the advantage of at least largely avoiding the “irrelevant
alternatives” problems to which all the normalization-point methods described above appear to be
subject (see note 40), though this must be examined on a case by case basis (e.g., not all axiomatic
bargaining solutions that have been defended in the literature satisfy the standard Independence
of Irrelevant Alternatives axiom (Kalai and Smorodinsky, 1975)). Such game-theoretic approaches
may also allow for a more natural integration of theories with unbounded value scales, theories
that recognize infinite values, theories with merely ordinal structure, etc (insofar as agents with
analogously structured utility functions can still engage in rationally guided, strategic interactions
with other agents with variously-structure utility functions). For these reasons, these approaches
strike me as more promising than Lockhart’s PEMT or MacAskill’s variance normalization.
On any parliamentary, bargaining-theoretic, auction-theoretic, or similar approach, the question
of scope remains important, though all these approaches seem to require scope wider than a single
choice situation if they are to yield interesting conclusions. My own hunch is that the most
attractive top-down normalization approaches will take as their scope all the lifetime choices of
either a single agent or all actual agents, and then model the deliberative process either as a
cooperative bargaining situation with the preferred outcome given by some axiomatic bargaining
solution or as a multi-item auction. Since I regard top-down normalization methods as a second-
best option, however, I will not attempt to develop or evaluate such an approach here.
44It is potentially misleading to speak of “decision procedures” in the context of a multi-stage
approach, because except at the final stage of the aggregation process, these procedures will not
issue a decision but rather simply an aggregation of the reasons offered by some comparability
class of moral theories. So it is perhaps more proper to call these rules “aggregation procedures.”
Nevertheless, the procedures themselves are the sorts of rules familiar from decision theory for
weighing objective reasons that would exist given various states of the world, under conditions of
uncertainty, to arrive at a decision: rules like statewise dominance, stochastic dominance, risk-
neutral or risk-weighted expected value maximization, maximin, and so on.
225
Credence O1 (risky) O2 (risk-free)
S1 .9 0 100
S2 .1 2500 100
T1 (risk neutral) .3 250 100
T2 (risk averse) .7 5 10
Table 6.3: George’s Gamble, Pre-Normalization
in structure and the difficulties of normalization) presents a difficulty for interthe-
oretic value aggregation and hence for a complete theory of decision-making under
moral uncertainty, let’s begin with an example: Suppose that George is faced with
two options, one risky and the other risk-free, and divides his belief between two
moral theories, one of which is risk-neutral and the other of which is risk-averse—
specifically, measuring the expected value of an option by the probability-weighted
sum of squares roots of its possible objective values.45
Table 6.3 represents George’s situation, where O1 and O2 are his options, T1
and T2 are moral theories, and S1 and S2 are possible states of the world. We see
that T1 prefers the risky option, since it has higher expected value, while T2 prefers
the risk-free option, which has higher risk-weighted expected value according to T2’s
risk aversion function.
There is, to begin with, a normalization problem here: If we take T2’s square
root risk aversion at face value, then T1 will end up carrying much greater weight
in the aggregated verdict of the two theories than T1 (as can be seen in Table 6.3,
where the difference in risk-weighted expected value between O1 and O2 is thirty
45This case is inspired in part by a similar case in Nissan-Rozen (2015).
226
Credence O1 (risky) O2 (risk-free)
S1 .9 0 100
S2 .1 2500 100
T1 (risk neutral) .3 250 100
T2 (risk averse) .7 50 100
Table 6.4: George’s Gamble, Normalized
times greater according to T1 than T2). Let’s overcome this problem by stipulating
the following: T1 and T2 are both egoistic ethical theories, and the values they assign
to states of affairs represent the agent’s lifetime utilities. T2 is risk averse because it
embraces a notion of adequacy according to which there is great value in approaching
or achieving an adequate standard of wellbeing, but additional units of wellbeing
become increasingly superfluous once one has already achieved an adequate level of
lifetime wellbeing. The balancing point at which one’s level of wellbeing is neither
inadequate nor superfluous is 100—that is, for a life with 100 units of wellbeing, T2’s
concern for adequacy is inert. Thus, T1 and T2 agree on the value of an option, like
O2, that guarantees 100 lifetime units of wellbeing. Table 6.4 re-represents George’s
situation according to this normalization.
So far so good, then: T1 and T2 are both ratio-scale theories, so we don’t need
to worry about structural diversity, and we have found a way of normalizing their
value scales. But now a new problem presents itself: If we simply calculate risk-
neutral expectations over T1 and T2, we find that EV(O1) = (.3× 250) + (.7× 50) =
75 + 35 = 110 while EV(O2) = (.3 × 100) + (.7 × 100) = 30 + 70 = 100, and we
conclude that George rationally ought to choose O1. But, if we instead calculate
227
risk-weighed expectation over T1 and T2, using T2’s preferred square root function,
we find something different: EVr(O1) = (.3×
√
250)+(.7×√50) ≈ 4.74+4.95 = 9.69
while EVr(O2) = (.3 ×
√
100) + (.7 × √100) = 3 + 7 = 10, and we conclude that
George rationally ought to choose O2.
In the Easy Case, I argued that risk-neutral expectational aggregation was
appropriate on the grounds that both theories in which Alice had positive credence
endorse a risk-neutral expectational principle of decision-making under uncertainty.
(See §B.1 for details.) But what should we say when the first-order theories we
are attempting to aggregate seem to disagree about the appropriate aggregation
principle?
One way to think of the difficulty is to recall the idea that second-order prin-
ciples of decision-making under moral uncertainty might “beg the question” against
the first-order theories they are meant to help us handle our uncertainty over: If
George is uncertain whether he ought to be risk-neutral or risk-averse, then in aggre-
gating the risk-neutral and risk-averse theories by means of a risk-neutral decision
procedure, he would seem to have forgotten that this is the very thing he’s uncertain
about.
Another feature of the difficulty is that either choice of aggregation procedure,
in a case like George’s, seems to give one of his first-order theories undue weight.
The risk-neutral theory insists that small chances of very large positive payoffs
should be taken as seriously as their expectations suggest, while the risk-averse the-
ory insists that one should generally not sacrifice the surety of a good thing for a
small chance of an extremely good thing, even when the expected value of doing so
228
is positive. Adopting a risk-neutral second-order aggregation procedure will allow
this sort of long-shot reasoning to dominate decision-making even when the agent’s
credence in the risk-neutral first-order theory is relatively low, while adopting a
risk-averse second-order aggregation procedure may result in almost complete in-
sensitivity to high-value, low-probability prospects, even when the agent’s credence
in the risk-neutral first-order theory is relatively high. The more the decision pro-
cedures diverge (e.g. if instead of the square-root risk averse theory we consider a
cube-root theory, or a theory that simply rounds off “rationally negligible proba-
bilities” to zero), the more extreme the effects of committing to the corresponding
second-order aggregation principle will become.
There are a great many things that might be said about this problem (which
we might call the “problem of decision-theoretic uncertainty”), but as in previous
sections, I will focus on a few general lines of response.
1. Least Common Denominator The aggregate value assignment of compa-
rability class {T1, T2, ...Tn} is defined only insofar as the decision procedures
endorsed by each of T1−n are in agreement. Where there is disagreement, the
result will be an imprecise assignment of cardinal values and/or an incomplete
ordinal ranking of options.
2. Open-Ended Internalism What procedure should be used to aggregate a
comparability class {T1, T2, ...Tn} depends entirely on the agent’s doxastic
state, but may depend on features of her doxastic state other than the de-
cision procedures endorsed by T1−n. For instance, an agent may be uncertain
229
whether risk-neutrality or risk-aversion is correct at the first-order level but
be justifiably certain, e.g., that risk-neutrality is correct at the second-order
level, or that she should adopt a second-order risk attitude that averages (in
some particular way) the risk functions of her various first-order theories, or
that the only valid second-order norm is MFO or statewise dominance or... In
any of these cases, she rationally ought to employ the second-order aggrega-
tion procedure that she judges to be correct, whether or not it is intuitively
“compatible” with the first-order theories being aggregated.
3. Externalism Some decision-theoretic principle(s) have belief-independent ra-
tional requiring force, such that any agent is rationally required to aggregate
over her uncertainties by means of these principles regardless of her doxas-
tic state, and in particular regardless of what decision-theoretic principles are
endorsed by the first-order theories in which she has positive credence.
An evaluation of these options must be put off to the next chapter, where
I will argue for my own preferred resolution to the problem of decision-theoretic
uncertainty. But to sketch the argument in advance: As I have already suggested,
I agree with Weatherson that the regress problem forces us to treat at least some
normative principle or principles as having external normative force. Hence, I will
argue in the next chapter, when an agent is uncertain about all higher-order as well
as first-order normative principles, open-ended internalism gets stuck in an infinite
regress and the least common denominator approach on its own yields the vacuous
result that everything is rationally permissible. So we must adopt at least a weak
230
form of externalism. However, if the principles to which we attribute external force
are sufficiently weak (e.g. statewise or stochastic dominance rather than maximizing
expected choiceworthiness), then the least common denominator approach may still
have a role to play, providing us with stronger aggregation procedures for certain
comparability classes of theories.
6.5 Rational Choice under Normative Uncertainty: A Tentative Gen-
eral Account
In this chapter I have identified three aspects of the problem of intertheoretic
value aggregation, each of which presents a distinct challenge for a complete the-
ory of rational decision-making under moral uncertainty. I have described what I
take to be the most plausible responses to each of these challenges, and argued for
my own preferred responses (though as yet in only a promisory way with respect
to the problem of decision-theoretic uncertainty). Taken in combination, these re-
sponses constitute the outline of a general theory of decision-making under moral
uncertainty. This theory can be summarized as follows:
General View multi-stage aggregation ∧ (universal content-based normalization
∨ top-down normalization) ∧ (externalism + least common denominator)
To put this into slightly better English: A morally uncertain agent should de-
cide what to do by aggregating the value assignments of the various moral theories
in which she has positive credence, beginning with minimal comparability classes
(united by shared structure, precise content-based normalization, and if necessary
231
by a shared decision theory), whose aggregate value assignments are taken as inputs
to broader comparability classes, and so on. The normalization and aggregation
procedures within comparability classes should be grounded in the agent’s own dox-
astic state as much as possible, but if content-based normalization turns out to be
insufficient then top-down normalization principles may have a role to play as well,
and some (perhaps quite weak) aggregation principle must have normative force
that is independent of the agent’s doxastic state.
At several points, I have defended various components of this view by compar-
ing costs, i.e., conceding that my own preferred approach fails or might fail to satisfy
some genuine desideratum while arguing that alternative views fail to satisfy more
important desiderata. It would be more persuasive, certainly, if I could convincingly
claim that my view had no substantial downsides. But as far as I can tell, there
is simply no such view on offer in this domain. Irrelevantism, MFT, MFO, and
every extant hedging theory all carry substantial costs, and if these costs cannot be
avoided, then they must simply be compared.
The view I have put forward remains highly schematic, of course, and therefore
its costs can be only imperfectly assessed. Among the details left to be filled in are (i)
the precise order in which the multi-stage aggregation procedure unfolds (i.e., which
comparability classes should be taken as input to which larger comparability classes);
(ii) what kinds of content-based normalization are possible, especially between very
disparate theories; (iii) whether there is a correct way to normalize theories when
content-based normalization fails, and if so what top-down normalization principle
gives the right account of these cases (e.g., by capturing the true purport of PES);
232
and (iv) what aggregation principle or principles have external normative force. (I
will have something to say about this last question in the next chapter.)
Especially in light of these as-yet missing details, it is natural to ask: Isn’t this
all just too complicated to be the right account of things? Doesn’t the comparative
simplicity of MFT, MFO, or even irrelevantism look like a powerful reason to prefer
one of these views over moral hedging, if this is the simplest and most cohesive
view the advocate of hedging can provide? This concern is worth taking seriously.
Nonetheless, the complexity of the view I have outlined proceeds from a simple idea:
that a rational agent should try to do the right thing, relative to her own imper-
fect state of belief about her objective reasons. The complexity of the deliberative
process that this entails for an ideal agent (and that non-ideal agents can only try
to approximate) results from the complexity of the agent’s doxastic state, i.e., from
her distribution of belief over moral theories with complex and diverse structure,
metaethical/explanatory content, and decision-theoretic commitments. Any theory
of decision-making under moral uncertainty that takes seriously the idea of respond-
ing rationally to one’s actual belief state will have to respect this complexity, not
attempt to eliminate it (e.g., by tacitly replacing all the theories in which the agent
has positive credence with simple cardinal surrogates). Nonetheless, the complexity
of the resulting decision-procedure has a unifying motivation.
One aim of the next chapter will be to give a more precise account of this
underlying unity: in particular, to suggest that a single norm, the enkratic principle
correctly understood, can explain both the internalist and externalist elements of
rational decision-making under moral uncertainty.
233
Chapter 7: The Regress Problem
The central question so far as been the following: What are the correct second-
order rational norms for an agent who is uncertain about first-order moral norms—
i.e., what is the correct deliberative procedure for such an agent to follow? We have
not yet considered the natural follow-up question: How should an agent deliberate
when she is uncertain not only about first-order moral norms but also about second-
order rational norms? If, for instance, in additional to having positive credence in
various first-order moral norms that offer conflicting guidance in the choice situation
before her, she also has some positive credence in various competing second-order
principles like MFT, MFO, and a variety of hedging principles involving different
approaches to the problem of structural diversity, the normalization problem, and/or
the problem of decision-theoretic uncertainty, then should she simply be guided
by whichever of these second-order principles is most plausible (the higher-order
equivalent of MFT), or choose the option supported by the greatest probability-
weighed sum of second-order principles (the higher-order equivalent of MFO), or
find some way to weigh the advice of these competing principles against each other
(the higher-order equivalent of hedging)? And what if she is uncertain in turn about
which of these third -order approaches to her second-order uncertainty is correct?
234
These questions introduce a fundamentally new difficulty, and suggest a final line of
objection with which any complete defense of moral hedging must contend.
In the next section, I will consider the regress problem as an argument for
normative externalism. I will then examine two possible lines of reply that avoid
conceding the existence of any external norms, and argue that these replies are
unsatisfactory. Instead, I’ll then argue that we should admit the existence of one
external norm: the enkratic principle, correctly formulated. I will consider some
variants of the enkratic principle that have different implications for decision-making
under normative uncertainty, without attempting to settle the potential debates
between them.
This “enkratic externalist” position satisfies the demands of the regress argu-
ment for externalism. But it is not obvious that this is enough to solve the regress
problem. In §7.2, I consider two aspects of the regress of higher-order norms that
might be thought to pose a problem, even given the concession of a limited form of
externalism. One concerns how an ideally rational agent should act when she is un-
certain about norms at every level of an infinite hierarchy. The other concerns how
a non-ideal agent should approximate the requirements of ideal rationality, when
she is only capable of contemplating finitely many levels of such a hierarchy.
I will argue, however, that enkratic externalism does in fact solve both of these
problems: If rationality requires full commitment to the enkratic principle, then
higher-order uncertainties take the form, not of uncertainty about competing final
principles, but of uncertainty about competing heuristics for predicting the demands
of the enkratic principle. For this reason, I argue in §7.3 that the problem of ideal
235
rationality simply doesn’t arise, given enkratic externalism: Any ideally rational
agent, capable of contemplating an infinite hierarchy of norms, would not need to
rely on heuristics to determine the probabilities of objective value propositions, but
could reason directly from her evidence. There are also at least two ways in which a
regress of ideal, non-heuristic principles might still be thought to arise, despite the
commitment to enkratic externalism, but I argue that both can be avoided.
But this does not resolve the problem of non-ideal rationality: Given that I
must rely on heuristics, for instance, to approximate the expected choiceworthiness
of my options, and that I must inevitably be at least somewhat uncertain about
what heuristic norms to follow, I and other non-ideal agents still seem condemned
to infinite regress. Taking up this problem in §7.4, I describe a model of non-
ideal practical deliberation on which agents like ourselves are not rationally required
to ascend an infinite hierarchy of heuristic norms, because at some point in the
process of deliberation—before we have ascended more than a very short way up
the hierarchy—we will justifiably conclude that the marginal cost of any further
deliberation exceeds the marginal benefit.
7.1 The Regress Argument for Externalism
We first encountered the regress problem as an argument for normative exter-
nalism in Chapter 3. But we should now take the time to set out this argument in
more detail. Its most compelling formulation, to my mind, is as follows.
236
The Regress Argument
1. If the rational requiring force of any normative principle N for agent A (i.e.,
whether A is rationally required to do as N advises) depends on A’s beliefs
about N , then in any choice situation where A is uncertain of the correctness
of N1, which advises her to choose option O, and assigns some positive degree
of belief to rival norms N2, N3, ...Nm, some of which advise against choosing
option O, A is not rationally permitted to choose option O straightaway on
the basis of N1, but is rationally required to deliberate in a way that accounts
for the possible correctness of N2−m.
2. In order to deliberate in a way that accounts for her uncertainty over N1−m,
A must make deliberative use of at least one higher-order normative principle
N ′1.
3. A rational agent will always be at least somewhat uncertain of any normative
principle or disjunction of normative principles, apart (perhaps) from unhelpful
trivialities like “It’s not the case that O is both rationally permissible and not
rationally permissible.”
4. If (3), then for any substantive normative principle N, whether first-order or
higher-order, and in any choice situation, any rational agent will have non-zero
credence in some normative principle of the same order that conflicts with N
(i.e., that regards a disjoint set of options as permissible).
5. Therefore, if the rational requiring force of any normative principle N for
agent A depends on A’s beliefs about N , then any rational agent in any choice
237
situation will be rationally required to make deliberative use of an infinite
series of higher-order normative principles. [from (1)-(4)]
6. Agents are not rationally required to make deliberative use of an infinite series
of higher-order normative principles.
C. For at least some normative principle N, the rational requiring force of N for
an agent A does not depend on A’s beliefs about N. [from (5), (6)]
Weatherson, I think, has something like this argument for externalism in mind
(Weatherson, 2014, pp. 155-7). I will argue shortly that the argument is sound,
but that contra Weatherson, it provides most support not for externalism about
first-order moral principles but rather for externalism about second-order rational
principles (specifically, the enkratic principle). But first we should consider potential
objections to the argument for externalism. Many such objections are possible.
Premise (1) may seem implausibly demanding, and Sepielli’s solution to the regress
problem, which we will consider in §7.1.2, can be interpreted as an attempt to reject
it. Premise (2) is plausible but non-obvious. Premise (3) depends on a commitment
to Bayesian regularity, which might be rejected for various reasons. I will focus,
however, on two general lines of response, which might be understood as attacking
the above argument at various points, but are best understood as strategies for
rejecting either (1), (4), or (6).
238
7.1.1 Fixed Point Solutions: Convergence or Contraction
The simplest way of responding to the regress problem is to claim that the
regress will stop after some finite number of iterations, because a rational agent will
in one way or another achieve certainty either about the relevant higher-order norms
or about what she subjectively ought to do. More precisely, there are two ways in
which the regress of higher-order norms might turn out to be less than maximally
vicious.
1. Convergence For any rational agent A in any choice situation, there is some
finite n such that all nth- and higher-order norms in which A has positive
credence will agree on which of her options are rationally permissible.
2. Contraction For any rational agent A in any choice situation, for any level
n in the hierarchy of higher-order norms, the set of options that some n+ 1st
order norm in which A has positive credence regards as rationally permissible
will be a subset of the set of options that some nth order norm regards as
rationally permissible—that is, as A ascends the hierarchy of norms, the set
of options that might be permissible will monotonically contract. Assuming
that A has only finitely many options, then there must be some level m such
that for any level o ≥ m, the set of possibly permissible options according to
oth-order norms in which A has positive credence will be identical to the set
of possibly permissible options according to mth-order norms in which A has
positive credence—that is, the set of possibly permissible options will stabilize,
239
remaining the same for all orders of norm greater than or equal to m. Then,
one might think, there will be some simple procedure for choosing among this
stable set of possibly-permissible options.
Convergence would offer the simplest and most desirable solution to the regress
problem, and would escape the argument for externalism by denying premise (4).
But I see no reason to think that it is correct. Surely in any choice situation, for any
level n in the hierarchy of norms, there is at least one pair of logically possible nth-
order norms that offer conflicting guidance (i.e., disagree about which options are
rationally permissible). And even if we grant, contra regularity, that it is sometimes
rationally permissible to have zero credence in a logical possibility, what reason is
there to think that we will always be able to eliminate as epistemic possibilities all
nth-order norms, for some n, except those that agree on a particular set of options
as permissible?
The more modest hypothesis of Contraction is suggested by Sepielli (2014b).1
It is not obvious to what extent Contraction, if true, would help us solve the regress
problem. One trivial way in which it could be true, for instance, is if we accept
the apparent implication of regularity, that for any option in any choice situation, a
1He writes: “I hope to eventually show that the further we ‘step back’ from states of uncertainty
[...] the fewer prospective actions will be left in the disjunctive set amenable to intentional expla-
nation[...] Suppose, for example, that I can do any of mutually exclusive actions A...Z. Perhaps
my uncertainty regarding objective normativity will intentionally explain my doing any of A...R
rather than S...Z [...] Now suppose that I consider what to do relative to the [minimal probabilities]
expressed in that uncertainty, and am uncertain as to the answer. Perhaps this new uncertainty
will intentionally explain my doing any of A...G, rather than H...Z [...] My hope is to show that, as
a general matter, potential actions will be hived off with each stepping-back, and that previously
hived off actions will never reappear among the set amenable to intentional explanation. I have
reasons to suspect that such a showing is possible, but I’ll have to make good on that suspicion
elsewhere.” (Sepielli, 2014b, p. 539)
240
rational agent should have some positive credence, at every level of the hierarchy of
norms, in one or more norms that regard that option as rationally permissible. In
this case, the set of possibly permissible options would never expand as we ascend
the hierarchy (satisfying Contraction), but it would never contract either. And this
presumably does not constitute a satisfactory solution to the regress problem.
Contraction might be helpful if it were guaranteed that, as we ascend the
hierarchy of norms, the set of possibly permissible options would contract in some
significant way (without necessarily reaching the point at which all theories agree
on which options are permissible), and never expand. Then it might be plausible to
hold that, whenever an agent realizes that, say, for every level of the hierarchy n ≥ 5,
the nth-order norms in which she has positive credence will designate exactly the
options O1, O2, ...Om as possibly permissible (i.e., at least one nth-order norm will
designate each of O1−m permissible, and no nth-order norm will designate any other
option as permissible), she can then invoke some simple “ω-order” norm (e.g., that
permits her to choose any of O1−m).2 This would amount to denying premise (6)
in the argument for externalism, and holding that although rationality does require
agents to consider the full infinite hierarchy of norms, the stabilization of the set of
possibly permissible options above some finite level renders this requirement feasible
for finite agents.
There is an obvious reason why this solution should seem unpromising: namely,
if we have followed the regress argument this far, why should we not suppose that an
agent can be rationally uncertain about ω-order norms just as much as she can be
2MacAskill appears to endorse this view (MacAskill, 2014, pp. 218-219).
241
about norms at finite levels of the hierarchy? For instance, along with the norm that
permits any of O1−m, she might consider a norm that instructs her to choose the
option most likely to be permissible, perhaps according to her fifth-order norms (the
first level of the hierarchy at which the set of possibly permissible options reaches
its stable minimum), or perhaps according to the norms at the highest finite level
of the hierarchy she is capable of considering, or perhaps somehow averaging the
probability of an option being permissible over the levels of norms ≥ 5 (e.g, for each
option O, taking the Cesa`ro sum of the probabilities that it is rationally permissible
given by the 5th-order norms, 6th-order norms, 7th-order norms, ...). Thus, so long
as the norms continue to disagree about which options are permissible at every finite
level (or more precisely, so long as for every level, there is some higher level at which
some norms disagree about which options are permissible), Contraction alone seems
unpromising, even if Sepielli’s monotonicity conjecture turns out to be correct.
7.1.2 Sepielli’s View
In other work, however, Sepielli has suggested a different approach to the
regress problem. This approach begins by distinguishing between dispositional un-
certainty and conscious uncertainty. A rational agent, Sepielli claims, may be dis-
positionally uncertain without being consciously uncertain, and may rationally act
on a norm N without considering alternative norms, even if she is dispositionally
uncertain about the correctness of N, so long as she is not consciously uncertain
(Sepielli, 2014a). This view avoids the argument for externalism by denying either
242
premise (1) or premise (3)—premise (1) if we interpret the references to “uncer-
tainty” in these premises as referring to dispositional uncertainty, premise (3) if we
interpret them as referring to conscious uncertainty. Since I will claim that disposi-
tional uncertainty is the more relevant condition, let’s adopt the first interpretation
of the argument, and understand Sepielli as challenging premise (1).
One immediate worry is that Sepielli’s view seems too permissive: We don’t
want to concede that an agent is rationally permitted to act on any norm, however
much her evidence calls that norm into question, so long as she doesn’t feel uncertain
about it. Sepielli offers a response to this worry, arguing that a rational agent will
always experience some pro tanto impetus to attend consciously to her dispositional
uncertainties, and that this will sometimes but not always lead her to experience
conscious uncertainty about a normative principle N about which she was previously
only dispositionally uncertainty, and therefore to seek some higher-order norm to
guide her action in the face of this newfound conscious uncertainty. She may become
consciously uncertain of this higher-order norm as well, and be forced to ascend the
hierarchy another level, but she also may not, and in any event will at some point
reach a level where her impetus to convert dispositional uncertainty into conscious
uncertainty runs out, ending the regress.
Nevertheless, Sepielli’s view is still implausibly lax. Consider where the line
between dispositional and conscious uncertainty must lie, for this view to succeed
in escaping the regress problem. We might draw a distinction between the merely-
dispositional uncertainty of an agent who doesn’t realize she has any reason to
doubt proposition P, but were she to consider it, would conclude that she does,
243
and the conscious uncertainty of an agent who realizes that P is less than certain.
But understood in this way, the distinction would not resolve the regress problem,
for a philosophically astute agent will realize that any (or nearly any) non-trivial
normative proposition is less than certain, and hence will in this sense be consciously
uncertain about any norm she considers guiding her action by.
Sepielli’s distinction between conscious and dispositional uncertainty is differ-
ent. Here is his explanation.
I think we need to distinguish between two types of uncertainty. The
first is dispositional, not necessarily conscious, the sort of attitude I have
towards any claim I wouldn’t bet my life on. The second is conscious,
involving a feeling of directionlessness, the kind that appears when I de-
liberate, and disappears when I’m “in the zone” [emphasis added]. I am
uncertain in only the first sense about what the strings on a guitar are;
I am uncertain in both senses about what the strings on a banjo are.
That is why I can simply play an A7 on a guitar, but can play an A7 on
a banjo only by trying. (Sepielli, 2014a, p. 91)
As Sepielli himself admits, however, it is unclear why the absence of this sort
of conscious uncertainty (“a feeling of directionlessness”) should make it permissible
for me to act straightaway on a norm N, without considering the possibility that N
might be mistaken. He concedes that “the waning of conscious uncertainty is only
a solution to the psychological problem of how we can act without [taking unguided
leaps of faith]. It’s not a solution to the normative problem of how we can manage
244
moral risks non-recklessly” (pp. 91-2).
Sepielli’s final conclusion with respect to the regress problem, then, is moder-
ately pessimistic.
I think the right thing to say is that meta-rules offer us a normative
advantage by forestalling moral recklessness, rather than by eliminating
it entirely. More precisely, there is a sense in which it is better to leap
[i.e., “take a leap of faith” by acting on a norm N the truth of which
is uncertain] in the face of uncertainty about meta-rules than to leap in
the face of uncertainty about ordinary moral rules, better still to leap
in the face of uncertainty about meta-meta-rules, and so on. (Sepielli,
2014a, p. 92)
On one interpretation, this view amounts to a concession that agents can
never fully satisfy the requirements of rationality. On another, Sepielli is still deny-
ing premise (1), and holding that at least in some circumstances when an agent is
dispositionally uncertain but not consciously uncertain about a norm N, she is ra-
tionally permitted to act on N straightaway without considering the possibility that
N might be wrong. The latter conclusion would seem unwarranted, however, absent
a reason why the feeling of certainty relieves an agent of her rational responsibility
to consider the possibility of error. The former conclusion would be disappointing,
and is worth avoiding if we can. In the next section, therefore, I will suggest that
we should concede a weak form of externalism, one that preserves the requirement
of moral hedging while avoiding an infinite regress of higher-order norms.
245
7.1.3 Enkratic Externalism
My proposal is that there is exactly one normative principle that has ex-
ternal, belief-independent normative force, namely, the enkratic principle, rightly
formulated. Let’s call this view enkratic externalism.3 The view, so stated, is in-
tentionally underspecified, for as we saw in §2.2, many formulations of the enkratic
principle as possible. We will shortly consider, without attempting to definitely re-
solve, which of these formulations is right. But first let’s make an initial assessment
of the plausibility of enkratic externalism in its underspecified condition.
By conceding the conclusion of the regress argument given in §7.1, enkratic
externalism allows us to accept the premises of that argument without paradox.
Whether this is enough to resolve the regress problem is another matter, which
we will take up in §§7.2-7.4. But on face it seems promising, for if an agent is
rationally permitted—indeed, rationally required—to do as EP advises, even when
she is uncertain of EP, and if EP can play the role of a second-order principle for
decision-making under first-order moral uncertainty, then the regress simply stops
at the second level: An agent may act on a second-order principle (EP) without
considering alternatives, and therefore without needing to make deliberative use of
any third-order principle(s).
3The idea that the principles of rationality that govern choice under first-order moral uncertainty
have belief-independent force has been suggested by Krister Bykvist, though his conception of
rationality seems to be Humean rather than enkratic: “Being rational, in the sense I have in
mind, has to do with the coherence between your preferences about simple outcomes, what you
fundamentally care about, and your preferences about prospects constructed out of these simple
outcomes...[M]y tentative conclusion is that in cases of uncertainty of rational matters there is an
answer to the question of what it is rational to prefer which is not sensitive to your own views
about rationality.” (Bykvist, 2013, p. 133)
246
But even if it can solve the regress problem, enkratic externalism faces chal-
lenges from two directions: from internalists like Sepielli, and from first-order moral
externalists like Weatherson. From the internalist direction, the basic question is:
How can an agent be rationally required to act on a principle regardless of her beliefs
or evidence with respect to that principle? To begin with, following premise (1) of
the argument for externalism, we might wonder why it is even rationally permis-
sible to act (straightaway) on a principle of which one is uncertain. But enkratic
externalism endorses a stronger claim than this: namely, that an agent is rationally
required to act on EP, even if she justifiably regards it as more likely (perhaps much
more likely) false than true. This seems to gratuitously offend the idea that favored
moral hedging in the first place, that agents should be guided by their normative
beliefs. Indeed, by this token, it seems paradoxically to offend the essential spirit of
the enkratic principle itself.
It seems to me, however, that there is no paradox in viewing EP as an excep-
tion from its own internalist motivations. Enkratic externalism claims that EP is
a constitutive or definitional requirement of rationality, i.e., that to be practically
rational just is to be enkratic. To see the plausibility of this claim, consider the lim-
iting case of EP in which an agent A is certain that she has most objective reason
to choose option O. If, to put it roughly, being rational just means that you respond
in a coherent way to your reasons as you understand them, it seems that rationality
must require A to choose O—for how could choosing anything other than O con-
stitute a coherent response to A’s reasons in this circumstance, as she understands
them? Even if A, as a good Bayesian, has some positive credence in the principle
247
EP−1 according to which she is rationally prohibited from choosing any option that
she is certain she has most objective reason to choose, there is plausibly nothing
irrational in ignoring that credence. By contrast, first-order moral principles like
maximize aggregate pleasure minus pain or don’t intentionally deceive other agents
are not plausibly constitutive requirements of rationality (pace certain Kantians)—
rather, they are theories of an agent’s objective reasons, which an agent might fully
or partially reject for purposes of practical deliberation without compromising her
rationality.4
This explains why enkratic externalism should be acceptable to proponents of
moral hedging, and also suggests a reason to prefer enkratic externalism to Weath-
erson’s first-order moral externalism. Are there any countervailing reasons to prefer
4The regress problem for normative uncertainty has a rough analogue in the regress problem
faced by conciliationist views of peer disagreement (how to handle “peer disagreement about peer
disagreement”), which has received somewhat more attention in the recent literature. Weatherson
(2007, 2013) has also taken the pessimistic view in this debate, arguing that conciliationist views
are incoherent since they recommend moderately one’s credence in conciliationism in the face of
disagreement while also adopting degrees of belief with respect to other propositions that can
only be justified by full credence in conciliationism. In response, Elga (2010) defends a version
of conciliationism that is self-exempting, arguing that any basic principle of inductive reasoning
“must be dogmatic with respect to its own correctness” (Elga, 2010, p. 185). That is, Elga
claims that an epistemically rational agent should be steadfast in the face of peer disagreement
about conciliationism and other basic inductive norms, believing with probability 1 that those
norms are correct, while adjusting her credences on other questions in response to disagreement as
conciliationism demands. This is not dissimilar from the enkratic externalist claim that a rational
agent should be fully committed to EP in deliberation, while hedging for her uncertainty about
other normative principles on the basis of EP. The important difference is that I am claiming, not
that a rational agent should be dogmatic about the truth of EP, no matter what her evidence,
but that she should comply with EP even when she is uncertain of its correctness. In the peer
disagreement context as well, this externalist line strikes me as the most plausible conciliationist
response to the threat of regress. That is, rather than claiming that an agent should always have
credence 1 in conciliationism, whatever her evidence to the contrary, the conciliationist should
instead claim that some basic norms of epistemic rationality, which include or imply conciliationism,
have belief-independent force, such that agents are rationally required to comply with these norms
even while rationally doubting their correctness. Like the enkratic externalist view I am defending,
this claim will be plausible to the extent that these epistemic norms can be plausibly characterized
as constitutive or definitional requirements of epistemic rationality, i.e., “just what it is” to be
epistemically rational.
248
first-order moral externalism, apart from those we considered in Chapter 3? Weath-
erson does not say much on this question. The relevant passage (worth quoting at
length in part for its particularly nice statement of the regress problem) reads as
follows.
There is a worry that externalism is not sufficiently action guiding,
and can’t be a norm that agents can live by. But any philosophical the-
ory whatsoever is going to have to say something about how to judge
agents who ascribe some credence to a rival theory. That’s true whether
the theory is the first-order theory that Jeremy Bentham offers, or the
second-order theory that Andrew Sepielli offers. Once you’re in the busi-
ness of theorising at all, you’re going to impose an external standard on
an agent, one that an agent may, in good faith and something like good
conscience, sincerely reject. The externalist says that it’s better to have
that standard be one concerned with what is genuinely valuable in the
world, rather than a technical standard about resolving moral uncertainty
[emphasis added]. But every theorist has to be a little bit externalist;
the objector who searches for a thoroughly subjective standard is going
to end up like Ponce de Leon. (Weatherson, 2014, pp. 156-7)
As the italicized sentence suggests, Weatherson’s major motivation for preferring
first-order moral externalism to any form of externalism that would allow for hedg-
ing is the fetishism argument, to which I have already replied. There is also a
suggestion, perhaps, that it would be arbitrary for some “technical standard” like
249
Lockhart’s PEMT to have external normative force. But enkratic externalism avoids
this worry insofar as it is plausible that enkrasia is uniquely constitutive of practical
rationality.5
7.1.4 Which Enkratic Principle?
As we’ve seen, many formulations of the enkratic principle are possible. The
standard formulation of EP says that if A believes that she ought to choose option O,
then she is rationally required to choose option O. Following Wedgwood, I pointed
out in Chapter 2 that this principle either fails to account for uncertainty (since
an agent might believe that she ought to choose O, and yet have positive credence
that she has very strong reasons to choose some other option instead, such that it
is on balance not rational for her to choose O), or else is almost never applicable
to actual agents (if belief is understood to require the absence of any uncertainty).
Wedgwood’s formulation of the enkratic principle as a requirement to maximize
expected choiceworthiness offers a more general principle of rationality that accounts
for uncertainty about objective oughts.6
One version of enkratic externalism, then, holds that agents are rationally
required, whatever their beliefs, to choose options with maximal expected choice-
worthiness. Let’s call this view EEMEC . We might reasonably wonder whether this
5One might worry that fully escaping the regress problem requires us to grant belief-independent
status to some normalization principle like PEMT, or other technical features of a theory of
decision-making under moral uncertainty that do not seem like constitutive requirements of ratio-
nality. I address this worry in §7.3 below.
6As we’ve also seen, Sepielli, MacAskill and others endorse maximizing expected
value/choiceworthiness as a general principle of rationality, though to my knowledge Wedgwood is
the first to have proposed it as a formulation of the enkratic principle.
250
view is too strong, though. Is it really constitutive of rationality that I maximize
expected value, with the very particular risk attitude (viz., neutrality) that this
implies? Many theorists have argued that that non-neutral attitudes toward risk
are rationally permissible (prominent examples include Keynes (1921) and Buchak
(2013)), and it seems implausible that an agent who justifiably believes this to be
the case and, rather than maximizing expected choiceworthiness simpliciter, maxi-
mizes some form of risk-weighted expected choiceworthiness (perhaps averaging over
risk-weighed and risk-neutral theories of rational choice under moral uncertainty)
thereby acts irrationally.
We might therefore prefer a weaker version of enkratic externalism. One salient
option is to replace Wedgwood’s expectational principle with a stochastic dominance
principle. A practical option O stochastically dominates an alternative P for agent
A iff, relative to A’s subjective probabilities
1. For any degree of choiceworthiness d, the probability that O is (or will turn out
to be) at least as choiceworthy as d is equal to or greater than the probability
that P is (“”) at least as choiceworthy as d, and
2. For some degree of choiceworthiness d, the probability that O is (“”) at least
as choiceworthy as d is strictly greater than the probability that P is (“”) at
least as choiceworthy as d.7
A stochastic dominance version of the enkratic principle would assert that, for any
7More precisely, these conditions define first-order stochastic dominance—I omit this qualifi-
cation in the main text for simplicity. More a more detailed explanation of stochastic dominance
reasoning, including its application to moral uncertainty over ordinally and lexically structured
theories, see Appendix A.
251
agent A and option O, if there is some alternative option P such that A’s probabil-
ity distribution over P ’s possible degrees of objective choiceworthiness stochastically
dominates A’s probability distribution over O ’s possible degrees of objective choice-
worthiness, then A is rationally required not to choose O. Let’s call the version of
enkratic externalism based on this formulation of the enkratic principle EESD.
Stochastic dominance is, in effect, expectation-maximizing minus any con-
straints on an agent’s risk attitudes. Unlike expectational reasoning, which is con-
troversial because of the constraint it imposes on risk attitudes, the principle that
agents ought to reject stochastically dominated options is largely uncontroversial
in decision theory. In formal terms, an agent will reject stochastically dominated
options so long as her utility function is monotonically increasing, i.e., so long as she
prefers a given chance of a higher-value outcome to the same chance of a lower-value
outcome, all else being equal (Hadar and Russell, 1969, p. 28). This is much more
plausibly a constitutive requirement of rationality than the risk-neutrality required
by MEC.
A further advantage of a stochastic dominance formulation of the enkratic
principle over Wedgwood’s MEC formulation is its ability to handle moral uncer-
tainty over ordinal and lexical theories. MEC is simply unable to handle ordinal
theories (since ordinal rankings can’t be multiplied by probabilities) and yields the
implausibly fanatical conclusion that any non-zero probability of a gain or loss on
the lexically prioritized value dimension according to a class of lexical theories takes
absolute precedence over any gain or loss on the lexically deprioritized dimension.
Stochastic dominance, while it is silent in a great many cases, can accommodate
252
and yield plausible conclusions with respect to both ordinal and lexical theories (see
Appendix A).
But there are drawbacks to EESD as well. For one thing, as just alluded to,
stochastic dominance offers a significantly incomplete ranking of uncertain prospects—
i.e., for most pairs of options, neither will stochastically dominate the other. Thus,
relying purely on stochastic dominance reasoning would leave us with a very weak
principle of decision-making under moral uncertainty, forcing us to concede that
in the typical case, many of an agent’s options will be rationally permissible. An
alternative is to adopt the hybrid approach suggested in the last chapter, treating
stochastic dominance as the only principle of rational requirement that has external
normative force while holding that an agent may be subject to internal (belief-based)
rational requirements to employ stronger principles, e.g., to aggregate expectational
theories expectationally. We will explore this issue at greater length in §7.3.
Additionally, while expectational reasoning requires only unit-normalizability
between moral theories, stochastic dominance reasoning requires level -normalizability,
i.e., it requires that we be able to say whether the value of option O according to
T1 is greater than, less than, or equal to the value of option P according to T2.
8
In light of the apparent limitations on the scope of both MEC and SD, we might
adopt the view that the right formulation of EP is a hybrid, e.g., requiring only that
agents reject stochastically dominated options at stages of the aggregation process
where level comparisons are possible, while requiring that they choose options with
8In general, any decision principle that requires or permits non-neutral risk attitudes will require
level-normalization, since risk-seeking and risk-aversion are impossible without the ability to make
such absolute level comparisons. MEC avoids the need for intertheoretic level comparisons by
requiring risk neutrality.
253
maximal expected choiceworthiness at stages where unit comparisons but not level
comparisons are possible.9
I have no strong view on which precise formulation of EP has external ratio-
nal requiring force, except that it is either MEC, stochastic dominance, or some
combination thereof. In the next three sections I will try to show that enkratic
externalism can resolve the regress problem, regardless of the precise formulation of
EP on which it is based.
7.2 Two Regress Problems
Although I initially framed the regress problem as an argument for normative
externalism, that does not mean that conceding any form of externalism is enough
to make the problem go away. As we will see in the coming sections, significant
difficulties remain. To address these difficulties, it will be helpful to first divide
them into two categories: those that concern ideal rationality, and those that concern
non-ideal rationality.
Intuitively, the distinction is this: First, suppose that an ideally rational agent
will have beliefs about normative principles at every level of an infinite hierarchy,
and that at every level she will have positive credence in principles that disagree
about which of her options in a given choice situation are permissible. Given full
9It is more plausible that risk neutrality is a requirement of rationality in such contexts since the
absence of level comparability simply precludes non-neutral risk attitudes. On this line of thought,
perhaps an agent who distributes belief over a comparability class of cardinal theories that are
unit-comparable but not level-comparable relative to one another should compute expectations
internal to each theory by that theory’s own preferred decision procedure (which might involve
non-neutral risk attitudes) but then aggregate the comparability class according to (risk-neutral)
MEC.
254
cognitive access to her own credences and unlimited cognitive resources, how should
such an agent decide what to do? This is the ideal regress problem.
But second, even if we had a fully satisfactory solution to the ideal problem,
we still face the following difficulty: Any finite, non-ideal agent like ourselves will in
general not be able to contemplate an infinite hierarchy of normative principles every
time she has to decide what to do. Unless her credal sate with respect to higher-
order norms is extremely simple, she either will not have well-defined credences
about higher-order norms or, at least, will not have immediate cognitive access to
those credences. Moreover, even if she did have such credences and could access
them in order to decide what to do, she certainly lacks the cognitive resources to
carefully consider the various competing norms at each level and apportion her
credence correctly to the evidence. So given any account of how an ideally rational
agent should act when she is uncertain over an infinite hierarchy of norms, there is
a further question how if at all a non-ideal agent can approximate the requirements
of ideal rationality. This is the non-ideal regress problem.
As we will see, both problems remain puzzling even if we accept enkratic
externalism. Nevertheless, I will argue this commitment ultimately enables us to
resolve both problems.
7.3 The Problem of Ideal Rationality
In §7.1.3 I suggested that, if the enkratic principle could be taken as a second-
order principle of choice under moral uncertainty on which an agent is rationally
255
permitted to act straightaway, despite her uncertainties, then there would be simply
no need for third-order norms of decision-making under second-order uncertainty,
and hence no regress problem to contend with. But does EP alone obviate the
need for any second-order normative principles about which an agent might be
rationally required to account for her uncertainties? In other words, is (the correct
formulation of) EP a complete principle of rational decision-making under first-order
moral uncertainty, or does it simply constrain the field of second-order principles? If
the latter, then the regress problem is unresolved: for even if an agent may rationally
disregard EP-incompatible second order norms in which she has positive credence
(e.g. MFT, MFO, or the norm that tells her to minimize expected choiceworthiness),
she must still account for her uncertainty about EP-compatible norms. And then,
it seems, she will be stuck contemplating the same infinite hierarchy of competing
higher-order norms, made somewhat skinnier but no less tall by the constraints of
EP.
There are two different worries about the completeness of EP as a second-order
norm that deserve separate attention.
7.3.1 Gappiness
Both MEC and SD have the potential to leave agents with rational options,
i.e., require of an agent in a particular choice situation not that she choose one
particular option but that she choose one of some disjunction of options. MEC
allows for rational options in two circumstances: (i) when the expected value of one
256
or more options is undefined and (ii) when two or more options are tied for maximal
expected value and hence are judged equally rational. SD allows for rational options
whenever there is more than one option in a given choice situation that is not
stochastically dominated by any other option. Since rational options are a more
general phenomenon under SD than under MEC, let’s focus on SD in this section.
Two interpretations of the stochastic dominance formulation of EP are possi-
ble. On one interpretation, EP tells me that I am rationally required not to choose
stochastically dominated options, and is simply silent with respect to options that
are not stochastically dominated. On another interpretation, EP tells me both that
I am rationally required not to choose stochastically dominated options, and that I
am rationally permitted to choose any option that is not stochastically dominated.
Let’s call these principles Incomplete SD and Complete SD respectively.10
Incomplete SD seems to leave the regress problem unresolved. Consider an
agent A facing options O1, O2, O3, O4. Suppose that O1 stochastically dominates
O3 and O4, but neither O1 nor O2 is stochastically dominated by any other option.
Enkratic externalism, with EP understood as Incomplete SD, says that A is permit-
ted (indeed, required) to rule out O3 and O4 straightaway on the basis of EP. But
it does not say that A is permitted to choose O1 or O2 straightaway on the basis of
EP, since EP (understood as Incomplete SD) doesn’t say that O1 is permissible or
that O2 is permissible—it simply refrains from saying that they’re impermissible,
10Likewise, there is an incomplete interpretation of MEC that simply prohibits me from choosing
any option for which there is an alternative option with greater expected choiceworthiness, while
remaining silent on all other options, and a complete interpretation of MEC, according to which
I am rationally permitted to choose all and only those options for which there is no option with
greater expected choiceworthiness.
257
and indeed, from saying anything about them at all.
Now, suppose that A has positive credence in (at least) the following two
second-order principles: N1 “Under first-order moral uncertainty, always choose
the option that minimizes the probability of the worst possible outcome,” and N2
“Under first-order moral uncertainty, always choose the option that maximizes the
probability of the best possible outcome.” Both of these norms are compatible with
Incomplete SD, so the external force of this principle does not allow A to rule out
N1 or N2. Suppose that according to N1, A is rationally required to choose O1 and
according to N2, A is rationally required to choose O2. Then it seems that A has
credence in two conflicting second-order principles, that the version of externalism
we have adopted does not permit her to simply ignore this uncertainty, and therefore
that she must decide what to do in light of this second-order uncertainty in light
of third-order principles. Since she will presumably be uncertain in similar ways
about third order principles, we’re off to the races and the regress problem has been
reinstated.
What this indicates is that the version of EP to which enkratic externalism
ascribes external normative force must be complete, i.e., it must not simply prohibit
certain options but rather classify every option in a given choice situation as either
permissible or impermissible.11 This allows formulations of EP like Complete SD,
Complete MEC (according to which any option than which there there is no alter-
native with greater expected choiceworthiness is permissible), or a hybrid principle
11Arguably it would be sufficient for EP to always designate at least one option in any choice sit-
uation as permissible, even if it was silent with respect to other options, since the agent would then
have an option that she was rationally permitted to choose straightaway despite her uncertainties.
But I’ll set this possibility aside for the sake of simplicity.
258
that applies stochastic dominance reasoning at some stages of the aggregation pro-
cess and expectational reasoning at others, Complete (SD + MEC). It also allows
the sort of hybrid view discussed in the last chapter, e.g. stochastic dominance +
least common denominator, so long as the “least common denominator” principle
is understood as having external normative force. Such a hybrid principle might
be expressed as follows, where the subjective reason to choose an option O relative
to a given comparability class of theories is determined by aggregating the objec-
tive reasons to choose O according to each theory, based on the strongest decision
procedure common to all theories in the comparability class.
Hybrid EP An agent is rationally permitted to choose option O if and only if
there is no alternative option P such that, for any degree of subjective reason
d, the total probability of all comparability classes of theories according to
which there is subjective reason to choose P greater than or equal to d is
greater than or equal to the total probability of all comparability classes of
theories according to which there is subjective reason to choose O greater than
or equal to d, and for some d, the total probability of comparability classes
according to which there is subjective reason to choose P greater than or equal
to d is strictly greater than the corresponding probability for O.
So long as we attribute external normative force to a complete principle like Com-
plete SD, Complete MEC, or Hybrid EP, therefore, we don’t need to worry about
the regress problem reappearing via incompleteness, since any of these principles
will permit an agent to conclude straightaway that certain options are rationally
259
permissible, despite her second-order uncertainties.
7.3.2 Normalization Uncertainty
We might worry, however, that even a complete version of the enkratic princi-
ple will not do away with second-order uncertainty and the possibility of regress. For
both expectational and stochastic dominance reasoning in the context of interthe-
oretic uncertainty require procedures for intertheoretic normalization, about which
agents may well be uncertain. And while it is plausible that rejecting stochastically
dominated options, and perhaps even maximizing expected value, is constitutive
of rationality and hence a belief-independent rational requirement, it is much less
plausible that a particular intertheoretic normalization procedure is among the con-
stitutive requirements of rationality as well.12
It seems to me, however, that uncertainty about normalization methods is not
enough to generate a vicious regress of higher-order norms. Although, as in the
last section, the problem of normalization uncertainty arises for both MEC and SD
versions of the enkratic principle in roughly the same manner, it will simplify the
exposition in this context to focus on MEC, where the role of normalization methods
is more familiar. Consider by analogy a kind of normalization uncertainty that
might afflict expectational reasoning under merely empirical uncertainty. Suppose
that Gerald is a hedonistic utilitarian and is concerned with the hedonic welfare
interests of non-human animals as well as humans. Suppose, however, that he is
12Here we see the force of Weatherson’s worry that a “technical standard about resolving moral
uncertainty” should not be among the normative standards to which we hold agents independent
of their beliefs.
260
uncertain to what extent animals with relatively simple nervous systems experience
pleasure and pain. Thus when he considers grasshoppers, for instance, he has some
suspicion that the phenomenal intensity of the pain quale a grasshopper experiences
when one of its C fibers fires is less than the phenomenal intensity of a pain quale a
human being experiences when one of her C fibers fires, but he is unsure whether this
is the case, and if so, how great the difference in phenomenal intensity of suffering
per C fiber firing is. Thus although he cares both about preventing human C fibers
from firing and about preventing grasshopper C fibers from firing, he is unsure how
to compare these two values—i.e., unsure how to normalize the scale of human C
fiber firings with the scale of grasshopper C fiber firings.13
Presumably this uncertainty about normalization methods will not involve
Gerald in a vicious regress. Rather, given his commitment to maximizing expected
hedonic utility, he will simply average over the various possible normalizations in
order to decide how to weigh human pain against grasshopper pain. Likewise, a
morally uncertain agent who is committed to maximizing expected choiceworthiness
but is unsure how to normalize the various moral theories in which she has positive
credence should, in light of her commitment to MEC, simply take a probability-
weighted average over possible normalizations. This does not involve the introduc-
tion of some extra principle beyond MEC—rather, this sort of averaging is just what
expectational reasoning amounts to in the first place.14
13This question might have practical significance if, say, Gerald is a researcher who must decide
whether to devote his time and energy to developing a more effective antidepressant for humans
or a more humane pesticide for grasshoppers.
14Likewise if the formulation of EP that has external normative force is a stochastic dominance
principle, computing probabilities that a given option has a degree of choiceworthiness greater than
or equal to d involves accounting for one’s uncertainties about how to normalize the value scales of
261
7.4 The Problem of Non-Ideal Rationality
It may seem at this point that we’ve successfully avoided the threat of regress.
But there is another sort of regress problem that confronts non-ideal reasoners under
moral uncertainty, and that even a complete form of enkratic externalism alone does
not obviously resolve.
7.4.1 Consequentialism and the “Heuristic Regress”
To see this non-ideal regress problem, consider again the analogy with empiri-
cal uncertainty. Suppose I’m certain that hedonistic utilitarianism is true and that,
when I am uncertain about the consequences of my actions, I ought to maximize
expected hedonic utility. In principle, this requires that for each option O, I consider
every possible world that has non-zero probability conditional on my choosing O,
assign conditional probabilities to each of those worlds based on my evidence, and
calculate the total quantity of hedonic utility (pleasure minus pain) in each of those
worlds.
Of course, for any agent with anything like human cognitive capacities and
limitations, this is impossible. So what I must do instead is come up with some
heuristic for approximating the expected value of each option, relative to my evi-
dence. I might, for instance, choose to only focus on certain kinds of consequences
(e.g. QALYs saved or lost). I might choose to restrict the scope of the consequences
a particular normative theory with the scale of objective choiceworthiness simpliciter over which
the stochastic dominance principle is defined.
262
I consider, ignoring as improbable or unknowable the effects on my actions on the
distant future or regions of the world outside some immediate radius of effect. I
might choose to only consider certain evidence, e.g. a small number of relevant
peer-reviewed studies. Or I might simply defer to certain experts, e.g. taking my
estimates of the effects of some philanthropic activity from the estimates of an or-
ganization like GiveWell.
The problem is that there are many possible heuristic procedures for estimat-
ing the hedonic consequences of my actions, some better than others. Some will
require me to invest more time in deliberation, others less. Some will ignore poten-
tially important consequences, others will divert my deliberative resources toward
consideration of trivial consequences. Just like my moral theory requires me, as best
I can, to choose the actions with greatest expected hedonic utility, so it requires me
to adopt the deliberative procedures with greatest expected hedonic utility. And just
as I cannot compute the expected utility of my actions directly, I cannot compute
the expected utility of adopting a given deliberative heuristic directly. So it seems
that, in order to decide what deliberative heuristic to use, I must employ some
meta-heuristic. For instance, I might attempt to mimic the deliberative processes
of another agent whom I regard as a particularly good deliberator, or I might try to
remember the consequences that ensued on previous occasions when I deliberated
in a particular manner and choose the deliberative heuristic with the best track
record, or... But just as I am uncertain which first-order heuristic is optimific in
expectation, I will be uncertain which of these second-order heuristics is optimific
263
in expectation. And so, it appears, I am doomed to regress.15
Just as a full commitment to maximizing expected hedonic utility still appar-
ently leaves me vulnerable to a regress of higher-order heuristics, so a full commit-
ment to maximizing expected choiceworthiness leaves me vulnerable to a similar
regress.16 As a non-ideal agent, I lack the cognitive resources to compute the ex-
pected choiceworthiness of my options directly, and so must reply on heuristics.
I must try to choose the heuristic with greatest expected choiceworthiness (i.e.,
presumably, the heuristic that will yield first-order choices with greatest expected
choiceworthiness). But there are many possible heuristics for estimating expected
choiceworthiness, and I am unsure which is best. And so on.
An ideal agent who has perfect access to her own credal states and unlimited
cognitive resources is not vulnerable to this regress, since she doesn’t need to rely
on heuristics in the first place—instead, she can compute the expected utility or
expected choiceworthiness of her options directly on the basis of her credences (and,
presumably, apportion those credences directly to her evidence without the need for
epistemic heuristics).
Structurally, this heuristic regress problem resembles the original regress of
higher-order norms that enkratic externalism was meant to avoid. The difference is
simply that the norms at each level are not final principles of rational requirement
which the agent believes to be true with certain probabilities, but rather heuristic
principles for approximating the expected value of options, given certain lower-order
15This problem was brought to my attention by Ben Holgu´ın, who to the best of my knowledge
is the first to have noticed it.
16Once again, it is convenient to focus on EEMEC rather than EESD, although the heuristic
regress problem arises on either view.
264
inputs, which the agent believes to be optimal with certain probabilities.
Luckily, I will argue, there is a solution to this new regress problem, uniquely
available to non-ideal agents. Describing this solution will allow us to address a
question on which I have so far been silent: namely, how non-ideal reasoners can
apply (or rather, approximate) complex decision procedures for choice under moral
uncertainty like those described in the last two chapters.
7.4.2 Non-Ideal Deliberation under Moral Uncertainty: The Power
of Priors
I propose that we adopt the following model of non-ideal practical delibera-
tion.17
1. A non-ideal (= cognitively limited) but rational (= enkratic) agent A deliber-
ating between options O1, O2, ...On enters deliberation with access to a number
of standing beliefs : in particular, (i) a belief about which of her options have
maximal expected choiceworthiness, (ii) a belief about the expected marginal
benefit of further deliberation to update her beliefs about the expected choice-
worthiness of her options in light of her evidence, and (iii) a belief about the
expected marginal cost of such deliberation.18 Call the set of options that an
17Again, for ease of presentation, I assume EEMEC throughout the following exposition, i.e., I
assume that a rational agent always chooses options that have maximal expected choiceworthiness,
whatever her beliefs about MEC. The exposition can be easily adapted to other versions of enkratic
externalism, e.g. by replacing all references to “maximal expected choiceworthiness” with “not
stochastically dominated.”
18These beliefs can be understood dispositionally. Thus, an enkratic agent’s standing belief about
which of her options is most choiceworthy in expectation can be identified with the answer to the
question, “If you had to choose right now, which of your options would you choose?” Likewise,
insofar as she is rational, her beliefs about the expected marginal benefit and cost of further
265
agent believes, at a given moment of deliberation, to have maximal expected
choiceworthiness her enkratic set.
2. Deliberation, for a non-ideal but rational agent, consists in updating her beliefs
about the choiceworthiness of her options in light of her evidence. While an
ideal agent can update costlessly (instantaneously and effortlessly), a non-ideal
agent cannot. Rather, she must invest time in considering the relevance of the
available evidence to her various empirical and normative beliefs. This might
consist, among other things, of applying heuristics, developing meta-heuristics
by which to assess those heuristics, applying the meta-heuristics, and so forth.
But each of these steps requires scarce resources of time and effort.
3. As an agent deliberates, her standing beliefs will change, sometimes con-
sciously and sometimes unconsciously. In particular, she may update the
content of her enkratic set (i.e., change her beliefs about which of her op-
tions have maximal expected choiceworthiness), and she will also change her
beliefs about the expected marginal benefit and cost of further deliberation.
(Strictly speaking, this latter process may not involve any belief change, since
her beliefs about the expected marginal value of deliberation are time-indexed,
i.e., the value of deliberating for one minute longer at time t is distinct from
the value of deliberating for one minute longer at time t+ 1.)
4. In nearly all circumstances (apart from cases of genuine decision-theoretic
paradox), the expected marginal cost of further deliberation will come to ex-
deliberation can be inferred from her dispositions to continue deliberating or to stop deliberating
and act straightaway.
266
ceed the expected marginal value after a finite period of deliberation. This
happens for at least two reasons: (a) Deliberation will in general have dimin-
ishing marginal benefit. A rational agent will first examine the strongest and
most important evidence, consider the most important and most uncertain
features of her various options, etc. The longer she deliberates, the more of
the low-hanging fruit of deliberation she will exhaust, and the less she can
expect further deliberation to improve her ultimate choice, per unit of time
invested. (b) Nearly all choices are to some extent time-sensitive—that is, the
quality of the agent’s options will start to decay if she deliberates too long.
While it can sometimes be rational to delay certain choices even for many
years in order to acquire new evidence and improve my beliefs (Wise, 2013),
it is never or almost never rational to delay forever, and typically not rational
to delay for very long at all. If I wait until I am 45 years old to choose a career
path, for instance, the deliberation this delay enables will be more than offset
by the decreased quality of my career options.
5. Whenever a non-ideal agent comes to believe that the expected marginal cost
of further deliberation would exceed the expected marginal benefit, she is ratio-
nally required to stop deliberating and choose based on her standing beliefs—
that is, choose one of the options presently belonging to her enkratic set.19
Figure 7.1 offers a graphical illustration of the model. An agent A enters
19Note that this model has both empirical and normative components. The empirical component
is the assumption that agents begin with certain standing beliefs that change in the course of
deliberation. The normative component is the claim that an agent ought to deliberate so long as
the expected marginal benefit of deliberation exceeds the expected marginal cost, and ought to
cease deliberating and choose straightaway as soon as this inequality is reversed.
267
into deliberation believing that some subset of her options have maximal expected
choiceworthiness, but also believing that the expected marginal benefit of further
deliberation exceeds the expected marginal cost, and hence that she should delib-
erate further before making her choice. As she deliberates, her beliefs about which
of her options have maximal expected choiceworthiness change, as do her beliefs
about the costs and benefits of further deliberation. After some finite period of
deliberation, the lines representing expected marginal cost and benefit of further
deliberation cross (at what I have labeled the “decision point”). At that point, A
rationally ought to choose one of the options in her enkratic set, i.e. one of the
options that she then regards as having maximal expected choiceworthiness.
{O1, O2, O3} {O1, O2} {O2} {O2, O3}enkratic set · · ·
← time →
←
va
lu
e
→
· · ·
expected marginal cost
of deliberation
· · ·
expected marginal ben-
efit of deliberation
decision
point
Figure 7.1: Non-Ideal Practical Deliberation
This model of deliberation strikes me as psychologically realistic. Consider a
simple case of non-moral deliberation: I walk into an ice cream shop intending to
buy a cone. I am immediately confronted with a large set of practical options. I
begin with some standing belief about which of my options is most choiceworthy
in expectation (choiceworthiness in this circumstance being, presumably, tightly
correlated with my own hedonic utility): If I had to choose straightaway then I
268
would choose, say, mint chocolate chip. But rather than choosing straightaway, I
stand a few steps away from the counter, deliberating, since I’m uncertain which
option will bring me most hedonic utility and it seems to me that the potential
improvement of my choice to be had by a brief period of deliberation outweighs
the cost of delayed ice cream. So I deliberate, imagining what each of my options
would be like, calling up memories of the last time I ate various flavors of ice cream,
perhaps recalling my slight disappointment the last time I chose mint chocolate chip.
After thirty second of deliberation my standing beliefs about my options change,
and I come to believe that chocolate raspberry is most choiceworthy in expectation.
After another ten seconds, I conclude that further delay is unlikely to significantly
improve my choice, and I choose chocolate raspberry straightaway. This seems to
me a realistic description of the sort of deliberation that actual, non-ideal agents
engage in all the time, even though most of us do not explicitly invoke terms like
“expected choiceworthiness” and “marginal value” when we deliberate about ice
cream.
How does all of this solve the heuristic regress problem, though? Simply thus:
A non-ideal agent is not rationally required to continue ascending the hierarchy of
higher-order heuristics until achieving certainty. It may turn out that what she ratio-
nally ought to do, in some circumstances, is compare the advice of various heuristic
decision procedures, evaluate those heuristics in light of higher-order heuristics, etc.
But at some point the rational thing to do is to cease deliberating and act, rather
than engage in any form of further deliberation.
To put it in terms of the argument for externalism with which we began this
269
chapter, we need not accept externalism with respect to heuristic norms for estimat-
ing expected choiceworthiness, because we can reject premise (1) of the argument for
externalism when considering non-ideal, cognitively limited agents. Such an agent
need not always account for her uncertainty about the optimality of a given heuristic
norm, because it may be positively irrational for her to invest the time and effort
needed to do so.
This doesn’t defeat the regress argument for externalism, because the premises
of that argument are still true with respect to ideal agents for whom the sort of delib-
eration I have just described (updating beliefs about the expected choiceworthiness
of options based on one’s credal state and evidence) is costless. The argument
therefore succeeds in showing that pure internalism has unacceptable consequences
with respect to the requirements of ideal rationality. But since the heuristic regress
problem only arises for non-ideal agents, it can be solved by the observation that,
with respect to such non-ideal agents, premise (1) is false.
In summary: The regress problem does force us to accept that at least one
normative principle has external, belief-independent rational requiring force. I have
argued that this one principle is the enkratic principle, rightly formulated, and
considered some candidates for that right formulation, without attempting to give
any final resolution of the question. Enkratic externalism, I have argued, solves the
regress problem for ideal agents by providing a second-order principle of decision-
making under first-order moral uncertainty that obviates the need for any hierarchy
of higher-order normative principles. A different regress problem seems to confront
non-ideal agents who must use heuristic norms to approximate ideal principles like
270
MEC. But, I have argued, a plausible model of non-ideal practical deliberation shows
that such agents are not rationally required to keep ascending the infinite hierarchy
of heuristic norms, resolving this problem as well.
7.5 Conclusion
I have had two main objectives in this dissertation. The first objective has
been to defend the rational relevance of moral uncertainty, and more specifically
the rationality of moral hedging (at least for many agents in many circumstances).
My positive case for these claims has centered on the enkratic conception of ratio-
nality, according to which the requirements of practical rationality are grounded
in an agent’s beliefs about objective value or choiceworthiness. But I have also
attempted to defend hedging against a variety of worries and objections from the
recent literature.
My second objective has been to develop, in outline, a theory of rational
decision-making under moral uncertainty. Here the central idea has been that of
content-based aggregation. I have argued that, if how an agent should aggregate
the value assignments of her competing theories depends on non-surface features of
those theories’ content, then the picture of rational decision-making that is likely
to emerge is one of multi-stage aggregation of a nested structure of comparability
classes bound together in various ways and to varying degrees by shared theoretical
propositions.
The philosophical investigation of rational choice under moral uncertainty is
271
still in its infancy, and even if the central claims I have advanced in this dissertation
are correct, many more questions remain to be answered. Nevertheless, I hope to
have shown that there is good reason to consider these questions and that we can
have some reasonable hope of arriving at satisfactory answers.
272
Appendix A: Moral Uncertainty for Deontologists
[Note: The following is the penultimate draft of “Moral Uncertainty for Deontolo-
gists,” forthcoming in Ethical Theory and Moral Practice (Tarsney, forthcoming b).
In it I argue that stochastic dominance provides a natural aggregation procedure
for deontological theories, supporting a threshold principle on which it is rationally
impermissible to choose options that carry a risk greater than .5 of violating a deon-
tological constraint when there are less morally risky alternatives available. Though
the paper focuses on a representation of deontological value assignments suggested
by Colyvan et al. (2010) that has the structure of the extended real number line
[−∞,+∞], the arguments are equally applicable to merely-ordinal or multidimen-
sional, lexical representations of deontological theories.]
A.1 Introduction
We are often uncertain about what we morally ought to do. Such uncertainty
can arise in two importantly distinct ways. On the one hand, some moral uncertainty
is grounded in empirical uncertainty: for instance, is this substance that I am about
to put in my friend’s coffee sweetener, or is it arsenic (Weatherson, 2014)? On the
other hand, there is what we might call purely moral uncertainty, uncertainty about
273
the basic principles of morality, which no empirical information could directly or
easily resolve: for instance, is it permissible to tell my friend a white lie about his
new haircut?
Moral consequentialists have typically had the most to say about uncertainties
of both kinds. With respect to empirically based moral uncertainty, consequential-
ists typically claim that agents should choose practical options that have maximal
expected value (the expected value of an option O being found by summing, over
all possible outcomes of that action, the value of that outcome times its probability
conditional on one’s choosing O). This expectational formula can be adjusted in a
variety of ways, e.g. to allow for risk aversion or risk seeking.
The study of purely moral uncertainty is still in its infancy, but to date the
dominant positive approaches in this literature have shared an almost exclusively
consequentialist, expectational flavor: Lockhart (2000), Ross (2006), Sepielli (2009),
and MacAskill (2014), for instance, all defend expected value approaches to decision-
making under purely moral uncertainty. The critics of this approach, on the other
hand, have predominantly been those who deny the need for a theory of rational
choice under purely moral uncertainty, because they deny that what an agent ought
to do depends on her purely moral beliefs or degrees of belief (e.g. Weatherson
(2014), Harman (2015), Hedden (2016)).1
Expectational approaches to uncertainty, however, seem ill-suited for if not
1It is standard in the literature to reserve the term “moral uncertainty” for what I am calling
purely moral uncertainty. I use the term more broadly in this paper both to avoid repeated
inelegant references to “morally relevant empirical uncertainty” and because one of my aims will
be to emphasize the continuity between empirically-based and purely moral uncertainty, so it is
helpful to have an umbrella term that refers to both at once.
274
actively incompatible with non-consequentialist moral theories like Kantian deon-
tology, most obviously because (unlike, say, classical utilitarianism) there is no nat-
ural way of interpreting these non-consequentialist theories as assigning degrees of
rightness and wrongness that are amenable to being multiplied by probabilities and
summed to yield quantitative expectations. Even setting aside any thought that
theories of choice under uncertainty must take an expectational form, we will see
in the next section that there are substantial reasons to doubt whether Kantians et
al can provide any plausible answer to certain inescapable questions about moral
choice under uncertainty.
The purpose of this paper, however, is to suggest one way in which deontol-
ogists can say something definite and precise about decision-making under moral
uncertainty, of both the empirically based and purely moral varieties. I will propose
that stochastic dominance reasoning provides powerful motivation for a threshold
principle of choice under uncertainty, which implies inter alia that when an agent
A faces a choice between practical options O and P, where P is known for certain
to be morally permissible and the status of O is uncertain, A subjectively ought
not choose O if her credence that O is objectively impermissible is greater than or
equal to .5.
In the next section I describe in greater detail the challenge to deontology
posed by both kinds of moral uncertainty, which some in the recent literature (e.g.
Jackson and Smith (2006, 2016) and Huemer (2010)) have characterized as a decisive
objection at least to absolutist forms of deontology. In §A.3 I introduce the idea of
stochastic dominance reasoning. In §A.4 I show that stochastic dominance reason-
275
ing yields substantive and plausible conclusions for how an agent who accepts the
general framework of deontological, agent-centered constraints (whether absolute or
defeasible) ought to act under both kinds of moral uncertainty, which in at least the
simplest cases will take the form of a straightforward threshold principle. Finally,
§A.5 addresses a powerful objection to threshold principles given by Jackson and
Smith, that such principles seem to violate “ought” agglomeration, since a pair of
actions each of which is below the threshold for acceptable moral risk can, in com-
bination, exceed that threshold. I argue that, in combination with the observation
that deontological moral evaluation is relativized to the particular choice situations
in which acts arise, the stochastic dominance approach can overcome this objection
and preserve “ought” agglomeration.
A.2 Absolutist Deontology and Uncertainty
Following Portmore (2016, p. 7), a deontological ethical theory may be charac-
terized as one that denies what all consequentialist theories affirm, namely, that the
objective deontic status and degrees of objective rightness/wrongness of any act are
determined by how the outcome or prospect of that act compares to the outcomes
or prospects of its alternatives.2 More precisely (to avoid accidentally counting as
deontological views that deny that acts have deontic status or degrees of rightness),
a deontological theory is one that at least sometimes attributes objective deontic
status and/or degrees of objective rightness to acts for reasons that do not depend
2A prospect is an objective probability distribution over outcomes; hereafter I will use “out-
come” to mean “outcome or prospect.” I will likewise speak of “degrees of rightness” when I mean
“degrees of rightness or wrongness.”
276
on how the outcomes of those acts compare to the outcomes of alternatives. An
imperfect litmus test for whether a theory is deontological is whether it endorses
the existence of agent-centered constraints, act types that are treated as prima fa-
cie objectively wrong or prohibited and such that a token of that act type is at
least sometimes objectively wrong or prohibited even when its performance would
prevent multiple future tokens of the same act type that are similar in all ethically
relevant respects to the prohibited token.3 Thus a deontological theory might hold,
for instance, that it is wrong for me to lie now even though my lie will prevent three
future acts of lying that are similar to mine in all ethically relevant respects (except
that they will not themselves prevent future instances of lying).4
An absolutist deontological theory holds that, for some act types characterized
non-relationally (i.e., whether an act is an instance of the act type does not depend
on any fact about how it relates or compares to its available alternatives), not only is
their deontic status not fully determined by how their outcomes compare to those of
alternatives, but all acts of that act type have the same deontic status regardless of
any facts about their outcomes or those of alternatives. For instance, an absolutist
theory might hold that all acts that have the non-relational property of being a lie
(asserting something that the agent believes to be false with the intent that another
3The litmus test is imperfect because ethical theories that incorporate these prima facie deonto-
logical features can be “consequentialized” (Portmore, 2011). As we will see below, a desideratum
for deontological approaches to uncertainty is to avoid a drift in the direction of expectational
reasoning about the risk of constraint violations, which, if fully embraced, amounts to evaluating
actions by a comparison of their prospects and hence to consequentialism, despite the presence of
agent-centered constraints.
4Agent-centered constraints may also be positive, e.g., I am required to keep a promise even if
my doing so will result in several fewer future instances of promise keeping, or more instances of
promise violation.
277
agent believe it, in a circumstance where the second agent has not consented to
being so deceived) is wrong even when telling the lie would prevent the deaths of a
thousand innocent people who will be killed if the agent does anything other than
lie.
Deontological theories in general and absolutist theories in particular face the
following challenge: How should an agent decide what to do when she is uncertain
whether a given act is an instance of a deontologically prohibited (or required) act
type? For instance, consider the following case.
Possible Promise I have unexpectedly found myself in possession of tickets to the
Big Game this afternoon. But as I am celebrating my good fortune, it occurs
to me that I may have promised my friend Petunia that I would help her
repaint her house later today, at just the time when the game is being played.
I have a vague recollection of making such a promise but can’t remember with
any confidence. I have no immediate way of contacting my friend and must
decide, here and now, whether to go to the game at the risk of breaking a
promise, or go to her house, avoiding the moral risk but missing the game.
Possible Promise is an instance of empirically based moral uncertainty, but we
can describe a closely related case of “pure” moral uncertainty.
Dubious Promise A week ago, Petunia sent me a text message asking if I would
help paint her house today. I replied, saying that I would. Unbeknownst to
Petunia, however, I was in the hospital at the time, recovering from a minor
operation and under the influence of a fairly strong narcotic painkiller. By the
278
time the influence of the painkiller subsided, I had completely forgotten my
conversation with Petunia, and only just remembered it a moment ago, while
planning my trip to the Big Game.
I take it that, under these circumstances, I may reasonably be uncertain whether I
am morally required to skip the game and help my friend, even if I am certain of a
background deontological conception of morality on which an ordinary, fully capac-
itated promise would have been morally binding.5 Because my moral uncertainty
does not trace back to any empirical uncertainty (e.g. about what I told Petunia,
my mental state at the time, or her expectations of my future behavior), it is pure
moral uncertainty, uncertainty about the content of a basic moral principle.
In either version of the case, the question is what I should do given my uncer-
tainty about what morality requires of me. To frame this problem rightly, we must
introduce a familiar distinction between objective and subjective normative proper-
ties. The objective normative properties of an act depend on facts about the world,
but not (in general) on facts about an agent’s beliefs or evidence. The subjective
normative properties of an act, conversely, depend on facts about the agent’s beliefs
and/or evidence but not on facts about the world independent of the agent’s be-
liefs/evidence. For instance (borrowing a helpful illustration from Portmore (2016)),
suppose you are tasked with defusing a bomb and must decide whether to cut the
green wire or the red wire. You believe, and your evidence overwhelmingly indi-
cates, that cutting the green wire will defuse the bomb while cutting the red wire
5Adjust the strength of the imagined painkillers as needed until this case strikes you as one of
substantial moral uncertainty.
279
will cause the bomb to explode. But in fact, cutting the red wire will defuse the
bomb, while cutting the green wire will cause the bomb to explode. In this case,
it is objectively right but subjectively wrong to cut the red wire, and objectively
wrong but subjectively right to cut the green wire.6
This distinction in hand, let’s return to the case of the Possible Promise. There
are various things a deontologist can say about this case. However, it is generally
agreed that she should not say any of the following: (1) It is subjectively wrong
for me to go to the game iff I have any positive credence that I promised to help
Petunia paint her house and hence that it would be objectively wrong for me to go
to the game. (2) It is subjectively wrong for me to go to the game iff I am certain
(i.e., credence 1) that I promised to help Petunia paint her house and hence that it
would be objectively wrong for me to go to the game. (3) It is subjectively wrong
for me to go to the game iff I did in fact promise to help Petunia paint her house
and hence it would in fact be objectively wrong for me to go to the game.
Answer (1) is unsatisfactory because I can rarely if ever be certain that a
6Among the possible objective normative properties of an act are being objectively right or
wrong; being objectively obligatory, permissible, or forbidden; being what the agent objectively
ought or ought not do; and being required, permitted, or forbidden in virtue of an agent’s all-things-
considered objective reasons. For each of these objective properties, we may define a subjective
counterpart. In general, whether an action has a given subjective normative property will depend
on the agent’s beliefs and/or evidence regarding the corresponding objective property. For our
purposes, the distinctions between the various normative properties in each category will not mat-
ter very much, so I will sometimes switch between talking e.g. of objective wrongness, objective
prohibition, and objective “ought not” as convenience dictates. Our central focus, however, will
be on the interaction between an agent’s beliefs about the objective properties of being (deonto-
logically) morally required or morally prohibited and the subjective property of being rationally
required or rationally prohibited, that is, being required or prohibited in virtue of an agent’s all-
things-considered subjective reasons. These concepts are related by way of objective reasons: a
rational agent who believes, e.g., that she is deontologically morally required to choose some prac-
tical option O will believe that she has decisive or at least very strong objective reason to choose
O, and (at least in general) it is an agent’s beliefs about her objective reasons that give rise to her
subjective reasons.
280
course of action does not violate a promise or some other source of deontological
obligation. Hence, the principle behind (1) implies that all or nearly all acts are
subjectively wrong. Answer (2) is unsatisfactory for a symmetrical reason: I can
rarely if ever be certain that an act would violate a deontological obligation, and
hence if those obligations are only ever relevant to the subjective normative prop-
erties of my actions under conditions of certainty, they are effectively inert. Answer
(3) is unsatisfactory because, by collapsing subjective deontic status into objective
deontic status, it fails to serve the purpose for which subjective normative concepts
are introduced, namely to provide epistemically imperfect agents with useful action
guidance in the face of uncertainty.
Absent a better answer than these, deontology is in trouble. It will seem either
to paralyze action entirely (as per (1)), to collapse into consequentialism insofar as its
agent-centered constraints are never practically relevant to epistemically imperfect
agents (as per (2)), or to be simply useless as guides to action for such agents (as
per (3)). Thus Jackson and Smith (2006, 2016) and Huemer (2010), inter alia,
have suggested that the problem of uncertainty poses a fatal objection at least to
absolutist forms of deontology.
There may be many lines of response more promising than (1)-(3) that the
deontologist can pursue in the face of this challenge. But the major focus of the
recent literature has been on variants of the threshold view, according to which, for
any objectively prohibited act type W there is some threshold t, 0 < t < 1, such
that it is always subjectively wrong for me to choose a practical option O if my
281
credence that O would be an instance of W is greater than or equal to t.7
The threshold view faces an important and widely acknowledged technical
objection, which we will address in §A.5. But it also faces much simpler worries
about motivation that have yet to be seriously addressed. First, why should there
be any credal threshold at which an act abruptly switches status, from subjectively
permissible to subjectively impermissible? What plausibility can there be, say, in
the idea that if I have .27 credence that going to the game would break a promise
to my friend, there is nothing at all wrong with going, but if I have .28 credence, I
am strictly prohibited from going? Second, and closely related, what could serve to
make any particular credal threshold the right one? Why should the threshold be
set here rather than there? And how can we ever hope to know where it has been
set?8
In the next two sections, I will propose an answer to these questions: namely,
that stochastic dominance reasoning both explains why there should be sharp thresh-
olds for subjective permissibility and fixes the value of those thresholds in a perfectly
non-arbitrary manner, thereby vindicating a version of the threshold view.
7Jackson and Smith (2006) treat the threshold view as the most natural response the absolutist
deontologist can give to the problem of uncertainty, though as we will see in §A.5 they argue
that it is subject to a conclusive objection. Hawley (2008) and Aboodi et al. (2008) both defend
versions of the threshold view against this objection. Isaacs (2014) suggests that it is subjectively
permissible for an agent to take a morally risky action only if she knows that that action would
not be an instance of a deontologically prohibited type, which is a close cousin of the threshold
view provided one accepts that belief above some credal threshold is a necessary condition for
knowledge.
8This objection from arbitrariness is pressed by both (Portmore, 2016, pp. 10-11) and (Jackson
and Smith, 2016, p. 284).
282
A.3 Stochastic Dominance
The idea of dominance reasoning is familiar to philosophers from game-theoretic
contexts like the prisoner’s dilemma. If I am uncertain about the state of the world,
but certain that, given any possible state of the world, option O is more choicewor-
thy than option P, then O is said to strictly dominate P. If I am certain (i) that,
given any possible state of the world, O is at least as choiceworthy as P, and (ii)
that given some state(s) of the world O is more choiceworthy, then O is said to
weakly dominate P.
Stochastic dominance extends these familiar ideas as follows. A practical op-
tion O stochastically dominates an alternative P iff, relative to the agent’s doxastic
or epistemic state
1. For any degree of choiceworthiness d, the probability that O is (or will turn out
to be) at least as choiceworthy as d is equal to or greater than the probability
that P is (“”) at least as choiceworthy as d, and
2. For some degree of choiceworthiness d, the probability that O is (“”) at least
as choiceworthy as d is strictly greater than the probability that P is (“”) at
least as choiceworthy as d.
An illustration: Suppose that I am going to flip a fair coin, and I offer you a
choice of two tickets. The Heads ticket will pay you $1 for heads and nothing for tails,
while the Tails ticket will pay you $2 for tails and nothing for heads. The Tails ticket
neither strictly nor weakly dominates the Heads ticket because, if the coin lands
283
Heads, the Heads ticket will yield a more desirable outcome. But the Tails ticket
does stochastically dominate the Heads ticket. There are three possible outcomes
of the game, which in ascending order of desirability are: winning $0, winning $1,
and winning $2. The two tickets offer the same probability of an outcome at least as
good as $0, namely 100%. Likewise, they offer the same probability of an outcome
at least as good as $1, namely 50%. But the Tails ticket offers a better chance of
an outcome at least as good as $2, namely 50%, versus the 0% probability offered
by the Heads ticket.
The principle that it is never rational to choose a stochastically dominated
option is extremely compelling. First, note that unlike the stronger principle that
a rational agent should always maximize expected utility or expected value, the
stochastic dominance principle places no unwelcome constraints on an agent’s atti-
tude toward risk. For instance, if I am offered a choice between a ticket that pays
$1 if a fair coin lands heads on a single flip and a ticket that pays $4 if a fair coin
lands tails on both of two flips, it may be rational for me to take the risk averse
option and select the first ticket even though the expected payoff of the second
ticket is twice as large. Unlike an expectational principle, stochastic dominance is
silent in this sort of case: the second ticket offers a better chance of a payoff at
least as good as $4 (namely, 25% rather than 0%), but the first ticket offers a better
chance of a payoff at least as good as $1 (namely, 50% rather than 25%), so neither
option is stochastically dominated.9 Relatedly (as we will see at greater length in
9Of course, expected utility theory will only requires that you choose the riskier ticket if the
payoffs are given in utiles or some other unit of pure value, rather than in dollars. But the point can
be made just as easily in these terms, since it is commonly believed that it is sometimes rationally
permissible to be risk averse with respect to utility.
284
the next section), stochastic dominance reasoning does not entail the implausible
results of expectational reasoning with respect to very small chances of very large
or infinite payoffs (as in Pascal’s Wager), since paying some definite cost for a very
small chance of a very large reward will not stochastically dominate the null option
of forgoing both the cost and the chance of reward.
Given these observations, it is hard to see how one can ever make the case
for selecting a stochastically dominated option. It is, in general, possible that the
stochastically dominated option will yield a more desirable payoff than the dominant
option, but whatever level of payoff one is most concerned with getting, the dominant
option offers an equally good or better chance of a payoff at least that desirable. The
way in which that payout comes about is, by definition, a matter of indifference: If,
say, I preferred winning $0 with a Heads ticket to winning $0 with a Tails ticket in
the first case of stochastic dominance reasoning given above, then the desirability
of the possible outcomes would depend on more than the monetary payout, and the
Tails ticket would no longer stochastically dominate the Heads ticket. If I do not
have such a preference, then the mere fact that the Heads ticket might turn out
to yield a more desirable outcome is not a sufficient reason for choosing it over the
Tails ticket.10
In the next section, I will show that the principle of eliminating stochastically
10In more formal terms, a stochastically dominant option will be preferred to a stochastically
dominated alternative by any agent whose utility function is monotonically increasing, i.e., for
whom a given chance of a higher-value outcome is preferred to the same chance of a lower-value
outcome, all else being equal (Hadar and Russell, 1969, p. 28). The notion of stochastic dominance
defined above, and which I employ throughout the paper, is sometimes called first-order stochastic
dominance, distinguishing it from higher-order stochastic dominance principles that place increas-
ingly stringent constraints on an agent’s risk attitudes.
285
dominated options can justify a threshold principle for choice under deontological
moral uncertainty.
A.4 Stochastic Dominance and Deontological Uncertainty
Let’s return to the pair of cases given in §A.2, in which I am unsure whether
I am required to skip the Big Game and help my friend paint her house, either
because I am empirically uncertain whether I told her I would help or because I
am purely morally uncertain whether some past act constituted a morally binding
promise.
Let’s provisionally suppose the following about these cases: (1) The back-
ground deontological conception of morality that I accept is absolutist, and therefore
treats the objective requirement to keep one’s promises as lexically stronger than
any non-moral reason (like the reasons stemming from my desire to see the game).
(2) This means that (a) an act of promise keeping is better or more choiceworthy
than any act that violates a moral obligation or is morally neutral (at least so long
as the agent is appropriately motivated, e.g. by considerations of duty), and (b) an
act of promise breaking is worse or less choiceworthy than any act that fulfills a
moral obligation or is morally indifferent (at least so long as no special exculpatory
condition applies).11 (We will later consider what happens when we relax these
11It may seem that we are some risk of describing deontological appraisal of actions in evaluative
rather than normative terms, when part of what distinguishes deontological approaches to ethics
is that they take normative notions to be primary, where consequentialists give primacy to the
evaluative. But nothing turns on the choice of evaluative or normative language—we can, for
instance, take the description of one action as “better” than another from the standpoint of a
given agent as mere shorthand for the normative claim that it is more strongly supported by the
agent’s reasons, without raising any new difficulties for the argument given in this section.
286
assumptions.)
Accepting these assumptions allow a decision-theoretic representation of de-
ontological moral considerations, along lines suggested by Colyvan, Cox, and Steele
(2010). In their model, absolutist deontology is characterized by the following ax-
iomatic extension of standard utility theory:
D1* If [outcome] Oij is the result of an (absolutely) prohibited act, then any ad-
missible utility function u must be such that u(Oij) = −∞.
D2* If [outcome] Oij is the result of an (absolutely) obligatory act, then any ad-
missible utility function u must be such that u(Oij) = +∞. (Colyvan et al.,
2010, p. 512)
On this model, while actions from duty have “infinite” positive value and actions
against duty have “infinite” negative value, actions that neither violate nor fulfill
duties have finite value determined, presumably, by prudential or desire-based rea-
sons and perhaps by other sorts of moral reason (e.g. consequence-based reasons of
benevolence) that do not generate absolute obligations or prohibitions.12
The mentions of infinite positive and negative value in Colyvan et al’s axioms
need not be taken too literally. As they point out (pp. 521ff), decision-theoretic
models of non-consequentialist ethical theories may be descriptively adequate with-
out being explanatory: A Kantian does not avoid lying, for instance, because she
regards acts of lying as having infinite disvalue. Nevertheless, the fact that, for a
12Talk of the “value” or “utility” of actions should be understood as simply a less cumbersome
stand-in for talk of choiceworthiness—i.e., it should be understood as denoting a normative rather
than an evaluative property of actions.
287
Kantian, moral obligations and prohibitions are lexically stronger than prudential
reasons can be accurately represented by treating the value of an action from duty
as the upper bound on the scale of reason strength, and the disvalue of an action
against duty as the lower bound.
This representation in hand, we can employ stochastic dominance reasoning
to draw conclusions about how a committed deontologist ought to respond to moral
uncertainty. Consider the case of the Possible Promise. Suppose that the prudential
value of seeing the Big Game is +20, while the prudential value of helping Petunia
paint is +5 (and that I know these prudential facts with certainty). It follows that
the option of helping Petunia will stochastically dominate the option of going to the
game if and only if the probability that I promised to help Petunia is greater than
or equal to .5.
Suppose, first, that the probability is exactly .5, that is, I regard it as equally
likely that I did as that I did not make a promise to Petunia. Then I may reason
as follows. My decision has four possible outcomes (Table A.1), which in order
from best to worst are: (i) I did make a promise to Petunia, but I go to the game,
violating that promise (−∞). (ii) I did not make a promise to Petunia, but skip the
game to help her paint anyway (+5). (iii) I did not make a promise to Petunia, and
go to the game, violating no obligation (+20). (iv) I did make a promise to Petunia,
and I skip the game to help her paint, thereby fulfilling that obligation (+∞).
Given my credences, helping Petunia paint (option P) stochastically dominates
going to the game (option G): P and G have the same chance of producing an
outcome at least as good as −∞ (100%); P has a better chance of producing an
288
outcome at least as good as +5 (100% vs. 50%); P and G have the same chance of
producing an outcome at least as good as +20 (50%); and P has a better chance of
producing an outcome at least as good as +∞ (50% vs. 0%) (Table A.2). Thus, if
I accept the principle that I should never choose a stochastically dominated option,
I am compelled to choose P rather than G.13
Promised (.5) Didn’t Promise (.5)
Painting +∞ +5
Game −∞ +20
Table A.1: Possible Promise v1, payoff matrix
−∞ +5 +20 +∞
Painting 1 1 .5 .5
Game 1 .5 .5 0
Table A.2: Possible Promise v1, probability of payoff ≥ x
On the other hand, suppose that my credence that I promised to help Petunia
paint is only .49. In that case, option G has a better chance than option P of yielding
an outcome at least as good as +20 (51% vs. 49%), so P does not stochastically
dominate G (Tables A.3-A.4).
Promised (.49) Didn’t Promise (.51)
Painting +∞ +5
Game −∞ +20
Table A.3: Possible Promise v2, payoff matrix
13The talk of “outcomes” should not be seen as illicitly consequentialist. Violating or fulfilling
a deontological obligation is an “outcome” only in the formal sense of being a distinct act-state
combination valued differently than other act-state combinations. Calling these “outcomes” does
not imply, for instance, that the wrongness of breaking a promise has anything to do with its
causal consequences.
289
−∞ +5 +20 +∞
Painting 1 1 .49 .49
Game 1 .51 .51 0
Table A.4: Possible Promise v2, probability of payoff ≥ x
Notice that this argument applies to Dubious Promise, the case of pure moral
uncertainty, just as it applies to Possible Promise, the case of empirically based
moral uncertainty. In either case, the option of helping my friend will stochastically
dominate the option of going to the game iff my credence that I am objectively
morally obligated to help my friend is greater than or equal to .5. Thus, it seems,
stochastic dominance reasoning gives the deontologist something definite, precise,
and well-motivated to say both about choices made under morally relevant empirical
uncertainty and about choices made under pure moral uncertainty.14
An immediate worry: We have so far assumed that the background deontolog-
ical conception is absolutist. But many philosophers who identify as deontologists
and endorse characteristically deontological moral phenomena like agent-centered
constraints do not think of deontological constraints as absolute.15 Fortunately,
14Interestingly, the view that it is permissible to take a morally risky act only when the proba-
bility that the act is not objectively wrong is greater than .5 has a history in the Catholic moral
theology literature, under the name “probabiliorism” (Sepielli, 2010, pp. 51-2). Stochastic dom-
inance, in these terms, implies that a deontological absolutist must adopt a position at least as
rigorous as probabiliorism in the cases of “asymmetrical” moral risk with which this literature is
chiefly concerned.
15The literature on deontological moral uncertainty is divided in focus between absolute and
non-absolute versions of deontology. Jackson and Smith (2006) direct their criticisms at absolutist
theories. Aboodi et al. (2008) suggest that this is a mistake, since “hardly any (secular) contempo-
rary deontologist is an absolutist” (p. 261, n5). But Huemer (2010, p. 348, n3) marshals a credible
array of apparent (contemporary or near-contemporary) deontological absolutists. Note that the
difficulties of absolutism arise for anyone who takes one class of reasons to be lexically stronger than
another: assuming that there is always, in any choice situation, a non-zero risk that the lexically
stronger reasons are in play (e.g. that a given course of action might cause some innocent person
torturous suffering), one may worry that the lexically weaker reasons (e.g. headache prevention)
will always be preempted by that risk and hence never rise to the level of practical relevance.
But in any case, the challenge posed by uncertainty is no less acute for non-absolute deontologists.
The non-absolutist must still answer the question of how an agent should decide what to do when
290
however, the stochastic dominance argument does not turn on the assumption of
absolutism. Following Colyvan et al, who conclude that the absolutist version of
deontology is implausible, we may represent the rightness of fulfilling an obligation
not as +∞ but as merely a very large finite positive number, i.e., as a reason of finite
strength though much stronger than ordinary prudential reasons. And likewise, we
may represent the wrongness of violating an obligation not as −∞ but as merely a
very large finite negative number.16 Suppose that keeping a promise to my friend
has a value of +9001 while breaking that a promise has a value of −9001. It is
easy to see that the stochastic dominance argument for helping Petunia paint will
go through mutatis mutandis.
Note, however, that a reliance on stochastic dominance rather than expecta-
tional reasoning lets us avoid one of the chief pitfalls for absolutist approaches to
uncertainty, namely, the risk of “Pascalian paralysis”: If violating a moral obliga-
she is unsure whether a given course of action would violate a deontological constraint. As with
the absolutist, the most natural answer she can give is some version of the threshold view, but
to defend this view she will have to overcome all the same obstacles as the absolutist (Jackson
and Smith, 2016, pp. 287-8). The only new option that the non-absolutist position seems to
open up is an expectational view on which the subjective reason-giving force of a deontological
constraint for an uncertain agent is the product of the inherent stringency of that constraint
(which is finite, since the constraint is non-absolute) times the probability that a given course
of action would violate the constraint. But if she goes this expectational route, it is no longer
clear that her view should be counted as a form of deontology, rather than simply a version of
consequentialism that incorporates agent-centered constraints. (Cf. discussion in Portmore (2016).
Portmore allows that a deontologist may give an expectational account of subjective rightness
without becoming a consequentialist, so long as the underlying account of objective rightness
remains deontological. But he argues convincingly that the challenges of uncertainty arise not
just in cases where we are ignorant of facts (e.g., what the consequences of some action will be
or would have been) but also in cases where there is no fact of the matter about whether some
course of action would have constituted a constraint violation (e.g., due to physical indeterminism
or counterfactual underdetermination), and that therefore the deontologist who wishes to go the
expectational route must give a partially expectational account of objective as well as subjective
rightness, which he takes it would amount to a form of consequentialism. For a different line
of argument against expectational approaches to deontological moral uncertainty, see Tenenbaum
(2017).)
16See Colyvan et al (2010, pp. 515-8) for an axiomatic characterization of such a non-absolutist
deontological theory.
291
tion is treated as an outcome with the value of −∞, then actions that are almost
certainly morally permissible but carry even a vanishingly small risk of violating
a moral obligation will carry an expected value of −∞—the same expected value,
in fact, as acts that are certain to violate a moral obligation.17 Worse still, if (as
seems plausible to me) every possible action has some non-zero chance of fulfilling
a moral obligation and some non-zero chance of violating a moral obligation, then
the expected value of every possible action is undefined (∞+ (−∞)).
One way to avoid Pascalian difficulties is to hold that a rational agent may
never have non-zero credence in any outcome with infinite positive or negative value.
But this seems implausible and has only ad hoc motivation. The better response is
to modify our decision theory, either weakening or amending expected utility theory
in a way that allows an agent with modest credence in infinite values and disvalues
to nevertheless remain responsive to finitary considerations. Stochastic dominance
is one such weakening of expected utility theory: As far as stochastic dominance
principles are concerned, it is rationally permissible for me to go to the game, despite
the risk of infinite moral turpitude, so long as that risk has a probability less than
.5.18
The .5 credal threshold for permissibility follows from stochastic dominance
17In the literature on “pure” moral uncertainty, this is often described as the worry that “fanat-
ical” moral theories will hijack expectational reasoning—i.e., moral theories that attribute infinite
value and disvalue to options, perhaps in very strange or counterintuitive ways, will take prece-
dence over all finitary moral theories, so long as one assigns them even the most vanishingly small
degree of positive credence (Ross, 2006, pp. 765-7).
18Of course, stochastic dominance need not be understood as the only or the strongest principle
of rational requirement under uncertainty (cf. the last three paragraphs of this section). Various
principles are possible that occupy an intermediate position between stochastic dominance and
expected utility maximization and that might impose a more rigorous form of moral caution on
agents acting under uncertainty while still avoiding the problem of Pascalian paralysis.
292
reasoning when an agent is certain of all her non-moral reasons. But when she
is uncertain which option her non-moral reasons favor, stochastic dominance may
become more demanding. Suppose, for instance, that in the Possible Promise case
I am uncertain whether I would have a better time at the game or painting with
my friend. Perhaps I believe that there is a one-in-three chance that my team
will lose, and I know that while seeing my team win would give me a utility of
+20, seeing them lose would give me a utility of −10. If I simply ignore the game
and help my friend paint, on the other hand, I am guaranteed a utility of +5. In
this case, the degree of belief that I promised my afternoon to Petunia at which
stochastic dominance will require that I do so is reduced—specifically, to .4 rather
than .5.19 But in simple cases of deontological moral-prudential conflict, where I
am certain that one option O is morally permissible but prudentially worse than
another option P, and certain that P is prudentially better than O but uncertain
whether it is morally permissible, we may say that stochastic dominance requires me
to choose option O iff the probability that P is objectively impermissible is greater
than or equal to .5.20
19Suppose my credence that I promised to help Petunia paint is exactly .4. There are five possible
outcomes, with values of −∞, −10, +5, +20, and +∞. Options P and G have the same chance
of producing an outcome at least as good as −∞ (100%); P has a better chance of producing at
outcome at least as good as −10 (100% vs. 60%); P has a better chance of producing an outcome
at least as good as +5 (100% vs. 40%); P and G have the same chance of producing an outcome
at least as good as +20 (40% — the chance that G produces an outcome at least this good is a
product of the .6 chance that I do not violate an obligation in going to the game and the .6 chance
that my team wins the game); and P has a better chance than G of producing an outcome at least
as good as +∞ (40% vs. 0%) (Tables A.5-A.6). If, however, my credence that I made the promise
were any lower, then G would have a better chance than P of producing an outcome at least as
good as +20 and P would no longer stochastically dominate G.
20Cases of this sort have been a major focus of the recent literature on “pure” moral uncertainty,
in particular the cases of abortion and vegetarianism, both of which seem to present agents (at least
in many cases) with a conflict of prudential reasons on one side and uncertain moral reasons on
the other. See for instance Guerrero (2007), Moller (2011), and Weatherson (2014) for discussion
of these cases.
293
Prms (.4) ¬Prms & Win (.4) ¬Prms & Lose (.2)
Painting +∞ +5 +5
Game −∞ +20 −10
Table A.5: Possible Promise v3, payoff matrix
−∞ −10 +5 +20 +∞
Painting 1 1 1 .4 .4
Game 1 .6 .4 .4 0
Table A.6: Possible Promise v3, probability of payoff ≥ x
Of course, even allowing for the possibility of a higher threshold in cases of
prudential uncertainty, the requirements of stochastic dominance will not rise to the
level of practical stringency that we intuitively expect at least some deontological
constraints to possess. The kind of reasoning I have described makes no distinction
between “weaker” constraints (like the constraint against telling white lies) and
“stronger” constraints (like the constraint against killing the innocent).21 And while
a threshold of .5 may seem plausible for the weaker constraints, it seems much
less plausible when we think about the stronger constraints—surely a deontologist
should not conclude, for instance, that a judge need not worry about the constraint
against punishing the innocent so long as the probability that her sentencing decision
violates that constraint is a mere .49 rather than .51.
Importantly, however, stochastic dominance is only a sufficient condition for
rational requirement, not a necessary condition. Everything I have said so far leaves
it open to the deontologist to argue for a higher threshold, at least for some cate-
21Except, on a non-absolutist view, when there is some chance that the constraint in question
is outweighed by other kinds of considerations—since the chance of a stronger constraint being
outweighed will presumably be smaller, ceteris paribus, the threshold of moral safety at which the
option that risks violating that constraint ceases to be stochastically dominated will be higher.
294
gories of constraint, on grounds other than stochastic dominance. And even if one
takes the laxity of the stochastic dominance threshold to be intuitively unaccept-
able with respect to any sort of deontological constraint, progress has still been
made insofar as the deontologist can now say with confidence that you must at least
believe it to be more likely than not that your morally risky action violates no con-
straint, even a relatively weak constraint, before non-deontological (e.g. prudential
or consequentialist) reasoning may take over.
Nevertheless, it seems to me that this objection from the variable stringency
of constraints represents a residual difficulty that it will be hard for deontologists
to fully overcome. It is intuitive to hold that the practical force of deontological
constraints must be sensitive to both (i) the probability of a constraint being violated
by a given course of action and also (ii) the seriousness, importance, or stringency
of the particular constraint in question. But taken together, these intuitions push us
in the direction of an expectational view that, while it may still incorporate agent-
centered normative considerations, has lost the rest of its distinctively deontological
character (see note 15 supra). The deontologist may resist this pressure, but to do
so she must be prepared at some point to deny at least one of the above intuitions.
A.5 Option Individuation and Ought Agglomeration
So far I have argued that stochastic dominance reasoning offers a foundation
for a threshold view of deontological moral choice under both empirically based and
pure moral uncertainty. But such threshold views face another important difficulty,
295
which we have yet to confront.
Jackson and Smith (2006) argue that threshold views violate the principle of
“ought” agglomeration, that ought(A) & ought(B) ` ought(A & B), in cases where
an agent is faced with the possibility of performing two acts, each of which individ-
ually is below the threshold for permissible moral risk, but which in combination
exceed that threshold. They illustrate this difficulty by way of the following thought
experiment.
Two Skiers Two skiers are headed down a mountain slope, along different paths.
Each, if allowed to continue, will trigger an avalanche that will kill ten innocent
people. (If both are allowed to continue, they will trigger two such avalanches
and kill both groups of ten.) The only way to save each group is to shoot
the corresponding skier dead with your sniper rifle. You can shoot either
skier individually or, being an extraordinarily crack shot, you can shoot both
with a single bullet. The moral theory you accept (with certainty) tells you
that you ought to kill culpable aggressors in other-defense, but are absolutely
prohibited from killing innocent threats. Unfortunately, you are uncertain
of the intentions of the skiers, assigning them each equal and independent
probabilities of acting obliviously, and thus as innocent threats, rather than
as ill-intentioned aggressors.
Suppose you subscribe to a threshold view on which you ought to kill a po-
tential aggressor in other-defense iff your credence that he is acting innocently is
less than t. And suppose that, with respect to each of the two skiers, your credence
296
that he is innocent is just less than t. Thus, it seems that you ought to shoot Skier
1 and you ought to shoot Skier 2. But, if you shoot both Skier 1 and Skier 2, the
chance that one of them is innocent and hence that you will have violated a deonto-
logical constraint is greater than t and hence, it seems, you ought not perform the
compound action: shoot Skier 1 and shoot Skier 2.
It is immediately tempting to point out that shooting Skier 1 and shooting
Skier 2 are two separate actions, and that all that should matter about the case
from a moral standpoint is whether the chance of either action individually violating
a deontological constraint is greater than t. And I will shortly argue that something
very much like this is the right response for the deontologist to give. But Jackson and
Smith seem to have headed off this response: While shooting each skier separately
looks like two separate actions, shooting both skiers with the same bullet, as they
have imagined you have the option of doing, is a single action, with a chance greater
than t of violating a deontological constraint. But surely it is implausible that you
are permitted (indeed, required) to shoot both skiers with separate bullets, but
prohibited from shooting them both with the same bullet.
A plausible solution to the agglomeration problem must therefore find some
appropriate criterion for “individuating” the shooting of Skier 1 and the shooting
of Skier 2 that allows us to distinguish the two sources of moral risk, even given
the option of shooting both skiers with a single bullet. Aboodi, Borer, and Enoch
(2008) offer one such proposal. They suggest that, if all deontological constraints
can be thought of as rights possessed by particular rightsholders, a deontological
agent acting under uncertainty ought to ensure that she does not run a risk greater
297
than the threshold t of violating the risks of any particular rightsholder, though she
may permissibly run a risk greater than t of violating the rights of some rightsh-
older. They suggest this might be justified, inter alia, by the broadly contractualist
thought that what determines the permissibility of my conduct is whether anyone
can reasonably or legitimately object to it: A rightsholder cannot reasonably object
to my conduct merely because it creates some risk that her rights will be violated,
nor because it creates a large aggregate risk that someone’s rights will be violated.
But she can reasonably object if I run an excessive risk of violating her rights in
particular. Thus, Aboodi et al conclude, in the Two Skiers case you ought to shoot
both skiers, because in so doing there is no rightsholder that you run a risk greater
than t of wronging. (Skier 1 and Skier 2 are each rightsholders, but there is no
compound entity, Skier 1 + Skier 2, that is a rightsholder or can lodge objections
against your conduct.)
This patient-based approach, however, has seriously counterintuitive conse-
quences, as Huemer (2010) points out. Suppose, as I have argued holds true in the
simplest sort of case, that the threshold for permissible risk of violating a deonto-
logical constraint is .5, and consider the following case (a slightly simplified retelling
of Huemer’s “War Options”).
The Weapon Minerva, a military officer, faces a situation in which she knows that, if
she does not act, a powerful weapon will fall into the hands of a ruthless enemy,
who will use it to kill 100,000 innocent people. Fortunately, there are two
surefire ways to prevent the weapon from falling into enemy hands: (1) Minerva
298
has received intelligence indicating that a certain scientist will shortly reveal
the plans for constructing the weapon to the enemy. It is unclear whether
the scientist is willingly helping the enemy or acting under exculpating duress
(say, threats on the life of his family). Minerva thinks it slightly more likely
than not (.55) that he is acting under duress and that it would therefore be
objectively wrong to kill him even to save a greater number. (2) In addition to
the weapon plans, the enemy needs certain materials to construct the weapon,
which are being kept in a small town of 25,000 people. Minerva knows that
if she orders the carpet bombing of this town, the weapon materials will be
destroyed, but 45% of the town’s population will be killed. Since she has no
specific information about which members of the town’s population will be at
risk, the epistemic probability that any individual townsperson will be killed
is .45.
On Aboodi et al’s patient-centered approach, it looks as though Minerva is permitted
to carpet bomb the town (since there is no rightsholder whom she runs a risk greater
than .5 of violating) but is not permitted to assassinate the scientist (since in doing
so she would run a risk greater than .5 of wronging him).22
Fortunately, there is another way of resolving the agglomeration problem that
both avoids these counterintuitive consequences and follows naturally from a central
feature of the deontological approach to ethics: Because deontological moral assess-
ment of options is relativized to the particular choice situations in which those
22Of course, the particular value of the threshold plays no role in this argument—the case can
be easily modified to accommodate any threshold greater than or less than .5.
299
options arise, it seems to me, deontological moral theories simply cannot and do not
assess compound options, i.e., combinations of options that arise in different choice
situations. Thus, while shoot Skier 1 and shoot Skier 2 are each individually sub-
ject to deontological moral assessment, the compound option shoot Skier 1 & shoot
Skier 2 is not. But since it is a theory’s moral assessment of options that provides
the basis for stochastic dominance reasoning, which in turn grounds the threshold
principle for deontological moral risk, this means that the threshold principle is sim-
ply inapplicable to compound options, and hence does not generate a prohibition
against shooting both skiers. Rather, we are left free to make the natural inference
from the fact that you ought to shoot Skier 1 and ought to shoot Skier 2 to the
conclusion that you ought to shoot both skiers, preserving “ought” agglomeration.
Let’s spell out this line of reasoning in a bit more detail. Deontological moral
assessment is relativized to choice situations in the following sense: Just as deonto-
logical theories (in contrast, for instance, to agent-neutral consequentialist theories
like classical utilitarianism) direct me to distinguish between the rights and wrongs
of my own actions and those of others, and hence, for instance, not to commit one
rights violation to prevent someone else from committing five, so likewise they direct
me to distinguish between the rights and wrongs of the choice immediately before
me and the rights and wrongs of my other choices, past, present, or future. Thus
for instance a deontological theory does not direct me to lie now even if it is the
only way of extricating myself from a situation in which I can predict with certainty
that I will tell five lies. To put this in terms of the quantitative representations
we employed in the last section, a deontological theory will assign the option of ly-
300
ing now the large negative value (either finite or infinite) associated with violating a
constraint, but will not assign this same degree of disvalue to the option of not lying,
even though this option carries with it the certainty of future constraint-violating
lies, which will themselves be assigned the disvalue of constraint violations in the
choice situations where they arise.23
Colyvan et al. (2010) put this point by suggesting that deontological moral
duties are both agent-relative and time-relative (p. 513), but this is not quite right,
since one can make multiple choices at the same time, and in such cases the rela-
tivization of deontological considerations remains to the individual choice situation,
rather than to the time at which both choices are made.24 For instance, consider
the following case:
The Buttons of Wrongness A mad ethicist has rigged a contraption that will test
your judgments about two ethical dilemmas at once. You may press one of
four buttons, colored red, green, yellow, and blue. If you press the red button,
a message will be sent from your phone to one of your friends, telling him a
white lie about his recent haircut. If you press the green button, $1000 will be
stolen from the accounts of a large corporation and donated to GiveDirectly.
If you press the yellow button, both these things will happen. If you press the
blue button, neither will happen. Pressing any of the buttons will deactivate
the rest, and you will have no other opportunity to reverse the effects of your
23A deontological theory should likely assign some disvalue to options that I know will lead me
to future constraint violations, but the point is that it does not assign the same kind or degree of
disvalue that it assigns to options that directly violate a constraint.
24In fairness, Colyvan et al may not have meant to assert otherwise, since later on the same page
they put the point in terms of choice situations rather than times.
301
selection once it has been made.
There is a sense, clearly, in which you make only one choice in this scenario,
namely which button to press. But in a more ethically relevant sense, you make
two choices, namely whether to tell a white lie and whether to steal money from the
large corporation for GiveDirectly—even though you will put both of these choices
into effect by means of a single action, namely, pressing a button. And to the
deontologist, what you will do with respect to one of these choices simply has no
bearing on what you are permitted to do with respect to the other—not because they
are potential violations of two different rightsholders (it would make no difference
if your friend with the new haircut was also the majority shareholder of the large
corporation) but simply because they are two different choices.25
What does this mean for Jackson and Smith’s Two Skiers case and the ag-
glomeration objection to threshold views? Simply this: Because deontological moral
assessment is choice situation relative, deontological moral theories are just not in
the business of assessing compound options like shoot Skier 1 & shoot Skier 2. Just
like the options of telling your friend a white lie and stealing for GiveDirectly in the
25This is not to say that deontologists are committed to metaethical possibilism, the view that
what I ought to do with respect to a particular choice situation can never depend on what I believe
I will do with respect to other choice situations. For instance, the deontologist might still allow
that in the classic test case for actualism vs. possibilism (Jackson and Pargetter, 1986, p. 235),
Professor Procrastinate should decline the invitation to review a recently published book, even
though he knows himself to be the person best qualified to review it, because he also knows that
given his inveterate tendency to procrastinate, he will not finish the review on time. The point is
rather the narrower one that the strength of a deontological moral prohibition or requirement is not
altered by one’s expectations about how one will behave in other choice situations. (Thus, what
allows the deontologist to take the actualist line with respect to Professor Procrastinate is that
reviewing the book is, one assumes, morally optional. If it were the case, for whatever reason, that
Professor Procrastinate were morally obligated to agree to review the book, then the deontologist
should agree with the possibilist that he ought so to agree, whatever his expectations about any
of his other choices or actions.)
302
Buttons of Wrongness case, these options inhabit different choice situations, even
if they are (or can be) effectuated by means of a single action (viz., shooting both
skiers with the same bullet). For this reason, we cannot say for instance that the
option of shooting neither skier stochastically dominates the option of shooting both
skiers: Stochastic dominance reasoning makes reference to a probabilistic assignment
of degrees of objective value or choiceworthiness to options, but since deontological
theories don’t assign degrees of value or choiceworthiness to compound options, they
can be neither stochastically dominant nor dominated. Since the threshold principle
of choice under deontological moral uncertainty is grounded in stochastic dominance
reasoning, this principle in turn is only applicable to simple options. Thus in the
Two Skiers case, the threshold principle does not imply that you ought not shoot
both skiers. Rather, we can and should reason, by the principle of agglomeration,
from the fact that you ought to shoot Skier 1 and ought to shoot Skier 2 to the
conclusion that you ought to shoot both skiers.
A.6 Conclusion
Moral uncertainty, of both the empirically based and purely moral varieties,
presents a challenge for deontologists. They must give an account of how agents
should deliberate and act in the face of such uncertainties, ideally an account that
is as precise and intuitively well-motivated as the expectation-maximizing account
available to consequentialists. I have suggested that a version of the threshold
account grounded in stochastic dominance reasoning may meet this need. If we
303
treat morally right/obligatory actions as having very large positive moral value
(whether finite or infinite) and morally wrong/prohibited actions as having very
large negative moral value (“”), then it will often turn out that morally risky options
are stochastically dominated. In the simplest sort of case, where only one of two
options carries moral risk, and that option is certain to be preferable but for the
risk of violating a moral requirement, it turns out that that option is stochastically
dominated if and only if the probability that it is objectively prohibited is greater
than or equal to .5, though this threshold may be less than .5 in more complex cases.
The principle that one rationally ought not choose stochastically dominated options
is extremely compelling, representing a weakening of expectational reasoning that
both permits a variety of risk attitudes (risk aversion and risk seeking as well as
risk neutrality) and avoids worries about moral fanaticism or Pascalian paralysis
arising from tiny probabilities of infinite value and disvalue. It therefore provides
an appealing basis for the threshold approach to deontological moral uncertainty,
one that moreover (when combined with the right understanding of deontological
moral evaluation) can avoid the most powerful objection to that approach. It seems,
then, that deontologists are better equipped to respond to the challenges of moral
uncertainty than many philosophers have thought.
304
Appendix B: What Grounds Intertheoretic Comparisons of Moral
Value
The purpose of this appendix is to explain in greater detail what grounds
the correctness of content-based value aggregation methods of the sort described in
Chapters 5-6. More specifically, I aim to show (i) that it is rational for agents in
certain doxastic states to aggregate the value assignments of the moral theories in
which they have positive credence in particular ways, based on the content of the
theories being aggregated, and (ii) that the propositions belief in which makes it
rational for agents to aggregate in these ways are plausibly also true, making the
content-based aggregation procedures they justify not only rational but correct.
The account I will give relies heavily on the Bayesian commitments taken on
in Chapter 1. Specifically, I will assume that an ideally rational agent has cre-
dences, satisfying the probability calculus, defined over all propositions that can be
constructed from her conceptual resources, and that those conceptual resources are
not unduly impoverished. This allows us to describe such an agent as distributing
belief over complete theories of the world, understood as maximal consistent sets of
propositions (“maximal” relative to her conceptual resources), with her credence in
all such theories summing to 1. For reasons that will become apparent, it will be
305
useful to think of morally uncertain agents as dividing their beliefs not simply among
moral theories (i.e., maximal consistent sets of moral propositions) but among the-
ories of the world simpliciter, containing empirical propositions, moral propositions,
logical/mathematical propositions, higher-order normative propositions about the
requirements of rationality, etc. As I have done through most of the dissertation, I
set aside questions of non-ideal or bounded rationality (except briefly in §B.5), and
content myself with arguing that ideally rational agents would respond in particular
ways to particular states of moral uncertainty.
For simplicity, I will take as my central example the Easy Case from Chapter 5,
in which Alice divides her beliefs between hedonistic utilitarianism and a pluralistic
theory that values both hedonic and aesthetic goods. We will introduce other cases
as necessary to illustrate potential complications absent from the Easy Case.
B.1 The Easy Case: Decision-Theoretic Compatibility
Let’s begin by giving a more thorough description of the Easy Case. Expanding
on and precisifying the description of the case from Chapter 5, we may say that Alice
has beliefs, inter alia, about the following propositions.
P1 The choiceworthiness of an option is equal to the value of the world that results
from it, where the value of a world is the cardinal sum of non-derivative value
that supervenes on value-bearing empirical phenomena in that world.1
P2 Hedonic experiences (i.e., experiences with positive or negative hedonic valence)
1“Value” being used in this context in a way not synonymous with “choiceworthiness,” since
only practical options and not worlds are direct objects of choice.
306
are non-derivative value-bearers.
P3 The explanation/grounding for the non-derivative value of hedonic experience
is ϕ.
P4 Aesthetic goods (beauty and the like) are non-derivative value-bearers.
P5 The explanation/grounding for the non-derivative value of aesthetic goods is ψ.
P6 The value of one hedon is equal to the value of one aestheton.
P7 Nothing apart from hedonic experiences and aesthetic goods is a non-derivative
value-bearer.
In the simplest version of the case, Alice then divides her beliefs between two
theories.
T1 (hedonism): {P1, P2, P3,¬P4,¬P5,¬P6, P7, ...}
T2 (pluralism): {P1, P2, P3, P4, P5, P6, P7, ...}
When we introduced the Easy Case, we also stipulated that T1 and T2 share
an expectational decision procedure. This fact might be represented in at least two
importantly distinct ways. First, it might be that T1 includes a proposition like
P8 Every agent is always rationally required (subjectively ought) to maximize ex-
pected hedonic value.
while T2 includes a proposition like
307
P9 Every agents is always rationally required (subjectively ought) to maximize
expected hedonic + aesthetic value.
But either of these propositions imply that what an agent rationally ought to
do is entirely insensitive to her moral beliefs, which I have argued is implausible
(Chapters 2-3).
More plausible versions of T1 and T2 might include the following propositions
respectively.
P10 Every agent who has perfectly true moral beliefs (i.e., assigns every true basic
moral proposition credence 1 and every false basic moral proposition credence
0) is always rationally required to maximize expected hedonic value.
P11 Every agent who has perfectly true moral beliefs (i.e., assigns every true basic
moral proposition credence 1 and every false basic moral proposition credence
0) is always rationally required to maximize expected hedonic + aesthetic
value.
Let’s refer to the versions of T1 and T2 that include P10 and P11 as T1a and T2a
respectively.
Neither P10 nor P11 say anything about how an agent who lacks perfectly
true moral beliefs is rationally required to act. So a theory that includes either
of these propositions could consistently include any number of further propositions
that speak to this question. That is, a theory that includes P10 or P11 might endorse
irrelevantism, MFT, MFO, or hedging, inter alia.
308
When I introduced the hedonism/pluralism case in Chapter 5, I claimed in
effect that it would be most natural for an agent who divides her moral beliefs
between T1a and T2a to accept a generalization of these principles like
P12 Every agent is always rationally required (subjectively ought) to maximize
expected choiceworthiness.
and more generally, that it would be most natural for an agent who has positive
credence in T1a and T2a to accept a principle like P12 conditional on the disjunction
T1a ∨ T2a.
If we wished, we could simplify our task by simply stipulating that Alice
divides her belief between versions of T1 and T2 that both include P12. But in any
case, for T1 and T2 to be maximal consistent sets of propositions, they must include
some proposition or propositions specifying the rational requirements that apply to
morally uncertain agents. And this illustrates an important and often-overlooked
point. Several critics of hedging have put forward some variant of the claim that
“It’s not the business of a moral theory to tell you what to do if you don’t believe
that theory” (e.g. (Hudson, 1989, p. 224), (Gracely, 1996, p. 331)). But this neglects
the fact that a moral theory is part of a complete theory of the world (though, for
a given agent, the same moral theory may figure in many complete theories of the
world, given her uncertainties about non-moral matters), and that components of an
agent’s moral theory may derive from principles that have implications outside the
domain of first-order morality. For instance, one way of being a classical utilitarian
is to accept the following two propositions.
309
P13 The choiceworthiness of an action is wholly determined by the net hedonic
utility of its consequences.
P12 Every agent is always rationally required (subjectively ought) to maximize
expected choiceworthiness.
Considered as the logical closure of these two propositions, this version of utilitari-
anism does tell an agent what to do when she is morally uncertain.2
Finally: If it is true, as I argued in Chapter 7, that a principle like P12 has exter-
nal, belief-independent normative force, then it is simply irrelevant to the question
of what Alice is rationally required to do whether the theories over which she dis-
tributes her belief contain P12 or other, conflicting principles of rational requirement
(like P8 or P9). She will be rationally required to maximize expected choiceworthi-
ness regardless—at least, so long as it is possible for her to make the comparisons
necessary to form an expectation.
I will therefore take it for granted that, insofar as it is possible for her to do
so, Alice ought to aggregate the reasons proffered by T1 and T2 by computing the
expected value of her options given the disjunction of the two theories. To summarize
the preceding paragraphs, there are three possible routes to this conclusion: Given
a normative internalist view on which the appropriate aggregation procedure for
a class of theories is the strongest procedure on which those theories can agree,
2One might quibble, of course, over whether P12 is really a component of an agent’s moral
theory, and claim that moral theories must contain only propositions like P10 that are conditioned
on an agent’s having perfectly true moral beliefs. But I cannot see why this sort of semantic
division between moral theories and the more general theories of the world in which they are
embedded should make any difference for our purposes, i.e., with respect to the requirements of
rationality for an agent in a particular epistemic/doxastic state.
310
we might (i) simply stipulate that the versions of T1 and T2 in which Alice has
credence both include P12 (the principle that Alice ought to maximize expected
choiceworthiness), or we might (ii) stipulate that the versions of T1 and T2 in which
Alice has credence include P10 and P11 respectively, and then argue that an agent
who accepts P10 conditional on the other propositions we have stipulated as elements
of T1 and accepts P11 conditional on the other propositions we have stipulated as
elements of T2 ought to accept P12, as a natural generalization of P10 and P11,
conditional on T1 ∨ T2. Finally, we might (iii) take the externalist line defended
in Chapter 7 and hold that the rational requirement stated by P12 is incumbent
on agents regardless of their beliefs (about P12 or anything else). To simplify the
following discussion, however, I will take option (i) and simply stipulate that Alice’s
beliefs are such that P12 ∈ T1 and P12 ∈ T2.
The principle of maximizing expected choiceworthiness stated by P12, how-
ever, is non-vacuous only so long as the expected value of options is defined. The
crucial question, therefore, is whether T1 and T2 meet the preconditions for defining
expectations over them, namely structural compatibility and scale normalizability.
Let’s now turn to this question.
311
B.2 The Easy Case: Structural Compatibility and Scale Normaliz-
ability
In Alice’s case, structural compatibility is trivial. Since P1 ∈ T1, T2, both
theories have one-dimensional cardinal structure.3 The real issue, then, is scale
normalizability. The crucial claim in Chapter 5 that allowed the Easy Case case to
go forward as an example of comparability was that the value of a hedon according
to T1 is equal to the value of a hedon according to T2. So the task now is to explain
this claim—partly to defend it, partly to generalize our conclusions to cases where
it or its analogues fail.
My strategy for defending scale normalizability has two parts. First, I will
claim that, conditional on T1 ∨ T2, Alice ought to accept
P14 The degree of value borne by a given unit of hedonic experience is independent
of whether aesthetic goods are non-derivative value-bearers.
But second, I will argue that, even if Alice does not accept P14, she ought to accept
some other proposition (or distribute her belief over some set of propositions) that
establish a different normalization between T1 and T2, which still allows her to
compute and maximize expected value.
To make the first claim more precise, I claim that P14 is very plausible given
{P1, P2, P3, P7} (i.e., given that hedons have non-derivative value, a complete and
3More specifically P1 seems to guarantee ratio structure, since there is a meaningful zero point
corresponding to any world that contains no value-bearers. But since an expectational decision
procedure requires only interval structure, the issue of interval vs. ratio structure is unimportant.
312
fixed story about what grounds or explains that value, and that nothing besides
hedons and aesthetons has non-derivative value), and therefore that, ceteris paribus,
an agent who accepts {P1, P2, P3, P7} ought to accept P14. The ceteris paribus clause
means that other features of an agent’s epistemic/doxastic state, which we have so
far left unspecified, might make it rational for her to reject P14, but that in the
default or normal case (leaving this intentionally vague), it will be rational to accept
it. This is to say, then, that in the default description of the Easy Case, P14 ∈ T1, T2
(just as I suggested in the last section that, in the default description of the case
where Alice accepts T1a and T2a, P12 ∈ T1a, T2a).4
Why is this so? One reason is P3, or more precisely, the fact that P3 ∈ T1, T2.
It is a stipulation of the Easy Case that T1 and T2 are in full agreement, not just
that hedonic experience has non-derivative value, but on everything to do with
hedonic value, and in particular with respect to the underlying moral metaphysics
that explains or grounds the value of hedonic experience. Given that, at least as far
as anything we have said up to this point is concerned, T1 and T2 do not disagree
about anything that has any direct connection to the issue of hedonic value, it is at
least prima facie plausible that they assign the same degree of value to a given unit
4Of course, given the rational requirement of regularity, Alice’s credence Cr(P14|{P1, P2, P3, P7})
cannot be 1—if rational agents ought to be at least a little uncertain about everything, then Alice
ought to be at least a little uncertain whether the value of a hedon is independent of the value
of aesthetons, even conditional on (the remaining propositions of) T1 and T2. Nevertheless, I
will set this detail aside for the moment, and suppose for simplicity that, if P14 coheres with the
remaining propositions of T1 and T2 better than any alternative, then Alice may simply distribute
the entirety of her belief over versions of T1 and T2 both of which include P14. Of course, the idea
that a rational agent could distribute her belief over just two complete theories of the world is
entirely inconsistent with the assumption of regularity, or indeed with much weaker assumptions
about the limits of rational certainty. But it is a convenient simplification for the moment. We
will later examine what happens when this assumption is relaxed, and consider a version of the
Easy Case in which Alice is uncertain not just about propositions like P4 and P5 (i.e., about the
value of aesthetons) but also about propositions like P14.
313
of hedonic experience.
To plausibilify this point, consider an analogous case of empirical uncertainty:
Suppose Bob is certain that hedonic, maximizing, expectational utilitarianism is
true, and indeed is certain of everything except for the empirical question of what
will happen if he presses the red button. More precisely, he divides the entirety
of his belief between two complete theories of the world: T3, according to which
utilitarianism is true and pressing the red button will save five innocent people,
and T4, according to which utilitarianism is true and pressing the red button will
kill ten innocent people. Call these two competing propositions about the effects
of pressing the red button Rs and Rk. In order to compute the expected value of
pressing the red button, Bob must make some assumption about how the value of
a hedon conditional on Rs compares with the value of a hedon conditional on Rk.
And of course, unless for some quite unusual reason his beliefs about the value of
hedonic experience are linked to his beliefs about the consequences of pressing the
red button, he will assume that the value of a hedon is independent of whether Rs
or Rk is true—that is, he will accept the analogue of P14.
Just as the value of a hedon is, prima facie or ceteris paribus, independent of
what would happen if Bob presses the red button, so it strikes me as natural to think
that the value of a hedon is, prima facie or ceteris paribus, independent of whether
aesthetons are value-bearers as well. Though questions about the value of hedons
and the value of aesthetons “belong to the same domain,” in a way that questions
about the value of hedons and the consequences of pressing the red button do not,
in both cases the questions are logically orthogonal. There is no obvious reason
314
why settling the question of whether aesthetons have non-derivative value should
affect Alice’s beliefs about the value of hedonic experience one way or the other.
The burden is on the opponent of intertheoretic comparability, then, to explain why
Alice should not accept P14—for if she accepts P14, and is rationally required to
maximize expected choiceworthiness (either in virtue of accepting P12 or in virtue of
P12 being an external rational requirement), then she will be rationally required to
make intertheoretic comparisons and engage in moral hedging (since accepting P14
enables her to compute the expected choiceworthiness of options notwithstanding
her uncertainty between T1 and T2). (This argument will be set out in greater detail
in the next section.)
Note that, since we have assumed that T1 and T2 are maximal consistent sets
of propositions, they cannot simply be silent on the question of how the value of a
hedon given P4 (that aesthetons have non-derivative value) compares with the value
of a hedon given ¬P4. Rather than P14, these theories might contain propositions
like the following.
P15 The value of a hedon if P4 is true is exactly half the value of a hedon if P4 is
false.
P16 The value of a hedon if P4 is true is determinately at least half, and no more
than twice, the value of a hedon if P4 is false, but indeterminate with respect
to any intermediate ratio.
P17 The value of a hedon if P4 is true is incomparable with the value of a hedon if
P4 is false.
315
A set of propositions that contained no proposition of this sort would not be max-
imal, since any of these propositions could be added to it while preserving consis-
tency.5
A proposition like P15, of course, would establish comparability just as much as
P14—it just normalizes the value scales of T1 and T2 differently.
6 A proposition like
P16 would establish rough comparability. Opponents of any kind of intertheoretic
comparability, on the other hand, must hold that Alice should accept P17. And
to do this, they must show a disanalogy with the case of Bob, who should clearly
accept the analogue of P14 and not the analogue of P17.
The only argument I am aware of that might support such a conclusion is
the following: A moral theory’s value scale is simply a representation of its ordi-
nal preferences over options in conditions of empirical uncertainty—derivable from
those preferences so long as they satisfy certain plausible constraints (e.g. the von
Neumann-Morgenstern axioms), as shown by the representation theorems of stan-
dard expected utility theory (von Neumann and Morgenstern, 1947). But these
representations are unique only up to positive affine transformation—that is, given
5One might deny this claim by denying that P14−17 are in fact coherent propositions, i.e., that
it is meaningful to talk about how the value of a hedon conditional on P compares with the value
of a hedon conditional on ¬P . But the case of Bob shows that this is false: If such propositions
were generally incoherent, then expectational reasoning would be impossible. One might also claim
that such propositions are incoherent only when the P on which we are conditioning the value of
a hedon is a necessary rather than a contingent proposition. I will address this objection in §B.5.
6Rebecca Stangl has suggested to me in conversation that, on certain stories about the nature
of value, discovering new sources of value in the world should lead us to regard other sources of
value as less valuable than we previously thought they were. For instance, if the explanation for
the value of hedonic experience (i.e., what fills in the variable ϕ in P3) is that it contributes to
living a good life, then learning that aesthetic experience also contributes to living a good life
might lead Alice to conclude that hedonic experience is less important than she would otherwise
have thought, since it is only one contributor to the good life rather than the sole contributor.
In this case, Alice should accept something like P15 rather than P14. But this is no obstacle to
my thesis: T1 and T2 are still scale-normalizable, only with a different normalization (that gives
comparatively more weight to T1 and less to T2).
316
any value assignment that represents a theory’s preferences over options, any trans-
formation of that value assignment of the form mx + b, for (positive) real-valued
constant m and (positive or negative) real-valued constant b, will result in another
value scale that represents the theory’s preferences equally well. So, the claim goes,
just as interpersonal utility comparisons are impossible in standard expected utility
theory, where a person’s utility function is just a representation of her preferences
over risky prospects, intertheoretic value comparisons are likewise impossible—and
just as the value of a hedon for person A is incomparable with the value of a hedon
for person B, on standard expected utility theory, so the value of a hedon according
to T1 must be incomparable with the value of a hedon according to T2.
7
This argument strikes me as extraordinarily unconvincing. To begin with, it
is simply wrong about the structure of many moral theories. Classical, hedonistic
utilitarianism, for instance, does not derive its value scale from preferences that it
enjoins on agents under empirical uncertainty. Rather, classical utilitarianism starts
from a claim like P1, that states of affairs have value in proportion to the value they
contain (more specifically, the balance of pleasure minus pain), and derives from
that fact the conclusion that agents in particular epistemic states ought to prefer
some options to others. That is to say, for theories like classical utilitarianism,
the axiological facts ground the normative facts, not the other way around. The
representation theorems therefore play no role in these theories.8
7This argument is suggested by Ittay Nissan-Rozen, who writes that “I cannot think of any
plausible alternative to the ‘representation of moral rankings’ explication of the term ‘moral value’,
and under this explication intertheoretic comparisons of moral value are...meaningless.” (Nissan-
Rozen, 2015, p. 18)
8This strikes me as the correct way of understanding the utilitarianism of Bentham and Mill,
but of course exegetical claims are beside the point. The point is that such a moral theory is
317
Moreover, the analogy to the problem of interpersonal utility comparisons
points out the simple implausibility of this argument. In the interpersonal context,
the implausible conclusion of standard expected utility theory that A’s slow and
agonizing death is not a greater harm to A than B ’s mild itch is to B shows us that
individual utility or wellbeing is not simply a representation of preferences over
risky prospects. Likewise in the intertheoretic context, the implausible conclusion
that, for instance, the value difference between universal/perpetual happiness and
universal/perpetual agony according to T1 is not greater than the value difference
between the actual world and the actual world plus one mild itch according to T2
might be taken to show us that moral value is not simply a representation of a moral
theory’s ranking of risky options.
But finally, Nissan-Rozen’s argument is simply question-begging as an argu-
ment against intertheoretic comparability or moral hedging. For he takes for granted
that an agent may have rationally justified credence in a subjective first-order moral
theory that ranks options under empirical uncertainty in a way that allows the con-
struction of a value scale for that theory via the representation theorems, while
ignoring the parallel possibility that an agent may have rationally justified credence
in a second-order theory that ranks options under moral uncertainty in a way that
allows the construction of an intertheoretic value scale. That is, if Alice can reason-
ably arrive at the ranking of options relative to her state of moral uncertainty that
the intuitive normalization of T1 and T2 would suggest, then P14 will simply emerge
from that ranking via the representation theorems, just as its analogue will emerge
possible, and furnishes a counterexample to Nissan-Rozen’s argument.
318
from Bob’s ranking of options in the red button case.
I conclude the following, then: (i) It seems prima facie rational for Alice,
given what else we have said about her epistemic state, to accept P14 (i.e., to divide
her belief between versions of T1 and T2 each of which includes P14). (ii) The most
likely scenario in which this is not the case is one in which Alice has reason to accept
propositions like P15 or P16, which still establish intertheoretic comparability and
the basis for hedging, rather than P17, which does not.
B.3 A Simple Argument for Hedging
We have now shown that Alice has the ability to make rationally well-grounded
intertheoretic comparisons. We can now suppose, as per the arguments of the pre-
ceding sections, that Alice divides her belief exclusively between final versions of T1
and T2,
T1f (hedonism): {P1, P2, P3,¬P4,¬P5,¬P6, P7, P12, P14, ...}
T2f (pluralism): {P1, P2, P3, P4, P5, P6, P7, P12, P14, ...}
such that her credence Cr(T1f ) = p and Cr(T2f ) = 1− p.
Since P14 ∈ T1f , T2f , both theories affirm that the value of a hedon according
to T1f is equal to the value of a hedon according to T2f . And since all it is for the
value of a hedon according to the two theories to be equal is for the theories to affirm
that equality, we may simply say that the value of a hedon according to T1f is equal
to the value of a hedon according to T2f . Further, since P6 ∈ T1f , the value of an
319
aestheton according to T2f is equal to the value of a hedon according to T2f . So by
transitivity of equality, the value of a hedon according to T1f is equal to the value
of an aestheton according to T2f .
The following argument, then, appears to be valid.
1. Alice has credences Cr(T1f ) = p and Cr(T2f ) = 1− p.
C. The expected choiceworthiness of an option O, for Alice, is equal to hedons(O)+
(1− p)(aesthetons(O)).
This establishes, a fortiori, the possibility of intertheoretic comparison and aggre-
gation for Alice.
The following argument is valid as well.
1. Alice is rationally required to maximize expected choiceworthiness.
2. The expected choiceworthiness of an option O, for Alice, is equal to hedons(O)+
(1− p)(aesthetons(O)).
C. Alice is rationally required to choose options that maximize hedons(O) + (1−
p)(aesthetons(O)).
This establishes, it seems to me, that the problem of intertheoretic value comparisons
does not constitute a general objection to moral hedging: If it is possible for an agent
to be under a rational requirement to maximize expected choiceworthiness (either in
virtue of accepting a normative principle to that effect, or in virtue of such a principle
320
having external, belief-independent rational requiring force), then in at least some
states of moral uncertainty, those expectations will be well-defined and she will
be rationally required to hedge. Someone who takes PIVA to constitute a general
objection to hedging, therefore, must first claim that the expected choiceworthiness
of options is sometimes undefined (i.e., there are some states of uncertainty that
preclude well-defined expectations) and then argue on this basis that agents cannot
ever be rationally required to maximize expected choiceworthiness. But it is not
clear how such an argument might go, and to my knowledge no such argument has
been put forward by the opponents of moral hedging.9
Notice that the arguments of the previous sections (to the effect that, given
Alice’s acceptance of {P1, P2, P3, P7}, she ought also to accept propositions like P12
and P14) were strictly inessential to the argument of the present section that it is at
least sometimes possible to make intertheoretic value comparisons and rational to
engage in moral hedging: We could, if we like, simply stipulate that Alice accepts P12
and P14 along with P1−7. The purpose of giving arguments rather than stipulations
is to show, not just that there are some pairs of theories such that an agent who
9Note that the fact that expected choiceworthiness is sometimes undefined does not obviously
problematize even the claim that rational agents are always rationally required to maximize ex-
pected choiceworthiness: In cases where expectations are undefined, this requirement is simply
inert—so long as it is understood, as it must be, to say that an agent is always required to choose
an option with maximal expected choiceworthiness whenever such an option is available, or alter-
natively, that an agent is always rationally prohibited from choosing an option O when there is
some alternative option P such that (i) P has greater expected choiceworthiness than O and (ii)
there is no option Q that has greater expected choiceworthiness than P. (Such qualifications are
necessary not just for the case of undefined expectations but also for cases in which an agent faces
infinitely many options with expected values of of 1, 2, 3, 4, ... or 1/2, 2/3, 3/4, 4/5, ... (Arntze-
nius et al., 2004; Landesman, 1995).) If we conclude that expected choiceworthiness is sometimes
undefined, we might then claim that agents are always required to choose options with maximal
expected choiceworthiness whenever such an option is available and that other requirements of
rationality govern the cases where such an option is not available.
321
divides her beliefs between those theories ought rationally to hedge for her moral
uncertainties and has rational grounds for the intertheoretic comparisons necessary
to do so, but also that these theories are reasonable and internally coherent, rather
than mere artificial combinations of jointly implausible propositions. This makes our
conclusions more relevant to real agents who, we hope, divide their beliefs mainly
among reasonable and internally coherent normative theories.
B.4 Generalizing (or Failing to Generalize) from the Easy Case
Nevertheless, the Easy Case is of course highly idealized, and we would like
to have some indication that its conclusions generalize, particularly to agents who
are more radically uncertain than Alice. The complications involved even in this
simplest of simple cases indicate, I think, that this cannot be done in a single stroke—
establishing content-based aggregation methods between theories takes work, in
particular a careful specification of the content of those theories that will, unless we
proceed by stipulatory fiat, involve substantive questions about the most plausible
way of explicating an initially under-described theory.
Nevertheless, in this section I will briefly address a few particularly salient
dimensions along which we might hope to generalize the idea of content-based in-
tertheoretic aggregation. The general claim I hope to make plausible is that in many
cases it is rational for an agent to accept an “independence assumption” analogous
to P14 that permits scale normalization between theories, or otherwise to distribute
her belief over propositions like P14, P15, and P16 that can be aggregated to yield
322
at least rough normalization and comparability. But I will also attempt to show, at
the end of the section, that this approach to intertheoretic comparisons leaves open
the possibility of incomparability between theories.
B.4.1 Indirect Comparability
A noteworthy feature of the Easy Case is the presence of moral certainty,
namely, Alice’s certainty with respect to the existence and basic nature of hedonic
value.10 If this were a necessary feature of the case—that is, a necessary condition
for content-based aggregation—that would be a discouraging result, since it seems
that an epistemically rational agent will be at least a little uncertain about any
feature of morality.
Fortunately, it is not. Consider the following, slightly less simple case: Car-
men, like Alice, is a committed consequentialist, who is uncertain what things in the
world are non-derivative value-bearers. The three candidates she takes as possible
value-bearers are hedonic experience, aesthetic goods, and knowledge. She divides
her beliefs between three theories: T5 values hedonic experience and aesthetic goods,
but not knowledge; T6 values hedonic experience and knowledge, but not aesthetic
goods; and T7 values aesthetic goods and knowledge, but not hedonic experience.
11
Representing things as we did in the Easy Case, we can say that Carmen has
beliefs about the following propositions, inter alia.
10If Alice accepts P14 then she is also certain, we might say, about the degree of hedonic value.
But even if she accepts P15 instead, or distributes her belief over propositions like P14−17, she is
still certain not only that hedons have value (P2) but also of the explanation/grounding for that
value (P3).
11Thanks to Andrew Sepielli for this case and for pressing me to consider the challenge it
represents.
323
P1 The choiceworthiness of an option is equal to the value of the world that results
from it, where the value of a world is the cardinal sum of non-derivative value
that supervenes on value-bearing empirical phenomena in that world.
P2 Hedonic experiences (i.e., experiences with positive or negative hedonic valence)
are non-derivative value-bearers.
P3 The explanation/grounding for the non-derivative value of hedonic experience
is ϕ.
P4 Aesthetic goods (beauty and the like) are non-derivative value-bearers.
P5 The explanation/grounding for the non-derivative value of aesthetic goods is ψ.
P18 Knowledge is a non-derivative value bearer.
P19 The explanation/grounding for the non-derivative value of knowledge is υ.
P6 The value of one hedon is equal to the value of one aestheton.
P20 The value of one hedon is equal to the value of one epistemon.
P21 The value of one aestheton is equal to the value of one epistemon.
P22 Nothing apart from hedonic experiences, aesthetic goods, and knowledge is a
non-derivative value-bearer.
Carmen then divides her beliefs between the following theories:
T5 (hedons + aesthetons): {P1, P2, P3, P4, P5, P6,¬P18,¬P19,¬P20,¬P21, P22, ...}
324
T6 (hedons + epistemons): {P1, P2, P3,¬P4,¬P5,¬P6, P18, P19, P20,¬P21, P22, ...}
T7 (aesthetons + epistemons): {P1,¬P2,¬P3, P4, P5,¬P6, P18, P19,¬P20, P21, P22, ...}
Are T5−7 intercomparable? As in the Easy Case, this will depend on what Car-
men believes about whether and how the value of potential value-bearers (hedons,
aesthetons, and epistemons) varies between theories that endorse them as value-
bearers. Suppose that Carmen believes the following propositions with certainty:
P23 Given that hedonic experience is a non-derivative value-bearer, the degree of
value borne by a given unit of hedonic experience is independent of whether
aesthetic goods/knowledge are non-derivative value-bearers.
P24 Given that aesthetic goods are non-derivative value-bearers, the degree of value
borne by a given unit of aesthetic goods is independent of whether hedonic
experience/knowledge are non-derivative value-bearers.
P25 Given that knowledge is a non-derivative value-bearer, the degree of value
borne by a given unit of knowledge is independent of whether hedonic experi-
ence/aesthetic goods are non-derivative value-bearers.
Carmen’s acceptance of these independence assumptions allows her to normalize
and compare T5−7: P23 says that the value of a hedon according to T5 is equal to
the value of a hedon according to T6, establishing a normalization of those theories.
Likewise, P24 says that the value of an aestheton according to T5 is equal to the
value of an aestheton according to T7, and P25 says that the value of an epistemon
according to T6 is equal to the value of an epistemon according to T7.
325
Note also that the presence of these independence assumptions—i.e., Carmen’s
belief that how much value is borne by one kind of value bearer does not depend
on what other kinds of value-bearers exist—imposes consistency constraints on the
theory-internal “exchange rates” between value-bearers. That is, given the indepen-
dence assumptions P23−25 and a knowledge of the theory-internal exchange rates for
any two theories, we can deduce what the theory-internal exchange rate of the third
theory must be. For instance, given the independence assumptions P23−25 plus the
facts that P6 ∈ T5 and P20 ∈ T6, we can deduce that P21 ∈ T7.
1. V T7(aestheton) = V T5(aestheton) [from P24]
2. V T5(aestheton) = V T5(hedon) [from P6 ∈ T5]
3. V T5(hedon) = V T6(hedon) [from P23]
4. V T6(hedon) = V T6(epistemon) [from P20 ∈ T6]
5. V T6(epistemon) = V T7(epistemon) [from P25]
6. V T7(aestheton) = V T7(epistemon) [from 1-5]
C. P21 ∈ T7 [from 6]
This sort of reasoning generalizes to cases in which independence assumptions are
replaced by well-defined dependence assumptions (e.g., if rather than accepting P23,
Carmen believes that the value of a hedon according to T5 is twice as great as the
value of a hedon according to T6). In either case, the independence/dependence
assumptions and the theory-internal exchange rates are constrained by the require-
ment of mutual consistency.
326
The sort of reasoning described above also allows us to handle cases in which
some pairs of theories share no overlapping value. For instance, consider an agent
who divides her belief between three theories: one that values only hedonic goods,
another that values only aesthetic goods, and a third that values both hedonic and
aesthetic goods and that regards one hedon as equal in value to one aestheton.
Given independence assumptions analogous to P23−25, to the effect that the value of
a hedon is the same according to the hedonistic theory as according to the pluralistic
theory and likewise for the value of an aestheton, we can infer that the value of a
hedon according to the hedonistic theory is equal to value of a hedon according
to the pluralistic theory, which is equal to the value of an aestheton according to
the pluralistic theory, which is equal to the value of an aestheton according to the
aesthetic theory, and hence that the value of a hedon according to the hedonistic
theory is equal to the value of an aestheton according to the aesthetic theory.12
B.4.2 Uncertainty about Value Independence
So far I have assumed that agents who are uncertain what sorts of things have
value are nevertheless certain either that the degree of value borne by a particular
value bearer is independent of what other value-bearers exist or that it is dependent
in some particular way (e.g., that the existence of a second value-bearer would
12This reasoning by chains of equalities is what vindicates the idea of pluralistic theories as
“bridges” between monistic theories introduced in Chapter 6: The transitivity of equality lets an
agent establish intertheoretic comparisons even when she is not certain of any particular value
claim that unites all the theories in which she has positive credence (as hedonic value does in the
Easy Case) and even when there are some pairs of theories in which she has positive credence that
have no value claim in common (like the hedonistic and aesthetic theories in the last example), so
long as any two theories are linked by a chain of theory-pairs each of which has some value claim
in common that allows for normalization.
327
reduce the value per unit of the first value-bearer by 50%). But what if an agent is
uncertain about the dependence relationship between value-bearers?
Consider Daniel, who divides his belief between classical utilitarianism and a
simple version of prioritarianism, according to which the interests of those above
some threshold of absolute wellbeing count for only half as much as the interests
of those below that threshold. Daniel wants to know how he should weigh these
theories against one another when they conflict. To do this, he must consider the
following question: Does prioritarianism value the interests of the less well off more
than classical utilitarianism, or the interests of the better off less, or both? That
is, what is the relationship between the value of, say, an increment of wellbeing to
an already-well-off person according to classical utilitarianism and the value of that
same increment of wellbeing according to prioritarianism? If Daniel were certain,
say, that if prioritarianism is true, then the interests of the less well off count for
twice as much as they would under classical utilitarianism, while the interests of
the better off have exactly the same degree of moral weight whether utilitarian-
ism or prioritarianism is correct, then he would have no difficulty in normalizing
utilitarianism and prioritarianism. But what if he is uncertain?
To make Daniel’s case precise, let’s suppose that his moral beliefs center on
the following propositions.
P1 The choiceworthiness of an option is equal to the value of the world that results
from it, where the value of a world is the cardinal sum of non-derivative value
that supervenes on value-bearing empirical phenomena in that world.
328
P2 Hedonic experiences (i.e., experiences with positive or negative hedonic valence)
are non-derivative value-bearers.
P26 Nothing apart from hedonic experience is a non-derivative value-bearer.
P27 An additional unit of hedonic experience (i.e. a hedon) has twice as much
value if the subject of that experience has a level of hedonic wellbeing below
threshold t than it does if the subject of that experience has a level of wellbeing
above threshold t.
P28 The value of an additional hedon for a subject below t is independent of whether
P27 is true.
P29 The value of an additional hedon for a subject below t is twice as great if P27
is true than it is if P27 is false.
P30 The value of an additional hedon for a subject above t is independent of whether
P27 is true.
P31 The value of an additional hedon for a subject above t is only half as great if
P27 is true as it is if P27 is false.
Daniel’s uncertainty about the dependence relations between utilitarianism
and prioritarianism, then, amounts to an uncertainty whether P28 ∧ P31 is true or
P29 ∧ P30 is true. Because of this uncertainty, Daniel divides his beliefs between
three theories rather than two.
T8 (utilitarianism): {P1, P2, P26,¬P27...}
329
T9 (“low prioritarianism”): {P1, P2, P26, P27, P28,¬P29,¬P30, P31, ...}
T10 (“high prioritarianism”): {P1, P2, P26, P27,¬P28, P29, P30,¬P31, ...}13
Given this description of the case, it is clear how Daniel ought to normalize
his moral theories for purposes of moral hedging: The value of a hedon for someone
below threshold t according to utilitarianism is equal to the value of a hedon for
someone below threshold t according to low prioritarianism, but only half as great
as the value of a hedon for someone below threshold t according to high prioritarian-
ism. The value of a hedon for someone above threshold t according to utilitarianism
is equal to the value of a hedon for someone above threshold t according to high
prioritarianism, and twice as great as the value of a hedon for someone above thresh-
old t according to low prioritarianism. This means, of course, that the value of a
hedon for anyone (above or below threshold t) is twice as great according to high
prioritarianism as it is according to low prioritarianism.
The more of his prioritarian credence Daniel assigns to high prioritarianism
rather than low prioritarianism, therefore, the greater the total influence of his
prioritarian credence on his decisions under moral uncertainty will be. And this is as
it should be: High prioritarianism says that there is more at stake than utilitarianism
recognizes, i.e., the interests of the less well off are more urgent. Low prioritarianism
13In fact, Daniel will presumably divide his beliefs between four theories, since one version of
classical utilitarianism will include P28∧P31 while another will include P29∧P30. This complication
is relevant if we suppose that Daniel is uncertain between a “high value” and a “low value” version
of classical utilitarianism just as he is uncertain between a “high value” and a “low value” version
of prioritarianism. For simplicity, I have assumed that Daniel is certain of the value of a hedon
given classical utilitarianism (i.e., given ¬P27), in which case the difference between the versions of
utilitarianism that include P28 ∧ P31 and P29 ∧ P30 are interchangeable for all practical purposes.
As far as I can see, uncertainty between high value and low value versions of utilitarianism would
not raise any interestingly new difficulties.
330
says that there is less at stake, i.e., that the interests of the already-well-off just
don’t matter that much, contra utilitarianism. High prioritarianism deserves more
weigh in a morally uncertain agent’s moral calculus than does low prioritarianism.
Of course, like the other cases we have examined in this appendix, Daniel’s
case is artificially simplified. In a more realistic case, Daniel might distribute his
prioritarian belief over a wide—perhaps infinite—range of prioritarian weighting
functions (more sophisticated alternatives to the “hedons count for twice as much
above threshold t” function used above), and for each of those functions, might
have credence in many “higher” or “lower” theories that normalize that particu-
lar weighting function with classical utilitarianism differently. Any attempt at a
realistic representation of such a belief state would yield a number of propositions
and theories much too large to be explicitly enumerated. Nonetheless, the lesson
from Daniel’s case should generalize to more complex cases: Once we make explicit
the agent’s distribution of belief over the various possible states of dependence or
independence that might be exhibited by a value-bearer (hedonic value, aesthetic
value, hedonic-value-for-subjects-below-threshold-t) relative to some set of theories
that assign it non-derivative value, the result will be a more fine-grained collection
of theories, each of which is connected to the others by some definite relations of
dependence or independence that allow for intertheoretic normalization and aggre-
gation.
331
B.4.3 “Infectious Incomparability”
I have so far assumed that, even if an agent divides her beliefs among propo-
sitions like P14−16 that posit different comparisons between the same pair of the-
ories, she does not assign positive credence to a proposition like P17 that posits
incomparability. But this is of course unrealistic—however strong the case for in-
tertheoretic comparability, Alice should have at least some positive credence in the
view that T1 and T2 are incomparable. MacAskill (2013) has pointed out that, given
an expectational decision rule, incomparability is “infectious”: An agent who has
any non-zero credence that incomparable values are at stake in the choice before her
will find that the expected value of her options is undefined. Does this not mean
that, so long as Alice has non-zero credence in P17, this incomparabilist view will
“swamp” whatever credence she has in comparison propositions like P14−16?
If so, then this would be bad news not just for proponents of moral hedging,
but for proponents of expectational reasoning more generally. For one thing, if the
true moral theory turns out to involve incomparable values (say, special obligations
to loved ones versus impersonal moral obligations, or a parent’s special obligations to
one of her children versus another), then any non-zero credence that a given choice
involves competition between those incomparable values will result in expectations
being undefined. But perhaps more importantly, even if the true ethical theory
involves no incomparability, the same problem presents itself from a different direc-
tion for agents attempting to maximize expected value under empirical uncertainty.
For MacAskill’s “infectious incomparability” problem has a close analogue in the
332
problem of “infectious infinities”: An agent deciding whether to perform a practical
option O who has some positive credence that O will result in infinite positive value
(say, by causing another agent to find the true religion and enjoy eternal salvation)
and some positive credence that O will result in infinite negative value (say, by
causing another agent to find a false religion and suffer eternal damnation) will dis-
cover that the expected value of O is undefined (since ∞−∞ is undefined). If an
epistemically rational agent should have positive credence in everything or nearly
everything, then it appears as though straightforward expectational reasoning is
simply impossible (or rather, entirely possible but disappointingly unhelpful in its
results). It seems to me, therefore, that the problem of small credences generating
undefined expectations is a general problem for decision theory, rather than a chal-
lenge to hedging under moral uncertainty specifically—that is, we have good reason
to require that the true decision-theoretic account of choice under uncertainty not
be simply expectational, but include some device for dealing with small probabilities
of extreme or undefined value.14
B.4.4 Incomparability Simpliciter
My focus so far in this appendix has been on cases in which any two theories,
even if they do not share any directly overlapping content (e.g. both attributing
non-derivative value to hedonic experience), are nevertheless connected by a chain
14Possible decision-theoretic approaches to this problem include Nicholas Smith’s theory of “ra-
tionally negligible probabilities” (Smith, 2014, 2016), on which agents are simply permitted to ig-
nore possibilities that fall below a certain probability threshold, and Lara Buchak’s “risk-weighted
expected utility maximization” (Buchak, 2013), on which agents are rationally permitted to adopt
any consistent set of risk attitudes, including attitudes that are less sensitive to extremely low
probabilities of extremely large gains or losses than risk-neutral expected utility maximization.
333
of theories in which each directly linked pair do share some such overlapping content
that allows for normalization. Thus, the approach I have described has limited scope:
It only applies to agents whose moral beliefs satisfy this condition. This leaves open
the possibility that, when two theories in which an agent has positive credence are
not connected by any chain of overlapping, normalizable theories, then they are
simply incomparable.
But there are two other ways in which incomparability can arise, even among
theories appropriately connected by common content: namely, when an agent judges
either that two value bearers are incomparable simplicter (that is, incomparable even
on the hypothesis that they are both genuine, non-derivative value bearers) or when
she judges that the dependency relation for a shared form of value between two
theories is one of incomparability. Let’s consider one final case to illustrate these
points.
In this case, Ella divides her beliefs mainly between classical utilitarianism,
according to which she ought to maximize global hedonic utility, and Nietzschean
perfectionism, according to which she ought to maximize a certain kind of self-
perfection. Considering the positive arguments for both of these views plausible,
she also has some slight credence that both global hedonic utility and Nietzschean
perfection have non-derivative moral value. Ella thus has beliefs about the following
propositions, inter alia.
P1 The choiceworthiness of an option is equal to the [possibly agent-relative] value
of the world that results from it, where the value of a world is the cardinal sum
334
of non-derivative value that supervenes on value-bearing empirical phenomena
in that world.
P2 Hedonic experiences (i.e., experiences with positive or negative hedonic valence)
are non-derivative value-bearers.
P32 Traits that contribute to an agent’s own self-perfection are non-derivative
(agent-relative) value-bearers.
P33 Nothing besides hedonic experience and an agent’s own self-perfection is a
non-derivative value-bearer.
Ella then divides her belief between three theories:
T11 (utilitarianism): {P1, P2,¬P32, P33...}
T12 (perfectionism): {P1, P2, P32,¬P33...}
T13 (pluralism): {P1, P2, P32, P33...}
If, say, Ella’s pluralistic theory T13 includes the proposition that the value of
one hedon is equal to the value of one perfecton, and if she is certain that the value
of a hedon (given that hedons have non-derivative value) is independent of whether
perfectons have value and vice versa, then the case is simple and her three theories
can be easily normalized.
But suppose instead that T13 includes the following proposition.
P34 The value of hedonic experience and the value of self-perfection are incompa-
rable.
335
If P34 ∈ T13, then the approach I have described offers no basis for normalizing
T11−13.
Alternatively, suppose that Ella believes the following with certainty.
P35 The value of a unit of self-perfection given P32 ∧ P2 is incomparable with the
value of a unit of self-perfection given P32 ∧ ¬P2.
P35 asserts that the value of a perfecton according to T12 is incomparable with the
value of a perfecton according to T13. Unlike an assumption of independence (or
any definite dependency, e.g. that perfectons have only half as much value accord-
ing to T13 as they do according to T12), this prevents any intuitive, content-based
normalization between T12 and T13 (and by extension between T12 and T11). So
if Ella accepts P35 (i.e., if P35 ∈ T11, T12, T13), then the approach to intertheoretic
normalization I have been describing will be of no help to her. (The same is true,
of course, if she accepts an analogous proposition to the effect that the value of a
hedon according to T11 is incomparable with the value of a hedon according to T13.)
In the last section, I claimed that an adequate resolution of the various decision-
theoretic puzzles involving small probabilities of incomparable values, infinite values,
etc, should let us avoid the problem of “infectious incomparability” for agents who
have very low credence in intertheoretic incomparability. But this is only true if
Ella’s unconditional credence in P35 and her conditional credence in (P34|P2 ∧ P32)
are very close to zero. If these credences are substantial, even if they are less than 1,
then presumably the decision-theoretic resolution of problems involving minuscule
probabilities will not allow Ella to ignore them.
336
How much this restricts the scope of the account I have given depends, there-
fore, on how plausible one finds propositions like P34 and P35 as opposed to the
competing assumptions of value comparability within pluralistic theories (contra
P34) and independence (or determinate dependence) of the value of one kind of
value-bearer on the existence of other value-bearers (contra P35). For my part, in-
comparabilist theories like the version of T13 that includes P34 seem fairly implausi-
ble, though they may have greater plausibility in the context of particular pluralistic
theories, like T13, that combine particularly irreconcilable sorts of value. And it is
unclear what would ever motivate an agent to assign substantial probability to a
proposition like P35. (In the next section, I will attempt to further strengthen the
case that agents ought to default to independence assumptions like P14 absent spe-
cific grounds for rejecting these assumptions, which will give us additional reason to
be skeptical of propositions like P35.) Nevertheless, the account I have given leaves
room for intertheoretic incomparability even in cases where the moral theories in
which an agent has positive credence are linked together by chains of overlapping
content.
B.5 What Grounds Credence in Comparisons?
I have attempted to show that the rationality of making certain intertheoretic
value comparisons, and of hedging for one’s moral uncertainties on the basis of
those comparisons, can be grounded in an agent’s doxastic state. But it might
appear that this only pushes the problem of intertheoretic value comparisons back
337
a step. An interlocutor might argue: “Granted, it may be rational for an agent who
believes certain intertheoretical comparison propositions—for instance, Alice who
believes that her T1 and T2 each assign the same value to a given unit of hedonic
experience—to act on the basis of those comparisons. But the deeper question is,
what could ever make such a comparison true? And if we lack any satisfactory
answer to this question, does it not undercut the claim that agents like Alice can
be rationally justified in believing such propositions?” In other words, even if we
have explained what grounds the rationality of acting as if certain intertheoretic
comparisons are correct, for agents who accept those comparisons, we have not yet
answered the question of what grounds the intertheoretic comparisons themselves.
In this section I will attempt to respond to this objection—not by giving a definite
answer to the question of what grounds or makes-true intertheoretic comparisons like
P14, but rather by drawing on and extending arguments from the preceding sections
to argue that there must be an answer of one kind or another, and suggesting a few
possibilities.
The argument is simply this: (1) In the context of various forms of non-
normative uncertainty, it is intuitively clear that propositions analogous to P14 are
not only reasonable to believe but also true (at least, insofar as any evaluative
propositions are ever true). And (2) there is no reason to expect that whatever
story we tell about the truth of these propositions will not extend straightforwardly
to propositions like P14 in the context of normative uncertainty.
To motivate (1), consider again the case of Bob, who is certain of the truth of
hedonistic utilitarianism but uncertain what will happen if he presses the red button
338
(and thus divides his belief between two theories of the world, T3 and T4). As we
noted, for Bob to calculate the expected value of pressing the red button, he must
assume something about how the value of a hedon according to T3 compares with the
value of a hedon according to T4—and presumably, he should assume that the value
of a hedon is the same according to both theories, i.e., that the value of a hedon
is independent of what would happen if he presses the red button. Furthermore, it
seems clear that this belief is true.15
What grounds or makes true the proposition that the value of a hedon is
independent of the consequences of pressing the red button? Various stories are
possible. Perhaps it is a fact about possible worlds, i.e., that hedons are equally
valuable in the possible worlds where pressing the red button will save five and
the possible worlds where it will kill ten. Perhaps it is simply the fact of logical
independence, i.e., that propositions about the consequences of pressing the red
button have no logical implications for the value of hedons. Perhaps it is some
weaker abstract truth to the effect that logically independent features of the world
are prima facie independent, i.e., should be treated as independent in the absence
of any facts particular to the case that would ground a dependency between them.16
The latter two hypotheses extend straightforwardly to the case of Alice and
15If hedons simply don’t have value, then it will be trivially true. It could only fail to be true,
as far as I can see, if concepts like value are so hopelessly confused that no proposition about
value can ever be true (a la` the version of moral nihilism considered in Ross (2006) and MacAskill
(2014)).
16This would for instance accommodate the idea that, if there are more sources of value in the
world, each source of value counts for less—even though the proposition that aesthetic goods are
non-derivative value bearers does not (at least on its own) logically entail that hedonic experience
is less valuable than it otherwise would be, there might be non-logical entailments that defeat the
presumption of independence.
339
other similar cases of moral uncertainty: The value of hedons is logically independent
of whether aesthetons have value, and this fact either on its own or coupled with the
absence of defeaters to the presumption of independence might ground the truth of
P14.
17 But the first hypothesis, that of equality across possible worlds, points out
a worry about the analogy between moral and empirical uncertainty: namely, that
basic moral propositions are (plausibly) necessary rather than contingent truths,
which presents an apparent obstacle to spelling out independence in terms of possible
worlds.
But this should not worry us too much, for there are cases of non-normative un-
certainty in which the objects of uncertainty are necessarily true/false propositions
that nevertheless seem clearly analogous to the case of Bob. Consider Frederick, a
precocious six-year-old utilitarian who finds himself faced with a trolley problem:
On the main track, he knows, are exactly 65 people, all of whom will be killed if the
trolley is allowed to continue. Frederick can divert the trolley onto a siding, but if
he does so it will collide with a train car loaded with highly volatile explosives, caus-
ing them to detonate. On a field next to the siding, Frederick can see a marching
band practicing, in a rectangular formation that (Frederick can quickly determine
by counting) is nine people wide and eight people deep. Frederick must make a
decision immediately, and has no time to count all the members of the marching
17Adding to the plausibility of these hypotheses, it is worth noting that the proponent of incom-
parability must explain what grounds the fact that the values of rival moral theories like T1 and
T2 are incomparable, and in doing so will presumably appeal to some similar general principle,
e.g. a prima facie principle that the values of rival theories are incomparable absent some special
fact that grounds their comparability. But why should incomparability be the default state of
affairs rather than independence (especially when independence seems to be the default in cases
of non-normative uncertainty)?
340
band. But fortunately, he has overheard his older sister practicing her multiplication
tables, and remembers that 9× 8 is either 63 or 72. He can’t remember which, and
has no time to improve his epistemic position on this question. As the trolley speeds
toward the junction, either answer seems equally plausible to him. Precocious child
that he is, it occurs to him that he should just split the difference between 63 and
72. While he can’t calculate it exactly, he realizes that whatever number is midway
between 63 and 72 will be greater than 65.
It seems clear to me that, in his state of mathematical uncertainty, Frederick
should not turn the trolley—even though doing so would in fact save more lives
(since 9 × 8 = 72). But, just as in the case of Bob, expectational reasoning in
the face of Frederick’s uncertainty whether 9 × 8 = 72 or 63 requires an implicit
assumption that the importance of the relevant values (hedons, lives saved, etc) is
independent of the mathematical facts, i.e., that Frederick should not value saving
a life any more or less on the hypothesis that 9 × 8 = 72 than on the hypothesis
that 9× 8 = 63.18
18If you think that there is some minimal threshold for rational agency that requires closure of
belief under the relatively simple deductive inferences required to compute 9×8, then of course you
can simply imagine an analogous case involving an adult reasoner who is relevantly uncertain about
some more difficult mathematical proposition. For instance, suppose that Georgiana, a utilitarian
and an accomplished logician whose beliefs meet any closure requirement that could be reasonably
imposed on human agents, must decide whether to kill one trillion people in order to save a number
of people equal to the billionth prime number. (Granted, summoning up intuitions about this case
may require suspending a deal of disbelief about certain implications of utilitarianism.)
There is nothing untoward, as far as I can see, in appealing to the mathematical reasoning
required to compute expectations under conditions of mathematical uncertainty. Granted, someone
who accepts the arithmetical principles required to compute expectations but holds or entertains
false arithmetical beliefs (e.g., that 9× 8 = 63) is in some sense not fully coherent, having at least
some credence in a contradiction. But to hold that ordinary principles of practical rationality do
not apply to an agent who has non-zero credence in a contradiction is to catastrophically restrict
the scope of those principles: For any finite reasoner, there will be some logical and mathematical
truths too difficult to prove in a fixed period of time. If such a reasoner is forced to make a choice,
within such a fixed period of time, that turns on the truth of such a proposition, she must form
some belief about it, and presumably that belief ought to be of a strength less than certainty.
341
If we are tempted to ground the independence assumption in facts about possi-
ble worlds in the case of Bob, then, what should we say about the case of Frederick?
Perhaps we can simply stay the course, appealing not to metaphysically or even log-
ically possible worlds (since there are no metaphysically or logically possible worlds
in which 9×8 = 63) but to epistemically possible worlds.19 Or perhaps the possibil-
ity of expectational reasoning in the face of mathematical uncertainty should lead
us to accept an alternative story about independence assumptions, in Bob’s case as
well as Alice’s and Frederick’s, that makes no appeal to possible worlds.
In any case, the question of what precisely grounds intertheoretic value compar-
isons—what makes it true in the Easy Case that the value of a hedon is independent
of whether aesthetons have value (or that it is dependent in some particular way)—
is one best left to the metaphysicians. But it seems clear that a morally uncertain
agent can reasonably accept propositions like P14, and therefore that the rationality
of hedging and of making the intertheoretic value comparisons that hedging requires
can be grounded in an agent’s reasonable beliefs.
19Of course, it seems that merely epistemic “possible worlds,” like that in which 9 × 8 = 63,
must be in some sense ersatz. There are no Lewisian possible worlds, as real as our own, where
9 × 8 = 63. But since few believe that there are any fully non-ersatz, Lewisian possible worlds,
this should not be seen as a great cost to most of those who are tempted to ground the truth of
Bob’s independence assumption in modal facts.
342
Bibliography
Aboodi, R., A. Borer, and D. Enoch (2008). Deontology, Individualism, and Uncer-
tainty: A Reply to Jackson and Smith. Journal of Philosophy 105 (5), 259–272.
Alpert, M. and H. Raiffa (1982). A Progress Report on the Training of Probability
Assessors. In D. Kahneman, P. Slovic, and A. Tversky (Eds.), Judgment Under
Uncertainty: Heuristics and Biases, pp. 294–305. Cambridge University Press.
Arntzenius, F. (2014). Utilitarianism, Decision Theory and Eternity. Philosophical
Perspectives 28 (1), 31–58.
Arntzenius, F., A. Elga, and J. Hawthorne (2004). Bayesianism, Infinite Decisions,
and Binding. Mind 113 (450), 251–283.
Arpaly, N. and T. Schroeder (1999). Praise, Blame and the Whole Self. Philosophical
Studies 93 (2), 161–188.
Bennett, J. (1974). The Conscience of Huckleberry Finn. Philosophy 49 (188), 123–
134.
Bennett, K. (2011). By Our Bootstraps. Philosophical Perspectives 25 (1), 27–41.
Bernoulli, D. (1954 (1738)). Exposition of a New Theory on the Measurement of
Risk. Econometrica: Journal of the Econometric Society 22 (1), 23–36.
Bhikkhu, T. (2013, November). San˜n˜a Sutta: Perceptions (AN 7.46), translated
from the Pali by Thanissaro Bhikkhu.
Bostrom, N. (2009, January). Moral Uncertainty — Towards a Solution? http://
www.overcomingbias.com/2009/01/moral-uncertainty-towards-a-solution.html.
Bostrom, N. (2011). Infinite Ethics. Analysis and Metaphysics 10, 9–59.
Broome, J. (1997). Is Incommensurability Vagueness? In Incommensurability,
Incomparability, and Practical Reason. Harvard University Press.
Broome, J. (1999). Normative Requirements. Ratio 12 (4), 398–419.
343
Broome, J. (2001). Normative Practical Reasoning. Aristotelian Society Supple-
mentary Volume 75 (1), 175–193.
Buchak, L. (2013). Risk and Rationality. Oxford University Press.
Bykvist, K. (2013). Evaluative Uncertainty, Environmental Ethics, and Consequen-
tialism. In R. I. Hiller, Avram and L. Kahn (Eds.), Consequentialism and Envi-
ronmental Ethics. Routledge.
Camerer, C., G. Loewenstein, and M. Weber (1989). The Curse of Knowledge in Eco-
nomic Settings: An Experimental Analysis. Journal of Political Economy 97 (5),
1232–1254.
Chang, R. (Ed.) (1997a). Incommensurability, Incomparability and Practical Reason.
Harvard University Press.
Chang, R. (1997b). Introduction. In Incommensurability, Incomparability, and
Practical Reason. Harvard University Press.
Colyvan, M., D. Cox, and K. Steele (2010). Modelling the Moral Dimension of
Decisions. Nouˆs 44 (3), 503–529.
Crouch, W. (2010). Moral Uncertainty and Intertheoretic Comparisons of Value.
Bphil thesis, University of Oxford.
Daniels, N. (1979). Wide Reflective Equilibrium and Theory Acceptance in Ethics.
Journal of Philosophy 76 (5), 256–282.
de Lazari-Radek, K. and P. Singer (2014). The Point of View of the Universe:
Sidgwick and Contemporary Ethics. Oxford University Press.
Diener, E., E. M. Suh, R. E. Lucas, and H. L. Smith (1999). Subjective Well-Being:
Three Decades of Progress. Psychological Bulletin 125 (2), 276–302.
Dreier, J. (2000). Dispositions and Fetishes: Externalist Models of Moral Motiva-
tion. Philosophical and Phenomenological Research 61 (3), 619–638.
Elga, A. (2010). How to Disagree About How to Disagree. In T. Warfield and
R. Feldman (Eds.), Disagreement. Oxford University Press.
Foot, P. (1972). Morality as a System of Hypothetical Imperatives. The Philosophical
Review 81 (3), 305–316.
Gauthier, D. P. (1986). Morals by Agreement. Oxford University Press.
Gert, J. (2003). Requiring and Justifying: Two Dimensions of Normative Strength.
Erkenntnis 59 (1), 5–36.
GiveWell (2013, November). Mass Distribution of Long-Lasting Insecticide-Treated
Nets (LLINs). http://www.givewell.org/international/technical/programs/
insecticide-treated-nets.
344
Gracely, E. J. (1996). On the Noncomparability of Judgments Made by Different
Ethical Theories. Metaphilosophy 27 (3), 327–332.
Greaves, H. and T. Ord (forthcoming). Moral Uncertainty about Population Ethics.
Journal of Ethics and Social Philosophy .
Greene, J. (2008). The Secret Joke of Kant’s Soul. In W. Sinnott-Armstrong (Ed.),
Moral Psychology, Vol. 3. MIT Press.
Greenwell, J. R. (1977). Abortion and Moral Safety. Cr´ıtica: Revista Hispanoamer-
icana de Filosof´ıa 9 (27), 35–48.
Guerrero, A. A. (2007). Don’t Know, Don’t Kill: Moral Ignorance, Culpability, and
Caution. Philosophical Studies 136 (1), 59–97.
Gustafsson, J. E. and O. Torpman (2014). In Defence of My Favourite Theory.
Pacific Philosophical Quarterly 95 (2), 159–174.
Hadar, J. and W. R. Russell (1969). Rules for Ordering Uncertain Prospects. The
American Economic Review 59 (1), 25–34.
Ha´jek, A. (2003). Waging war on pascal’s wager. Philosophical Review 112 (1),
27–56.
Ha´jek, A. (2012). Interpretations of Probability. In E. N. Zalta (Ed.), The Stanford
Encyclopedia of Philosophy (Winter 2012 ed.).
Harman, E. (2015). The Irrelevance of Moral Uncertainty. In R. Shafer-Landau
(Ed.), Oxford Studies in Metaethics, Volume 10. Oxford University Press.
Harman, G. (1975). Moral Relativism Defended. Philosophical Review 84 (1), 3–22.
Hawley, P. (2008). Moral Absolutism Defended. Journal of Philosophy 105 (5),
273–275.
Hedden, B. (2016). Does MITE Make Right? On Decision-Making under Normative
Uncertainty. In R. Shafer-Landau (Ed.), Oxford Studies in Metaethics, Volume 11.
Oxford University Press.
Hobbes, T. (2004 (1651)). Leviathan. Clarendon Press.
Horgan, T. and M. Timmons (2010). Untying a Knot From the Inside Out: Reflec-
tions on the “Paradox”of Supererogation. Social Philosophy and Policy 27 (2),
29–63.
Horty, J. F. (2014). Reasons as Defaults. Oxford University Press Usa.
Hudson, J. L. (1989). Subjectivization in Ethics. American Philosophical Quar-
terly 26 (3), 221–229.
345
Huemer, M. (2010). Lexical Priority and the Problem of Risk. Pacific Philosophical
Quarterly 91 (3), 332–351.
Isaacs, Y. (2014). Duty and Knowledge. Philosophical Perspectives 28 (1), 95–110.
Jackson, F. (1991). Decision-Theoretic Consequentialism and the Nearest and Dear-
est Objection. Ethics 101 (3), 461–482.
Jackson, F. and R. Pargetter (1986). Oughts, Options, and Actualism. Philosophical
Review 95 (2), 233–255.
Jackson, F. and M. Smith (2006). Absolutist Moral Theories and Uncertainty.
Journal of Philosophy 103 (6), 267–283.
Jackson, F. and M. Smith (2016). The Implementation Problem for Deontology. In
E. Lord and B. Maguire (Eds.), Weighing Reasons, pp. 279– 291. Oxford Univer-
sity Press.
Joyce, J. M. (2002). Levi on Causal Decision Theory and the Possibility of Predicting
One’s Own Actions. Philosophical Studies 110 (1), 69–102.
Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
Kahneman, D. and A. Deaton (2010). High Income Improves Evaluation of Life
but not Emotional Well-Being. Proceedings of the National Academy of Sci-
ences 107 (38), 16489–16493.
Kahneman, D., B. L. Fredrickson, C. A. Schreiber, and D. A. Redelmeier (1993).
When More Pain is Preferred to Less: Adding a Better End. Psychological sci-
ence 4 (6), 401–405.
Kalai, E. and M. Smorodinsky (1975). Other Solutions to Nash’s Bargaining Prob-
lem. Econometrica: Journal of the Econometric Society 43 (3), 513–518.
Kant, I. (1960 (1793)). Religion within the Limits of Reason Alone. Harper and
Row. tr. T. M. Greene and H. H. Hudson.
Kavka, G. S. (1983). The Toxin Puzzle. Analysis 43 (1), 33–36.
Kerstein, S. J. (2002). Kant’s Search for the Supreme Principle of Morality. Cam-
bridge University Press.
Keynes, J. M. (1921). A Treatise on Probability. London: MacMillan and Co.
Kolodny, N. (2005). Why Be Rational? Mind 114 (455), 509–563.
Korsgaard, C. M. (1996). The Sources of Normativity. Cambridge University Press.
Landesman, C. (1995). When to Terminate a Charitable Trust? Analysis 55 (1),
12–13.
346
Lockhart, T. (1977). Another Moral Standard. Mind 86 (344), 582–586.
Lockhart, T. (2000). Moral Uncertainty and Its Consequences. Oxford University
Press.
MacAskill, W. (2013). The Infectiousness of Nihilism. Ethics 123 (3), 508–520.
MacAskill, W. (2014). Normative Uncertainty. Ph. D. thesis, University of Oxford.
Marquis, D. (1989). Why Abortion is Immoral. The Journal of Philosophy 86 (4),
183–202.
Moller, D. (2011). Abortion and Moral Risk. Philosophy 86 (03), 425–443.
Nissan-Rozen, I. (2015). Against Moral Hedging. Economics and Philosophy 31 (3),
1–21.
Norcross, A. (2006). The Scalar Approach to Utilitarianism. In H. West (Ed.), The
Blackwell Guide to Mill’s Utilitarianism, pp. 217–32. Wiley-Blackwell.
Nover, H. and A. Ha´jek (2004). Vexing Expectations. Mind 113 (450), 237–249.
Oddie, G. (1994). Moral Uncertainty and Human Embryo Experimentation.
Medicine and Moral Reasoning 3, 144.
Oddie, G. and P. Menzies (1992). An Objectivist’s Guide to Subjective Value.
Ethics 102 (3), 512–533.
Parfit, D. (1971). Personal Identity. Philosophical Review 80 (January), 3–27.
Parfit, D. (1973). Later Selves and Moral Principles. In A. Montefiore (Ed.), Phi-
losophy and Personal Relations. Routledge and Kegan Paul.
Parfit, D. (1984). Reasons and Persons. Oxford University Press.
Parfit, D. (2011). On What Matters. Oxford University Press.
Pascal, B. (1852/1669). Pense´es. Dezobry et E. Magdeleine.
Pascal, B. (1997 (1657)). The Provincial Letters. Wipf and Stock Publishers.
Pfeiffer, R. S. (1985). Abortion Policy and the Argument from Uncertainty. Social
Theory and Practice 11 (3), 371–386.
Portmore, D. (2008). Are Moral Reasons Morally Overriding? Ethical Theory and
Moral Practice 11 (4), 369–388.
Portmore, D. (2011). Commonsense Consequentialism: Wherein Morality Meets
Rationality. Oxford University Press USA.
Portmore, D. (2016). Uncertainty, Indeterminacy, and Agent-Centered Constraints.
Australasian Journal of Philosophy , 1–15.
347
Quinn, W. (1993). Putting Rationality in Its Place. In Morality and Action, pp.
228 – 255. Cambridge University Press Cambridge.
Railton, P. (1986). Moral Realism. Philosophical Review 95 (2), 163–207.
Rawls, J. (1951). Outline of a Decision Procedure for Ethics. Philosophical Re-
view 60 (2), 177–197.
Raz, J. (1975). Permissions and Supererogation. American Philosophical Quar-
terly 12 (2), 161–168.
Reisner, A. (2013). Is the Enkratic Principle a Requirement of Rationality? Organon
F 20 (4), 436–462.
Roese, N. J. and K. D. Vohs (2012). Hindsight Bias. Perspectives on Psychological
Science 7 (5), 411–426.
Rosen, G. (2002). Culpability and Ignorance. Proceedings of the Aristotelian Soci-
ety 103 (1), 61–84.
Rosen, G. (2004). Skepticism About Moral Responsibility. Philosophical Perspec-
tives 18 (1), 295–313.
Ross, J. (2006). Rejecting Ethical Deflationism. Ethics 116 (4), 742–768.
Sepielli, A. (2006). Review: Moral Uncertainty and Its Consequences. Ethics 116 (3),
601–604.
Sepielli, A. (2009). What to Do When You Don’t Know What to Do. Oxford Studies
in Metaethics 4, 5–28.
Sepielli, A. (2010). ‘Along an Imperfectly-Lighted Path’: Practical Rationality and
Normative Uncertainty. Ph. D. thesis, Rutgers University Graduate School - New
Brunswick.
Sepielli, A. (2012). Normative Uncertainty for Non-Cognitivists. Philosophical Stud-
ies 160 (2), 191–207.
Sepielli, A. (2013). Moral Uncertainty and the Principle of Equity among Moral
Theories. Philosophy and Phenomenological Research 86 (3), 580–589.
Sepielli, A. (2014a). Should You Look Before You Leap? The Philosophers’ Maga-
zine 66, 89–93.
Sepielli, A. (2014b). What to Do When You Don’t Know What to Do When You
Don’t Know What to Do. . . . Nouˆs 48 (3), 521–544.
Sepielli, A. (2016). Moral Uncertainty and Fetishistic Motivation. Philosophical
Studies 173 (11), 2951–2968.
348
Sidgwick, H. (1962 (1874)). The Methods of Ethics (7th ed.). Palgrave Macmillan.
Singer, P. (1972). Famine, Affluence, and Morality. Philosophy & Public Af-
fairs 1 (3), 229–243.
Singer, P. (2005). Ethics and Intuitions. The Journal of Ethics 9 (3-4), 331–352.
Smith, M. (1994). The Moral Problem. Blackwell.
Smith, M. (2002). Evaluation, Uncertainty and Motivation. Ethical Theory and
Moral Practice 5 (3), 305–320.
Smith, N. J. J. (2014). Is Evaluative Compositionality a Requirement of Rationality?
Mind 123 (490), 457–502.
Smith, N. J. J. (2016). Infinite Decisions and Rationally Negligible Probabilities.
Mind (500), 1–14.
Soll, J. B. and J. Klayman (2004). Overconfidence in Interval Estimates. Journal
of Experimental Psychology: Learning, Memory, and Cognition 30 (2), 299.
Talbott, W. (2015). Bayesian Epistemology. In E. N. Zalta (Ed.), The Stanford
Encyclopedia of Philosophy (Summer 2015 ed.).
Tarsney, C. Normative Uncertainty and Social Choice. unpublished.
Tarsney, C. (forthcoming a). Intertheoretic Value Comparison: A Modest Proposal.
Journal of Moral Philosophy .
Tarsney, C. (forthcoming b). Moral Uncertainty for Deontologists. Ethical Theory
and Moral Practice.
Taylor, C. (1999). The Atomists: Leucippus and Democritus. Fragments, A Text
and Translation with Commentary. Toronto: University of Toronto Press.
Tenenbaum, S. (2017). Action, Deontology, and Risk: Against the Multiplicative
Model. Ethics 123 (3), 674–707.
Thomson, J. J. (1976). Killing, Letting Die, and the Trolley Problem. The
Monist 59 (2), 204–217.
Twain, M. (1884). Adventures of Huckleberry Finn. Chatto & Windus.
Urmson, J. O. (1958). Saints and Heroes. In A. I. Melden (Ed.), Essays in Moral
Philosophy. University of Washington Press.
Vallentyne, P. and S. Kagan (1997). Infinite Value and Finitely Additive Value
Theory. The Journal of Philosophy 94 (1), 5–26.
Van Fraassen, B. C. (1989). Laws and Symmetry. Oxford University Press.
349
von Neumann, J. and O. Morgenstern (1947). Theory of Games and Economic
Behavior (2nd ed.). Princeton University Press.
Weatherson, B. (2007). Disagreeing about disagreement.
Weatherson, B. (2013). Disagreements, Philosophical and Otherwise. In J. Lackey
and D. Christensen (Eds.), The Epistemology of Disagreement: New Essays, pp.
54. Oxford University Press.
Weatherson, B. (2014). Running Risks Morally. Philosophical Studies 167 (1), 141–
163.
Wedgwood, R. (2013). Akrasia and Uncertainty. Organon F 20 (4), 484–506.
Williams, B. (1979). Internal and External Reasons. In R. Harrison (Ed.), Rational
Action, pp. 101–113. Cambridge University Press.
Williams, B. (1981). Persons, Character, and Morality. In James Rachels (Ed.),
Moral Luck. Cambridge University Press.
Wise, J. (2013, July). Giving Now vs. Later: A Summary.
Wittgenstein, L. (1965). A Lecture on Ethics. The Philosophical Review 74 (1),
3–12.
Wolf, S. (1982). Moral Saints. The Journal of Philosophy 79 (8), 419–439.
Zimmerman, M. J. (1997). Moral Responsibility and Ignorance. Ethics 107 (3),
410–426.
Zimmerman, M. J. (2008). Living with Uncertainty: The Moral Significance of
Ignorance. Cambridge University Press.
350