ABSTRACT 
 
 
 

 
Title of Dissertation: REPOSITIONING COGNITIVE KINDS   
  
 Aida Roige Mas, Doctor of Philosophy, 2022 
  
Dissertation directed by: Distinguished University Professor Peter 

Carruthers, Department of Philosophy 
 
 
 
This dissertation puts forward a series of theoretical proposals aimed to advance our 

understanding of cognitive kinds. The first chapter introduces the general debates that provide 

the philosophical underpinnings for the topics addressed in each of the following chapters. 

Chapter two compares and distinguishes between modules of the mind and mechanisms-as-

causings, arguing that they should not be conflated in cognitive science. Additionally, it provides 

a novel “toolbox” model of accounts of mechanisms, and discusses what makes any such 

account adequate. Chapter three addresses the question of whether there is a role within the new 

mechanistic philosophy of science for representations. It advances a proposal on how to carve 

working entity types, so that they may include representational explanans. Chapter four offers an 

account of mental disorders, one that captures the regulative ideal behind psychiatry’s inclusion 

of certain conditions as psychopathologies. Mental disorders are alterations in the production of 

some mental outputs (e.g. behaviors, beliefs, emotions, desires), such that their degree of 

reasons-responsiveness is extremely diminished with respect to what we would folk-

psychologically expect it to be. 



  
 

 

 
 
 
 
 
 
 
 
 

REPOSITIONING COGNITIVE KINDS   
 
 
 

by 
 
 

             Aida Roige Mas 
 
 
 
 
 

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 

2022 
 
 
 

 
 
 
 
 
 
 
Advisory Committee: 
Professor Peter Carruthers, Chair 
Professor Lindley Darden 
Professor Eric Saidel 
Professor Harjit Bhogal 
Professor Bob Slevc 

 
 
 



  
 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

© Copyright by 
Aida Roige Mas 

2022 
 
 
 
 
 
 
 
 
 
 
 
 
 



 

 ii  

Dedication 

For my parents and àvia 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

 



 

 iii  

Acknowledgements 

I am deeply grateful and indebted to many for their support throughout the years, not 

only throughout the writing of this dissertation, but throughout my life more 

generally. I am profoundly indebted to my parents, who always believed in me and 

encouraged me to pursue my path even if it was out of the ordinary. Also to my 

beloved àvia, to whom I dedicate this dissertation I promised to her on her last days. I 

am most indebted to Peter Carruthers, who is not only a brilliant philosopher and 

cognitive scientist from whom I learned a lot and who helped me tremendously in this 

project, but is the best supervisor I could have ever asked for. I don’t think there is 

any mentor that is more kind, diligent, insightful and dedicated than Peter. I am very 

fortunate that he guided me so generously throughout this dissertation and my career, 

and I have accumulated a profound debt of gratitude to him. I owe many thanks as 

well to Lindley Darden, who has in effect come to be another advisor to me. Lindley 

has been incredibly supportive and encouraging of my work, revising and 

commenting on it countless times, and guiding me throughout with her knowledge of 

philosophy of science and biology. I am deeply indebted to both of them for my 

intellectual and professional development.  

Many thanks are owed as well to Eric Saidel, whom I had many stimulating 

discussions with and has provided insightful comments on my work. I am also 

indebted to Harjit Bhogal and Bob Slevc, from whom I learned much and I am very 

grateful to count with as committee members. At the University of Maryland, College 

Park, I was fortunate to find an engaging and nurturing intellectual community. Many 

thanks to Sam Kerstein for introducing me to biomedical ethics and for his support 



 

 iv  

more generally. To Georges Rey, for many enjoyable conversations full of insights. 

Special thanks to Andrew Fyfe for his companionship throughout our shared doctoral 

journey. To him, Christopher Masciari, Julia Janczur, and Heather Adair, whom I am 

lucky to call friends. I am grateful as well to the many faculty members, colleagues 

and fellow graduate students from which I benefited during my graduate career, 

including Dan Moller, Elizabeth Schechter, Christopher Morris, Aiden Woodcock, 

Julius Schöenherr, Mike McCourt, Lia Curtis-Fine, Shen Pan, Kyley Ewing, Cody 

Gomez, Evan Westra, Jeremiah Tillman, Xintong Wang, Zhaoqi Hu, Ken Glazer, 

Louise Gilman, and many others. I also thank the DC History and Philosophy of 

Biology (DCHPB) group (including Makmiller Pedroso, Joan Straumanis, Kalewold 

Kalewold, and other past and current members) where I found a vibrant community. 

I am also deeply grateful to those who encouraged me during my academic journey, 

including faculty and colleagues at Universitat Autònoma de Barcelona (specially 

Daniel Quesada, Thomas Sturm, Josep Manuel Udina, and Èric Arnau Soler) and 

CSIC (Mario Toboso and Fanny Brotons). I also have countless people to thank for 

their comments as I have presented my ideas at conferences, workshops, reading 

groups and works-in-progress. A Fulbright-Spain fellowship supported me my first 

two years in the program and allowed me to take extra courses. 

Last, but not least, I would like to give many heartfelt thanks to Andrés, for his 

patience, love and support throughout the making of this dissertation. Thanks as well 

to the rest of my family, especially Ferran, Sergi, padrí, and in loving memory of 

those who unfortunately, passed away while I was away. 



 

 v  

Table of Contents 
 
 
Dedication ................................................................................................................... ii  
Acknowledgements .................................................................................................... iii 
Table of Contents ........................................................................................................ v 
Chapter 1: Introduction ............................................................................................... 1 

1. How is the mind/brain composed? Two sorts of mechanisms in cognitive 
science ................................................................................................................... 2 
2. Can the New Mechanistic Philosophy of Science find a role for 
representations? ..................................................................................................... 7 
3. A material girl in a normative world: an account of mental disorders ....... 11 

Chapter 2. How is the mind/brain composed? Two sorts of mechanisms in cognitive 
science ................................................................................................................ 17 
1. Introduction ................................................................................................. 17 
2. Modules or mechanisms-as-systems ........................................................... 21 
3. An illustration of a mechanism-as-system: face recognition ...................... 33 
4. Mechanisms-as- causings ............................................................................ 36 
5. An illustration of a mechanism-as-a-sort-of-causing: placebo analgesia ... 39 
6. The relationship between mechanisms-as-systems and as-causings .......... 43 
7. How to treat work in progress ..................................................................... 48 

Chapter 3: Can the New Mechanistic Philosophy of Science find a role for 
representations? .................................................................................................. 52 
1. Introduction ................................................................................................. 52 
2. Representations in cognitive science .......................................................... 58 
3. Fitting representations in the MDC account ............................................... 62 
4. Working-entity-hood ................................................................................... 74 
5. Getting it ARIght ........................................................................................ 90 
6. Conclusion .................................................................................................. 95 

Chapter 4: A material girl in a normative world: the extreme reasons-
irresponsiveness account of mental disorders .................................................... 96 
1. Introduction ................................................................................................. 96 
2. Two questions about mental disorders ........................................................ 98 
3. What makes a disorder mental? ................................................................ 106 
4. Considerations favoring my account of mental disorders ......................... 117 
5. Why does psychiatry restrict itself to disorders of reasons-responsiveness?
 123 
6. Conclusion ................................................................................................ 126 

Bibliography ............................................................................................................ 128 
 
 
 
 
  



 

 

1 
 

Chapter 1: Introduction 
 
 
 
 
One of the more useful distinctions in philosophy is that of tokens versus types: we 

distinguish concrete particulars from general sorts of things. Science is mostly in the 

business of providing explanations and predictions of general (as opposed to 

particular) phenomena, and it employs general taxonomical categories (“types” or 

“kinds”) to do so. However, in cognitive science, what exactly can we say about those 

kinds? And what makes some such categories better than others for induction and 

explanation? 

Instead of abstract, general recipes that don’t capture the particularities of the 

cognitive domain, I believe that the work that remains to be done is closer to the 

scientific ground in which the discipline develops. Rather than providing a general 

account of kind-hood for the cognitive sciences, I believe we should answer this 

question on a case-by-case basis. Thus, my dissertation aims to provide three different 

—but related— contributions to debates involving cognitive kinds. The first paper 

distinguishes between two sorts of kinds that produce cognitive phenomena: modules 

of the mind (“systems” or “mechanisms”) and mechanisms-as-causes. The second 

paper provides an account of how to carve out a mechanism’s working entities such 

that these may include representational ones. The third paper provides a novel 

account of what sort of thing mental disorders are. 

 



 

 

2 
 

The aim of this introduction is to sketch the philosophical debates in which these 

three papers are situated. After providing the relevant philosophical background, I 

will briefly summarize the thesis of each paper as well as the contribution it makes to 

those debates. 

 

1. How is the mind/brain composed? Two sorts of mechanisms 
in cognitive science 

 
 
The question of mental architecture, or of what the underlying structure of the human 

mind/brain consists in, has involved at least two major debates.1 The first debate 

concerns the type of processing by which the mind/brain converts inputs into outputs: 

classicists (e.g. Chomsky, Fodor, Pylyshyn) considered that this was done via symbol 

manipulation —analogous to symbolic computation in digital computers. Meanwhile, 

connectionists (e.g. Hinton, McClelland) held that this processing occurred via 

dynamic, parallel, and patterned activity in networks that connect simple processing 

devices; a proposal they saw as more biologically plausible (Dawson, 1998).2 The 

debate around type of cognitive processing slowed as two things became clear: first, 

that classical architectures could be implemented by neural networks and vice-versa; 

and second, that the mind/brain could involve more than one type of information 

 
1 In these papers (and more generally), I assume that some version of physicalism is true.  
2 Arguably, a third contender appeared later on: embodied and extended cognitive science, which 
emphasized the acting and interacting in the world aspects of information processing. 



 

 

3 
 

processing, and that indeed it seemed to do so for different information-processing 

tasks.  

This debate between classicists and connectionists paved the way for a second debate: 

one concerning the functional architecture of the mind/brain. It was, and still is, 

common for cognitive scientists to talk about the “modules,” “systems,” or 

“mechanisms” composing the mind/brain: long-term memory, face recognition, visual 

perception, and so on. This approach of breaking down the mind/brain into its 

functional components is informed by faculty psychology, and has as precedent Franz 

Gall’s phrenology (1835). Both faculty psychology and Gall viewed the mind as 

compositional: i.e., as something that could be best understood when explained in 

terms of separate functions, powers, or faculties. In philosophy of cognitive science, 

during the 1990s and 2000s a major point of contention was the extent of, and the 

characteristics of, modules of the mind. Regarding the extent of mental modules, 

Jerry Fodor (1983) held that only the peripheral systems of the mind were modular: 

that is, only the input (perception and language) and output (action) systems were 

modular, while central cognition (higher-order processing) was not. Modular systems, 

according to him, are informationally encapsulated: they cannot rely on information 

held elsewhere in the mind during the course of its processing. In contrast, central 

cognition is isotropic: it had access to all domains and could potentially use any 

relevant available information.  Fodor’s view was countered by many authors who 

argued for massive modularity. Massive modularity entails that the entirety of the 

mind/brain is modular. For this to be possible, information encapsulation (as Fodor 

described it) had to go. Many authors disputed as well other aspects of Fodor’s 



 

 

4 
 

characterization of modules: content domain specificity, strong localizability, shallow 

outputs, innateness, fast processing and automaticity (see e.g. Barrett & Kurzban, 

2006; Carruthers, 2006; Coltheart, 1999). 

Peter Carruthers provided, in his 2006 book, what I consider to be the correct account 

of modules —or at least, the closest to the notion of modularity actually used by 

cognitive scientists. He characterized modules as “isolable function-specific 

processing systems, all or almost all of which are domain specific (in the content 

sense), whose operations aren’t subject to the will, which are associated with specific 

neural structures (albeit sometimes spatially dispersed ones), and whose internal 

operations may be inaccessible to the remainder of cognition” (Carruthers, 2006, p. 

12). He distinguished between narrow-scope information encapsulation (Fodor’s) and 

wide-scope information encapsulation —which is true of a system if it has access to 

some, but not all, exogenous information during the course of its operations. 

Carruthers (2006) also argued that comparative psychology alongside evolutionary 

considerations make the most plausible case for the massive modularity of mind 

hypothesis.  

Independently of these developments, philosophers of science were identifying a 

plurality of types of explanation in the sciences. New Mechanists argued that 

explanations in certain special sciences were in fact mechanistic, a notion which they 

aimed to elucidate. At a minimum, mechanistic explanations work by modelling the 

mechanism in the world causally responsible for the phenomenon of interest. 



 

 

5 
 

Following Krickel (2019) and Nicholson (2012), I will classify their proposed 

accounts of mechanism into two sorts: 

- Mechanisms-as-systems: mechanisms that consist of stable arrangements, structures, 

and interacting parts such that their combined operation produces predetermined 

outcomes. These will include early accounts of mechanisms by Stuart Glennan (1996, 

2002, 2010), William Bechtel (2008a, 2008b), Bechtel and Robert C. Richardson 

(1993), and Bechtel and Adele Abrahamsen (2005). 

- Mechanisms-as-causes: mechanisms that aren’t object-like but process-like, in the 

sense that their operation is a manifestation of the causal processes involving several 

entities that act and interact.  These will include the accounts of Peter Machamer et al. 

(2000), Phillis Illari and Jon Williamson (2012), and recently Glennan (2017). 

Recent debates in the literature involving New Mechanisms have been about the 

suitability (or lack thereof) of one or other account to a certain (sub-)discipline within 

the special sciences. For example, does Machamer et al.’s (2000) account capture the 

mechanisms involved in explanations in evolutionary biology? (Skipper & Millstein, 

2005). What about neuroscience? (Craver, 2007; Chemero & Silberstein, 2008). The 

assumption seems to be that if any such account is successful, it is so because it 

captures the specific type of explanation deployed in a given subdiscipline. 

At this point, it is worth noting that debates about modularity and debates about 

mechanisms are related, even if I haven’t seen them connected in the literature. 

Insofar as New Mechanists assume that the mind/brain is composed of mechanisms, 



 

 

6 
 

are they providing a genuine alternative to modules as functionally dissociable units, 

or are they presenting modules under a different guise? If the latter, which notion of 

modularity were they employing? Similarly, do discussions on modularity give 

support to one account of mechanism over the others? And, do all explanations of 

cognitive phenomena rely on modules, or on mechanisms? 

Chapter 2, “How is the mind/brain decomposed? Two sorts of mechanisms in 

cognitive science” aims to bridge this gap, providing an answer to these questions. I 

argue that post-Fodorean modules of the mind (such as those characterized in 

Carruthers’ account) can be identified with mechanisms-as-systems (such as those 

characterized in Bechtel’s account), since both families of views place similar 

conditions on modules/mechanisms being functionally dissociable, stable 

architectural components of the mind/brain. However, not all explanations of 

cognitive phenomena involve identifying a module(s) performing its proprietary 

function: some cognitive phenomena are the causal product of the workings of 

different systems in ways that go beyond their proper functions. In other words, some 

cognitive phenomena are produced by mechanisms-as-causes, and as such, their 

production does not entail the existence of a stable, dedicated system for their 

production. Moreover, I argue that at early stages of research, scientists should use 

the “minimal notion” of mechanism, which allows one to cash partial explanations in 

mechanistic terms without carrying over metaphysical commitments on what the 

target mechanism is like.  



 

 

7 
 

This paper makes a contribution to at least two different debates: (1) the debates on 

modularity of mind, and (2) the debates on what is the “best” account of mechanisms 

for a given scientific discipline. I address (1) by highlighting the minimal conditions a 

system must have in order to be a “module,” and pointing out how these overlap with 

the conditions proposed by the mechanists-as-systems hypothesis; and (2) by arguing 

that the proprietary domain of a discipline may involve the action of more than one 

sort of mechanism, and so it is misguided to favor a single mechanistic account as the 

blanket answer to explanations in that domain. 

 

2. Can the New Mechanistic Philosophy of Science find a role 
for representations? 

 
 
Central to philosophy of mind has been the mind-body problem: the question of how 

mind and body are (and can be) related and how they affect (and can affect) one 

another. Most contemporary authors favor physicalism —the thesis that only physical 

particulars exist, such that a physical duplicate of this world would be a duplicate 

simpliciter. Within physicalism, there is a debate about the ontological status of 

putative mental kinds: are they reducible to underlying physical ones, or on the 

contrary, do they constitute an autonomous domain? 

The position that mental kinds are reducible to physical ones is known as “reductive 

physicalism.” Two objections have traditionally been raised against reductive 

physicalism. First, the absence of bridge laws connecting the mental to the physical 

suggests that one domain is not reduceable to the other (Davidson, 1970). Second, 



 

 

8 
 

reductive physicalism entails that creatures with different brains cannot have the same 

mental kinds as we do. So, if mental states can be multiply realized, reductive 

physicalism is false (Putnam, 1967; Block & Fodor, 1972).  

Within non-reductive physicalist theories, a popular position is psycho-functionalism. 

Psycho-functionalism can be construed as the conjunction of two theses: that relevant 

ontological kinds are those posited by the best cognitive scientific theories, and that 

functionalism is true. Since the first tenet is self-explanatory, I will elaborate on the 

second. Functionalism is the view that what makes a mental state, event, process or 

property3 (for simplicity, I will just talk of “states” in what follows) the kind of 

mental state it is (e.g. pain) is its functional role —that is, its causal profile— in the 

system of which it is a part. An internal state of an individual is an instance of type of 

mental state if, given a certain input, it performs the relevant causal role in relation to 

other states of the nervous system, and is causally efficacious in contributing to the 

subsequent behavior of the organism that possesses it (Putnam, 1975). It is worth 

clarifying that mental states are not abstracta,4 according to the functionalist. When 

one is characterizing things by their functional role, one is not describing nonmaterial 

entities —but merely omitting5 certain implementational details from the 

 
3 Functionalism doesn’t have a commitment as to whether states, processes, events, properties, etc. are 
the correct kinds or metaphysical units in the mind/brain. The idea is the same regardless of 
metaphysical unit: what makes a process/event/state an instance of a given kind is the role it plays in 
the system in which it is part.  
4 Here I am referring to the distinction between concrete vs abstract things. Concrete things are those 
that are located in space and time, or —for an alternative characterization— that have causal powers 
(e.g. an atom, this book). In contrast, abstract stuff does not have spatiotemporal locations, and 
couldn’t possibly be physical. Others define abstracta by what lacks causal powers -i.e. can’t cause 
anything (e.g. the number 7; English) (see Rey, 1997; Falguera et al., 2022). 
5 Abstraction or omission of details to characterize a kind is pervasive in the natural sciences, and if 
properly conducted is not problematic.  



 

 

9 
 

characterization. Items with the property of having a certain functional role (kind 

instances) also have physical properties6. Although mental states are recognized, in 

part, by the behavior to which they are a contributing cause, they are not identical to 

that behavior. Likewise, a state may be caused by, and play, certain causal role(s), but 

it is not just these causal role(s). Functionalism entails multiple realizability—the 

view that a single mental kind can, in principle, be implemented or realized by 

multiple distinct physical kinds. According to functionalism, there may be a one-to-

many relations between mental and physical kinds. 

Although psycho-functionalism does not, in principle, entail that the mind is 

representational (that is, that it contains intentional items that are about or refer to 

things), in practice its adoption involves accepting that the mind contains 

representations. This is because many successful theories in cognitive science involve 

explanations and generalizations ranging over representational kinds. Historically, 

some of the most successful theories in cognitive science have been developed from 

classical approaches, which emphasize the manipulation of physical symbols or 

representations according to some rules. Even those deriving from connectionist and 

embodied-extended approaches are representational (Dawson, 2013; Calvo & 

Gomila, 2008).  

Yet, not everyone is onboard with granting the reality of functional and 

representational kinds. Some authors view theories involving representations in the 

 
6 Note that functionalism is compatible with interactionist dualism. However, most functionalists are 
physicalists (they think that there is nothing “over and above the physical”, and that a physical 
duplicate of this world would be a duplicate simpliciter of this world), so I will just talk about 
physicalist functionalism in what follows. 



 

 

10 
 

cognitive science as suspicious, or even worse, deficient. For example, Gualtiero 

Piccinini and Carl Craver (2011) argued that functional (and other non-mechanistic) 

explanations in cognitive science are a temporary “patch,” a product of our present 

ignorance of the underlying mechanisms operating in cognitive phenomena, which 

will inevitably dissipate as neuroscience becomes more and more integrated with 

psychology and other “higher-level” branches of cognitive science. Since Craver and 

Piccinini use functional kinds in their mechanistic models (e.g. “neurotransmitter”, 

“selection pressure”), but not representational ones, it is to be supposed that what they 

feel uneasy about is that such explanations involve representations.  

Carl Craver is one of the main articulators of the account of mechanisms that came to 

be known as “MDC”, according to which: 

Mechanisms are entities and activities organized such that they are productive of 

regular changes from start or set-up to finish or termination conditions (Machamer et 

al., 2000, p. 1). 

So, one might ask: does the MDC account require that the kinds involved in cognitive 

scientific explanations are neuronal, or otherwise physical? 

A positive answer would situate MDC in the reductive physicalist side of the mind-

body problem. To my knowledge, the only attempt to answer this question (Krickel, 

2019) did not arrive to a convincing conclusion for either view. So, in this paper, I 

amend (or maybe interpret) MDC to be compatible with representationalism. I do so 

by developing an account on which properties determine whether an individual is an 



 

 

11 
 

instance of certain working entity-kind, which I call the “ARI” account —for 

Activity-Enabling, Robust and Individuating properties.  

My paper situates the mechanistic account provided by MDC within the reductive vs 

non-reductive physicalism debate. Also importantly, it makes a contribution to the 

question of which properties determine working entity status. That is, what makes 

certain entities suitable to be part of a mechanism producing a phenomenon, and what 

excludes other entities from this categorization? In so doing, I answer Franklin-Hall’s 

(2016) carving problem: the question of how to carve a mechanism’s components in a 

way that is appropriate, non-arbitrary, and non-gerrymandered. 

 

3. A material girl in a normative world: an account of mental 
disorders 

 
Cognitive science is the discipline that systematically studies information-processing 

systems such as the human mind/brain. There is a related discipline that also studies 

the human mind/brain, but its goal is not merely to gain knowledge, but also to 

promote mental health: psychiatry. Psychiatry studies, classifies, and researches 

causes as well as possible venues for intervention in mental disorders. Yet, few 

taxonomical systems or kinds have had the assumption of realism questioned as 

vehemently as that of mental disorders has. 

What sort of things are mental disorders? This question can be decomposed into three 

more nuanced ones. The first one asks about the metaphysical nature of mental 

disorders: are they natural kinds, socially constructed kinds, broken normal kinds, 



 

 

12 
 

pragmatic kinds, etc.? The second asks about what makes them disorders, as opposed 

to non-pathological features of persons. The third asks what makes them mental 

disorders, as opposed to, say, neuronal or somatic disorders. My paper addresses the 

third question. 

What distinguishes the mental from the physical? In the philosophy of mind, there 

have been various attempts to find a feature or set of features that all mental states, 

processes, properties and so on have, and that all non-mental stuff lack. These 

feature(s) are sometimes called “the mark(s) of the mental” or criteria for mentality. 

The challenge is to provide such “mark(s)” in such a way that does not immediately 

presuppose the truth of a particular stance on the mind-body problem. There is still an 

ongoing debate about this, but two features have continued to be discussed since they 

were first proposed. First, phenomenal consciousness: it is the property which mental 

states have when it is like something to undergo them (Nagel, 1974). A problem for 

this proposal is that there are some mental states, such as standing beliefs or 

subliminal perceptions, that aren’t conscious. Second, intentionality: the property of 

being about something, of being directed at something, of standing for something 

(Brentano, 1874). A difficulty for this proposal is that some qualitative states (e.g. 

tickles, pains) are mental states but they don’t seem to be about anything (although 

representationalists have made a convincing case that those states also have content). 

Another difficulty is that it seems that some non-mental stuff can be intentional too: 

for instance, tree rings stand for a tree’s age, and this introduction is about my 

dissertation. There are also other proposals, such as direct access, incorrigibility, and 

transparency, but I won’t discuss them here.  



 

 

13 
 

 

The way psychiatrists have approached the mental, however, does not seem to be 

informed by those philosophical debates. Many conditions that cause distress and/or 

impairment affect phenomenal consciousness and intentionality, yet they are not 

mentioned in the Diagnostic and Statistical Manual of Mental Disorders (DSM) —

most notably, those affecting perception such as prosopagnosia, blindsight or 

phantom limb syndrome, but also some affecting other cognitive functions, such as 

Alzheimer’s, Lou Gehrig’s disease, or migraines. The notion of the “mental” they use 

isn’t characterized negatively either, that is, by capturing what does not have a known 

physical cause: Down syndrome is a mental disorder despite having a known physical 

cause (having a third copy of the 21st chromosome), and so is narcolepsy, which is 

caused by hypocretin deficiency.  

The inclusion and exclusion of disorders in the DSM follows a complicated collective 

decision-making process, where social, political, technologic, economic and 

pragmatic considerations play important roles. It is plausible that the disorders 

currently in the DSM simply don’t form a systematic, coherent hole, and that there 

are exceptions to any characterization of mental disorders —since there are no 

consistently applied criteria, metaphysic or epistemic, for something to be a mental 

disorder. Despite this, I think it is possible to provide a general characterization of 

mental disorders that captures the paradigmatic conditions listed in the DSM, drawing 

on a folk-psychological conception of the mind as containing reasons-responsive 

mechanisms. A mental disorder will thus be a significantly diminished (compared to a 



 

 

14 
 

non-disordered person) degree of the reasons-responsiveness of the mechanisms 

producing certain intentional states, actions and/or emotions. 

Some clarifications: I don’t take folk-psychology to be true (also, there isn’t a single 

folk psychology but folk-psychologies, since some mental concepts are acquired 

through socialization and acculturation). I consider the “mental”, so understood, as a 

generally useful fiction produced by the mind-reading system. If “the mental” vs 

“non-mental” distinction had some sort of interest-independent reality, we would 

probably fail to get the division right, given the large amount of evidence that we 

interpret, and to certain degree confabulate, the reasons behind our actions and those 

of others (Carruthers, 2011). I also hold the view that all mental phenomena are 

explainable causally. Moreover, this view of “the mental” entails that what falls upon 

its domain changes through time, to the extent that reasons-discourse and normative 

standards evolve alongside socio-cultural factors. 

In philosophy, reasons-responsiveness has received the most attention from ethics. In 

a Frankfurtian compatibilist spirit, Fischer and Ravizza (1998) have an account of 

moral responsibility according to which a person is morally responsible for her 

behavior if she has guidance control over it, that is, if the mechanism producing these 

behaviors is responsive to reasons. Whatever the merits of this view to capture moral 

responsibility, I use this as the basis for an account of what makes certain disorders 

mental (as opposed to neurological or somatic conditions).  

Many mental disorders can be characterized by such a lack of reasons-

responsiveness: the severity of mood disorders, for instance, depends on the degree to 



 

 

15 
 

which one’s mood is unresponsive to the things we consider provide reasons to 

change it. (Relatedly, an exclusionary factor for depression diagnoses is whether the 

person has an appropriate reason for experiencing its symptoms, such as the death of 

a spouse, in which case the persistent symptoms would not indicate a failure of 

reasons-responsiveness). A person with arachnophobia has an unfitting fear emotion 

to spiders that is resistant to sort of convincing that would modulate another person’s 

fear. Delusions are resistant to evidence. This doesn’t mean the mechanisms 

responsible for these states cannot be modulated: maybe they can, but venues for 

intervention would likely depend on causal, as opposed to reasons-giving, 

interventions (e.g. using antidepressants or cognitive behavioral therapy).  

The connection between “the mental” in psychiatry and “moral responsibility” in 

ethics also explains why it is plausible that mental illness at least mitigates moral 

responsibility. The debate as to whether people with mental disorders are morally 

responsible for their actions is a contested one in ethics (see, e.g. Pickard, 2011; King 

& May, 2018; Kozuch & McKenna, 2016) and in law (e.g. Elliott, 1996; Kalis & 

Meynen, 2014). My account of mental disorders makes sense of why these debates 

exist. 

Coming back to our original question: what can we say about cognitive kinds? My 

dissertation project tries to develop a piecemeal approach to what makes for a good 

kind in the cognitive sciences. The three papers comprising this dissertation aim, thus, 

to make contributions on this regard. In chapter 2, I distinguish between two sorts of 

mechanisms that should not be conflated in cognitive science. In addition, I provide a 



 

 

16 
 

“toolbox” model of mechanistic accounts of explanation: the adequacy of any 

particular account depends on both the target mechanism, as well as our current 

knowledge about it. In chapter 3, I vindicate the appropriateness of the New 

Mechanical Philosophy (and in particular, Machamer et al.’s 2000 account) to capture 

representational explanans, by providing a general account of how to carve working 

entities in the context of a mechanism. In chapter 4, I address a certain taxonomic 

kind —“mental disorders”— , with idiosyncratic characteristics that make them work 

differently than most other scientific kinds, and I provide an account of what guides 

the inclusion or exclusion of certain diagnostic categories in the psychiatric 

classification system. At the end of this dissertation, I hope, we will have reached a 

greater understanding of what cognitive mechanisms, representations and psychiatric 

disorders are.  

 
 

 

 

  



 

 

17 
 

 

Chapter 2. How is the mind/brain composed? Two 
sorts of mechanisms in cognitive science 
 
 
 
 
 
 
 
 

1. Introduction 

Cognitive science is a multidisciplinary scientific field; its domain of interest is 

cognition, broadly understood—how systems (especially human nervous systems) 

represent, store and process information. Cognitive science is not only interested in 

what our cognitive systems do, but also in how the relevant parts of the mind/brain 

make these things occur.  A way to shed light on the workings of a cognitive system 

like the human mind/brain is by decomposing it to see how the relevant parts produce 

the phenomenon of interest. When the goal is to produce general explanations, this 

decomposition often involves types (as opposed to particulars).  

In philosophy of cognitive science, two separate streams of literature that have taken 

on the project of answering the question “what should the mind/brain be decomposed 

into?” or “what are the relevant units when it comes to analyzing how cognitive 

phenomena are produced?”. These two traditions have proposed different blanket 

answers: one tradition emphasizes the notion of “modularity”. First introduced by 

Fodor (1983), this notion has been substantively revised in the hands of massive 

modularists (e.g. Cosmides & Tooby, 1992; Carruthers, 2006; Barrett, 2015).  



 

 

18 
 

Massive Modularists argued that the mind/brain is entirely composed by “modules” 

or "isolable function-specific processing systems" (Carruthers, 2006, p. 12). These 

systems are stable architectural components of the mind/brain, often characterized by 

appeal to domain specificity and a dedicated neural architecture (but without 

necessarily committing to the central component of Fodorian modularity, the 

encapsulation of modules). On the other hand, the New Mechanical Philosophy of 

science (after a widely-cited paper of Machamer et al., 2000, published some years 

after Glennan, 1996) proposes that many phenomena of the special sciences are the 

product of mechanisms. Some mechanists (e.g. Piccinini & Craver, 2011; Kaplan & 

Craver, 2011) have argued that all cognitive scientific phenomena are underlain by 

mechanisms-as- causings (or entities, activities and organization in causal continuity 

to produce the phenomenon). According to those mechanists, to explain cognition one 

ought to identify the relevant mechanism(s) involved. 

So, is it modules or mechanisms that we should be decomposing the mind/brain into? 

The dispute is not merely a verbal one, they are genuine alternatives.  Although one 

may be tempted to consider these accounts to capture two sides of the same coin, I 

argue that many times their targets are different in nature. What is described by each 

of those accounts is distinct and usually ought to not to be conflated with the other. At 

a minimum, there are metaphysical differences between the two: systems or modules 

are machine-like in that they continue to exist even when they aren't acting; while 

mechanisms-as- causings only exist while they occur.  



 

 

19 
 

This has implications for how we quantify and count cognitive components and for 

our treatment of deviant cases. For instance, consider John’s ability to recognize 

faces. Treating face recognition as a module implies that John’s face-recognition 

system is the same when he recognizes Mary and when he recognizes Thomas, and 

that same token module (m1) explains both these occurrences. On the other hand, 

from a mechanism-as-sort-of-causing standpoint, causes are only tokened while they 

occur, so each token instance of John’s recognizing a friend (Mary, Thomas, Mary at 

a later time) is produced by a different token mechanism (call them mc1, mc2, 

mc3 …), which explains a particular occurrence of recognizing. Token systems 

(token modules) thus persist over time while token mechanisms-as- causings do not. 

Moreover, were John to acquire prosopagnosia, which is the inability to recognize 

faces as a result of a brain injury, the modularist would talk about a “damaged 

module”  while the mechanist-as-a-sort-of-causing would talk about the different 

mechanisms that now underlie his response to facial stimuli7.  

This is not, however, the only difference to be found between mechanisms-as-systems 

and mechanisms-as- causes. The former are often universal among humans, with a 

distinct developmental path, a dedicated brain network, and a distinct proprietary 

function, while for the latter we may expect the involvement of distinct (and not so 

closely related) networks, greater variability in their starting conditions and/or parts 

involved, and invariance with respect to the production of the phenomena.  

 
7 The mechanisms will differ by belonging to different mechanism-types. The two different 
mechanism-types may be close together in similarity space, in the sense of having a good number of 
components in common, but they are still different because they produce different phenomena. 



 

 

20 
 

I argue that both approaches to mechanisms properly capture part of the functioning 

of mind/brain. I claim (section 2) that cognition involves modules or mechanisms-as-

systems, which I illustrate by presenting the face recognition system in section 3. I 

argue that cognition also involves mechanisms-as- causings in section 4, which I 

illustrate by presenting a mechanism for placebo analgesia in section 5—a 

mechanism which, like many mechanisms-as-causings, doesn’t have clear spatio-

temporal boundaries nor dedicated processing structures. I address the relationship 

between both sorts of approaches in section 6. However, this raises questions about 

how to treat research in progress (section 7). Most research involves fragmentary 

sketchy models, that include some but not all the parts of the target mechanism. If 

both sorts of mechanisms (system-like and cause-like) exist in cognitive science, how 

do researchers decide  into which category their proposed mechanism fits? How 

should they be treating their target system in the meantime?8 

This leads us to the last section of the paper. I argue that we ought to understand 

hypotheses or proposals in the early stages as non-committal as to whether the 

phenomenon is underlain by a mechanism-as-system or mechanism-as-sort-of-

causing. A good approach is to make use of Glennan and Illari’s “minimal notion of 

mechanism” (2017) as a tool to start thinking about the target mechanism and 

incorporate the different findings. Their minimal notion has the advantage of being 

neutral with respect to the metaphysical commitments made by the other two 

accounts. I argue that at early hypothesis-generating stages, researchers are not saying 

 
8 This is important methodologically, as it would determine the methods researchers should choose to 
study the mechanisms of interest.  



 

 

21 
 

anything metaphysically “thick” about what is behind the phenomenon in question. It 

would be an error to assume, for instance, that every time evolutionary psychologists 

hypothesize that there is a “system” behind the observed phenomena of cheater 

detection, that what they have in mind a stable functional component or “system” in 

Carruthers’ (2006) or Bechtel’s (2008b) sense. Even if that ended up being the case, it 

wouldn’t be epistemically justifiable to make that assumption until there is more 

evidence for the claim.  

 

2. Modules or mechanisms-as-systems 
 

Cognitive scientists often try to describe mental mechanisms, but they aren’t always 

explicit about how they use myriad terms like “mechanism”, “module”, or “system”. 

There seem to be at least two different ways in which they do so: very roughly, 

sometimes they treat them as systems (as in memory systems, language module, face 

recognition system, the visual system, etc.); other times, they mean the sort of causing 

that explains how something comes to be the case9 (here, we could find the sequence 

of causally interacting parts that gives rise to instances of the decoy effect, McGurk 

effect, Thatcher effect or change blindness). Although the two senses are intuitively 

related (or may even overlap when a system is performing its proper function), they 

 
9 “How” in a causal sense, not in an evolutionary one. 



 

 

22 
 

capture different aspects of reality. In this section, I will discuss the first sense, i.e. the 

notion of mechanism-as-system or module. 

When discussing face perception, Nancy Kanwisher wondered:  

“Is face perception carried out by domain-specific mechanisms, that is, by modules 

specialized for processing faces in particular? Or are faces handled by domain-

general mechanisms that can operate on nonface visual stimuli as well?” (Kanwisher, 

2000, p. 759) 

Like her, many cognitive scientists have discussed the units composing the 

mind/brain to be mechanisms in the “system” or “module” sense: 

"An association may be found between tasks X and Y because the mechanisms on 

which they depend are adjacent in the brain rather than because they depend on the 

same underlying mechanism. Gerstmann’s syndrome is an example. It is defined by 

four very different symptoms: problems of finger identification; problems in 

calculation; impaired spelling; and left–right disorientation. It is improbable that the 

same mechanisms or modules are involved in all four tasks. What is much more 

likely is that these four symptoms depend on different mechanisms that happen to be 

anatomically adjacent in the brain.” (Eysenck & Keane, 2015, p. 20) 

As these quotations illustrate, mechanisms are sometimes treated as stable systems 

composing the brain, with a specific function, which perdure even when they aren’t 

acting, and may be (perhaps loosely) localized.  



 

 

23 
 

This sense of modularity is implicit in Fodor’s original formulation in his book 

Modularity of Mind (1983), and is the common ground between Fodor’s view and 

those of other modularists who later disputed his characterization and provided their 

own (e.g. Carruthers, 2006; Coltheart, 1999; Barrett & Kurzban, 2006; Cosmides & 

Tooby, 1992; Pinker, 1997).  

Fodor took modules to be systems characterized by being informationally 

encapsulated (that is, that during processing modules do not share, and cannot be 

affected by, information held anywhere else in the mind), as well as by exhibiting 

eight other characteristics10. According to him, modules are restricted to systems at 

the “periphery” of the mind —those that deal with perception, language and action. 

Many after him have argued that the entirety of the mind/brain is modular, a thesis 

known as “massive modularity”. Most notably, Peter Carruthers (2006) argued that 

comparative psychology alongside evolutionary considerations make the most 

plausible case for the massive modularity of mind hypothesis. For massive modularity 

to be possible, modules cannot be informationally encapsulated —and unsurprisingly, 

this is the feature of Fodor’s list that has received the most pushback. Other features 

 
10 Here is the full list of characteristics Fodor (1983, p. 37) proposed a module had to have “to some 
interesting extent” (meaning to an appreciable degree): (1) Domain specificity: modules are restricted 
to certain kind of inputs.(2) Mandatory operation: modules operate in a mandatory (or automatic) way. 
(3) Limited central accessibility: the operations and representations occurring within a module are not 
accessible (or accessible only in a very limited way) to higher cognitive processes. (4) Fast processing: 
modules generate outputs quickly.(5) Informational encapsulation: during its processing modules do 
not share, and cannot be affected by, information held anywhere else in the mind. (6) ‘Shallow’ 
outputs: the outputs generated are basic, simple. (7) Fixed neural architecture: modules are realized in 
a dedicated neuronal architecture. (8) Characteristic and specific breakdown patterns: modules can be 
selectively damaged, allowing for phenomena like double dissociations. (9) Characteristic ontogenetic 
pace and sequencing: modules are hard-wired in the brain and have a characteristic development. 
Among the features of this list, Fodor took information encapsulation to be a module’s most important 
feature (Fodor, 2000), probably because he thought (mistakenly) it would solve the frame problem—
the problem of how to restrict computations to what is relevant, “given that relevance is holistic, open-
ended, and context-sensitive” (Shanahan, 2016). 



 

 

24 
 

of the list also got disputed: content domain specificity, strong localizability, shallow 

outputs, innateness, fast processing and automaticity (see e.g. Barrett & Kurzban, 

2006; Carruthers, 2006; Coltheart, 1999). Nowadays, cognitive scientists largely 

operate outside Fodor’s assumptions about the extent and characteristics of modules. 

The important point for our purposes is that the family of modular views of the mind 

all share a common notion of module: modules are stable systems composing the 

brain, with a specific function, which perdure even when they aren’t acting, and may 

be (perhaps loosely) localized. This basic characterization is a minimum common 

denominator among modular accounts, to which different authors add further 

characteristics11. This is also the common denominator among a certain family of 

accounts of mechanisms, as I will discuss next. 

Daniel Nicholson’ describes a family of theories of mechanisms that he calls 

“machine mechanisms”: “systems conceived in mechanical terms; that is, as stable 

assemblies of interacting parts arranged in such a way that their combined operation 

results in predetermined outcomes” (Nicholson, 2012, p. 153). Beate Krickel (2019) 

labels it the “Complex Systems Approach” to mechanisms: accounts that “speak of 

mechanisms in terms of stable arrangements, structures, or objects” and “highlight the 

machine analogy in arguing that mechanisms are like machines” (Krickel, 2019, p. 

 
11 For example, Carruthers (2006) considers the following characteristics help us identify modules: 
modules have a function, something they are supposed to do with the information they receive. They 
are domain-specific. Modules are usually associated with particular areas of the brain, although they 
may be using a set of neural pathways that may be scattered in multiple brain regions. The processing 
of information that occurs in modules is largely independent of information stored elsewhere. Modules 
are entity-like in that they can be selectively damaged, allowing for double dissociations. Modules 
operate automatically, not at will. Most importantly, the mind/brain contains several modules; those 
are treated as its architecture. 



 

 

25 
 

22). These will include early accounts of mechanisms by Stuart Glennan (1996, 2002, 

2010), William Bechtel (2008a, 2008b), Bechtel and Robert C. Richardson (1993), 

and Bechtel and Adele Abrahamsen (2005).   

As I will spell out below, there are some commonalities between Complex Systems 

Approaches to mechanisms and accounts of modules, not only in their minimum 

common denominators discussed above, but also among other putative characteristics 

in maximally articulated accounts (e.g. Carruthers 2006 among modular accounts, 

Glennan 2002 and Bechtel 2008b among Complex System Approaches). The 

common features in these approaches make up what I will call “the mechanisms-as-

systems” view.  

The targets for both, Complex Systems and modular accounts, are stable entity-like 

systems that persist. For instance, Glennan talks about mechanisms as stable 

arrangements of parts that have dispositions: “Perhaps the most notable difference 

between the complex-systems and Salmon/Railton approach is that Salmon/Railton 

mechanisms are sequences of interconnected events while complex-systems 

mechanisms are things (or objects)" (Glennan 2002, S345, emphasis in original). 

Bechtel and Abrahamsen say that "[a] mechanism is a structure performing a function 

in virtue of its component parts, component operations, and their organization" 

(Bechtel & Abrahamsen 2005, p. 423). I will start first by discussing the two features 

that come most to mind when discussing systems as object-like things: their spatial 



 

 

26 
 

localization, on the one hand, and their “function” or “predetermined outcomes”12, on 

the other. 

Mechanism-as-systems or modules are often associated with a particular part of the 

brain. Modularists typically speak of this as a module’s “dedicated neural 

implementation”, but that dedication isn’t always exclusive. We should interpret 

identifications of a module with certain brain areas as an heuristic, not something to 

be taken literally. Bechtel describes it in the following way:  “A common first step in 

such research is to attribute to a part of the system that produces a given phenomenon 

full responsibility for that phenomenon (in Bechtel & Richardson, 1993, we referred 

to this as direct localization). In the case of vision, this involved treating an area of 

the brain as the visual center. Such attributions of complex activities to single 

components of a system seldom turn out to be correct. As heuristic identity claims, 

however, they often contribute to productive research by facilitating the discovery 

that in fact the area is associated with only one or a few of the component operations 

required to produce the phenomenon.” (Bechtel, 2008b, p. 89). In other words: spatial 

localization is a rough heuristic that helps to identify part of the implementation of a 

module —by looking at what areas selectively activate when performing a certain 

function. But more often than not, a module’s functioning will require an integrated 

network of brain areas working together, sometimes including (sub-)modules 

scattered across the brain. For instance, face recognition comprises a brain network 

not only involving left and right fusiform face areas (which seem to be involved in 

 
12 The former is the term Bechtel and Carruthers use; the latter, Nicholson’s. 



 

 

27 
 

discrimination of faces from non-social stimuli; Lopatina et al., 2018), but also 

occipital face areas (more involved in detailed recognition; Atkinson & Adolphs, 

2011), the medial temporal lobe (specially the perirhinal cortex for familiarity-based 

recognition; Eichenbaum et al., 2007), and probably others. For another example, the 

affective system involves brain circuits that include nearly the entire brain: positive 

valence is processed in a network that involves subcortical regions of the ventral 

striatum linked with regions of orbitofrontal and ventromedial prefrontal cortex, 

while negative valence is processed in the amygdala and subcortical regions of the 

ventral striatum linked with the anterior insula, and cortically within the anterior 

insula and anterior cingulate (Berridge & Kringelbach, 2013; Yarkoni et al., 2011; 

Grabenhorst & Rolls, 2011). The point illustrated here is that the spatial localization 

heuristic is an over-idealization of a module’s implementation: except for very simple 

and low-level sensory and motor systems, we can expect modules to comprise 

integrated networks of what may be different brain areas with specific sub-functions, 

which comprise a stable perduring system.  

Modules or mechanisms-as-systems also have a proper function. This proper function 

is to be distinguished from things a token “merely does”, does “by accident”, or does 

when “malfunctioning”. For instance, a clock is a mechanism-as-system for time-

tracking, and an air conditioner, for cooling air. Even if they both take up space in 

one’s apartment, that (taking space) is not their (proprietary) function, just something 

they do. The notion of proprietary function is not that of “mere causal role” employed 

by those who articulate mechanisms-as- causings: it is historical, and has an at least 

minimal normative dimension. There is something the module is supposed to do with 



 

 

28 
 

the information it receives, there is a set of phenomena mechanisms of that type are 

supposed to produce. While I do not want to tie modularity claims to a specific 

account of proper function, I will point out that the account must be one of the 

teleological family that can be traced back to Larry Wright.  In Wright’s 

characterization, the function of X is Z means (a) X is there because it does Z, and (b) 

Z is a consequence (or result) of X's being there (Wright 1976, p. 81). For Ruth 

Millikan (1984, 1989), Z is the proper function of X iff Z has had certain effects on 

how the tokens of X were “copied” or reproduced. These effects on the ancestors of a 

given token of X “have helped account for the survival, by continued reproduction, of 

the item's lineage” (Millikan, 2002, p. 8). For instance, hearts have been selected for 

by Darwinian natural selection (given that they did pump blood in “historically 

normal conditions”), and reproduced genetically. These are “proper biofunctions”. In 

addition, the proper function of some modules may be a “secondary adaptation” or 

exaptation, if their preservation in recent evolutionary past is due to their doing G and 

not their F-ing they were originally selected for (Griffiths, 1993). 

However, not all modules are adaptations, and thus not all functions of modules are 

grounded in natural selection: some modules acquire their function via learning 

mechanisms —for instance, those of the exact number system and the print reading 

system. Recently, Justin Garson (2017, 2019) has provided a historical account of 

function (which he calls the General Selected Effects, or GSE) that captures these two 

ways in which a mechanism comes to have a proper function. According to GSE, the 

proper function(s) of a mechanism is the activity that historically “contributed to its 

bearer’s differential reproduction, or differential retention, within a population” 



 

 

29 
 

(Garson, 2017, p. 523). The “or” here is disjunctive: if a selection process has taken 

place resulting in differential reproduction or retention of a function, that is enough to 

make it a proper function —it doesn’t matter if that selective process was natural 

selection, learning or competitive retention. For instance, suppose a learning process 

results in a neuronal disposition to Z to be retained, while another neuronal 

disposition to Y is eliminated, given a competitive process that takes place between 

them —a “zero sum game” (Garson, 2017, p. 532). In that case, Z-ing is the proper 

function of that mechanism.  Something like this seems to be the process by which 

the print reading module is acquired: the left fusiform gyrus specializes in reading 

over the course of learning, resulting in the area known as “visual word form area” 

(VWFA) to acquire the proper function of recognizing letters and printed words, at 

the expense of conducting other tasks such as face processing. Supporting this 

hypothesis, Dehaene et al. (2010) and He et al. (2009) found that left fusiform 

responds to faces in illiterate adults, but such sensitivity is reduced when they learn to 

read. Among child beginner readers, Centanni et al. (2018) found that the greater the 

size of letter-sensitive cortex in left fusiform, the smaller the left face fusiform area 

(left FFA); and the greater the sensitivity of left fusiform to letters, the better the kid’s 

reading ability.  

Nicholas Shea (2018) has a similar account of proper function (which he calls “task 

function”). Roughly, an output F is a proper function of S if F is (i) a robust outcome 

function of S; and (ii) a stabilized function of S. An output F is a robust outcome 

function of S iff S produces F in response to a range of different inputs, and in a range 

of relevant external conditions (Shea, 2018, p. 55). An output F is a stabilized 



 

 

30 
 

function of S iff producing F has been systematically stabilized in at least one of the 

following manners:  “(i) by contributing directly to the evolutionary success of 

systems S producing F; or (ii) by contributing through learning to S’s disposition to 

produce F; or (iii) where S is an organism, by contributing directly to the persistence 

of S” (Shea, 2018, p. 64). The idea here is similar to Garson’s: proper functions are 

grounded in processes that historically select Ss because they do F.  

Like Garson and Shea, I contend that the processes of differential learning or 

selective retention are proper function-endowing. I suspect a majority of modules of 

the mind have acquired proper functions in this way: they start as partially innately-

specified learning systems that become elaborated and built through domain-specific 

learning, a process that endows them with specialized proper functions. 

The important point for our purposes is that the notion of “function” employed in 

mechanisms-as-systems is a teleological one, and not that of “mere causal role” 

employed by those who defend mechanisms-as-causings. This is true both for 

modularist authors as well as those who come from the mechanisms-as-systems 

traditions. For instance, William Bechtel and Oron Shaghir argue that, in order to 

describe a system computationally (in Marr’s 1982 sense), one needs to specify what 

a system’s function is, i.e. what it is meant to do, as well as its proprietary domain 

(see Bechtel & Shaghir, 2015). 

 



 

 

31 
 

Cognitive systems have proprietary domains; they are only sensitive to information of 

a certain sort. The “sort” of information a module is sensitive to is set by the 

properties of the inputs that reliably start a module’s performance of its function, and 

not whether things in the world belong to a “sort”. For instance, the proprietary 

domain of the face-recognition system are roughly face-like things, regardless of 

whether, in the world, these naturally conform to a kind. The generality or specificity 

of a proprietary domain doesn’t have anything to do with how many instances of the 

thing there are, either. It is not the case that the proprietary domain of the face 

recognition system keeps expanding as more and more people populate the world.  

It is important to emphasize that a module’s sensitivity to its proprietary domain of 

information is due to its physical properties13, and exempts the module from being 

actively working all the time. As an illustration, consider a photovoltaic cell. This cell 

is sensitive only to photons touching its surface, because of its physical-chemical 

properties and those of photons. Similarly, a module’s physical properties may make 

it sensitive only to representations with properties of a certain sort.  

That mechanisms-as-systems are stable and entity-like is also relevant in that it 

becomes possible to talk about a system as having a distinct development path, or 

being selectively damaged (thus allowing for double dissociations, i.e. cases when 

you have two related functions, X and Y, and after brain damage one subject can 

perform function X but not Y, and another subject can perform Y but not X; see 

 
13 “Physical” as opposed to “mental” properties, not the properties of physics. See Shaghir and Bechtel 
(2017) and Bechtel and Shaghir (2015) for discussion of the role physical or contextual constraints 
play in perception and computational tractability.  



 

 

32 
 

Shallice 1988). Carruthers (2006) took findings of double dissociation and/or a 

distinct development path to provide (non-conclusive) evidence for the existence of 

the module. We can get to know more about the proprietary domain of the module in 

question also by manipulating the input and then check whether the specific function 

is produced (that is, by manipulation; Woodward, 2003). Among the New 

Mechanists, Bechtel also discusses lesion research and double dissociations as 

providing prima facie evidence that the brain area in question is part of a certain 

mechanism-as-system. If damaging a certain brain area results in impaired 

functioning, then that would seem to suggest that the brain area in question was 

involved in the mechanism’s operations (Bechtel, 2008b, p. 43). One thing to note is 

that lesion research can only provide evidence that something is a component of a 

system if we presume that the identity of the system is preserved despite the damage; 

that is, if we treat the system as by default perduring in time despite the fact that, in 

cases of damage, it may be missing some components and/or its functioning may be 

impaired. 

 The persistence of modules also makes them subject to selective pressures. An 

argument that is sometimes provided for the modularity of the mind is that modules 

are the best candidates to be units of selection for psychological traits. The idea here 

is that complex systems (like biological ones) need to be organized in a pervasively 

modular way, that is, as a hierarchical assembly of separately modifiable, functionally 

autonomous components, for the system to be constructed incrementally. Since the 

human mind/brain is a complex biological system that has evolved incrementally 

from animal mind/brains, it is plausible that it is modular (Carruthers, 2006). 



 

 

33 
 

In summary, mechanisms-as-systems or modules are stable systems composing the 

brain, with specific functions, which perdure even when they aren’t acting, and may 

be loosely localized.  Properties such as having a distinct developmental path, giving 

rise to distinct phenomena, having proprietary parts and operations, being 

distinctively sensitive to certain domain of information, and being dissociable add to 

the evidence for the mechanism in question being a system.  

While there are some discussions as to whether modules are peripheral (meaning, 

only in perceptual and motor systems) or everywhere in the mind (including for 

higher cognitive functions), what seems clear is that cognitive scientists have 

provided numerous examples of mechanisms-as-systems: for example, memory 

systems, the language module, the visual system, and face recognition. In the next 

section, I will discuss the latter as an example of a mechanism-as-system. 

 

3. An illustration of a mechanism-as-system: face recognition 
 
 
 
When we talk about face recognition, we are referring to the identification of 

someone previously met using his/her face as the input. We are not talking about the 

recognition of a face as a face, as compared to another kind of object, nor of the 

recognition of facial expressions nor lip speech. The face recognition system is a 

mechanism-as-system which functions to recognize faces. Its proprietary inputs are 

face-like stimuli, and its output is a sometimes described as a sense of familiarity or 

recognition when we see a face of someone we know. 



 

 

34 
 

The most widely accepted model of face processing, by Bruce and Young (1986) 

separates the distinct processes that take place after encountering a face-like stimulus. 

Briefly: first, different codes are extracted from the input: pictorial codes (imagistic) 

and structural codes (more abstract). Those structural codes are then analyzed in order 

to obtain visually derived semantic codes (gender, age, personality), and identity-

specific semantic codes (meaningfulness of a face). If the face is represented as 

familiar, the Face Recognition Unit (FRU) specific to that face becomes activated, 

producing a feeling of familiarity. This sense of familiarity takes place prior to the 

retrieval of any information about that person. Face-identity recognition ends, thus, 

with the activation of the FRU (or the lack of it) when presented with the stimulus 

face. Posteriorly, the FRU serves as a gateway to the information about that person, 

via stimulation of the Person Identity Node (PIN). In turn, that can activate the related 

Semantic Information Unit (containing information about the face’s owner), which in 

turn can stimulate the Name Unit (containing that person’s name). Starting from the 

early visual analysis and until the activation of a PIN, face-identity is processed and 

recognized independently of expression and facial-speech/lip-reading analysis.  

A variety of evidence jointly supports the claim that face recognition comprises a 

mechanism-as-system or module. For example, there are developmental impairments 

of face recognition as well as (non-pure) double dissociations. For double 

dissociations, we have some cases in which brain damage has caused severely 

impaired facial recognition while recognition of other objects is around control levels 

(acquired prosopagnosia), while other patients exhibit severely impaired object 

recognition and normal levels of face recognition (object agnosia). There are also 



 

 

35 
 

documented cases of developmental prosopagnosia (see Bate, 2017, for an overview), 

which appears to be ‘congenital’ or ‘hereditary’. Selective impairments of the face 

recognition system are better explained by assuming the functional dissociability of 

this system from others. 

There is also evidence that face recognition has an ontogenic developmental path of 

its own. Babies exhibit a visual preference for faces (Valenza et al., 1996), which can 

be explained by both an innate facial module, and to what are good stimuli for 

immature visual systems (in terms of contrast, spatial frequency, top-heavy layout, 

etc.; ). Face recognition develops and improves until one reaches approximately 10 

years of age, becoming perceptually narrower (e.g. the other-race effect appears 

during this development). This effect is explained by both the face-specific module 

hypothesis and the expertise hypothesis (Hole & Bourne 2001). 

In addition, there are face processing-specific computational simulations. Several 

computational systems have successfully emulated face-identity processing while 

exhibiting the same errors patterns humans do. Most are based on neural networks 

(e.g., Lawrence et al., 1997).  

Finally, facial recognition involves the activation of brain regions different from non-

facial recognition14.  Neuroimaging studies have related it mainly to the fusiform face 

area (FFA) and the occipital face areas, which are regions that don’t seem to have 

 
14 Maybe with the exception of the areas for visual expertise. There is some debate as to whether 
the FFA is specific to faces (but gets recruited for visual expertise in object recognition), to visual 
expertise (but gets recruited for face recognition), or both. See e.g. Tarr and Gauthier (2000), 
Gauthier et al. (2003), Kanwisher (2000). 



 

 

36 
 

other unrelated functions. But the module is likely to comprise a broader network of 

brain regions, including the hippocampus and the perirhinal cortex in the medial 

temporal lobe, some patches of the anterior inferior temporal cortex, and the 

amydgala (Lopatina et al., 2018).  All this evidence supports the claim that face 

recognition is a mechanism-as-a-system or module. 

 

4. Mechanisms-as- causings 
 
 
So far, I have defended the position that mind/brain contains some mechanisms-as-

systems. However, there isn't an entity-like system behind every cognitive 

phenomenon15. Sometimes, researchers individuate mechanisms of interest on the 

basis of how a phenomenon is causally produced—regardless of whether it forms a 

"system" or not. In this second sense, a cognitive mechanism is a specific "sort of 

causing":  

“Researchers […]  specify a mechanism for the causal relationship and combine the 

results from a variety of research questions.” (Morling, 2017, p. 266, my emphasis) 

 Or: 

 
15 Doing so would be committing a version of what Bedford (1997) calls the "Not-The-Liver Fallacy," 
i.e. assuming that there must be a system responsible for every dissociable function of the organism. 



 

 

37 
 

“[T]here are many questions of mechanism, such as How is object recognition 

accomplished by the visual system?, that the lesion method, alone, is ill-suited to 

address.”  (Polsner, 2015, p. 56) 

This is the second, mechanisms-as -causings notion. Explanations invoking 

mechanisms in this second sense address questions of “how does X occur?” by 

appealing to repeatable causal chains producing X. Here, mechanisms are 

individuated solely on the basis what happens to produce a certain phenomenon—

they don’t need to have natural boundaries or form a stable system16.  

This second notion is captured by  Machamer, Darden and Craver’s account of 

mechanisms-as- causings: “entities and activities organized such that they are 

productive of regular changes from start or set-up to finish or termination conditions” 

(Machamer et al. 2000, p. 1; henceforth, “MDC”). In the rest of this section, this is 

the account I will be discussing because it is the most articulated, although it is not 

the only account of mechanisms-as- causings. The accounts of Illari and Williamson 

(1992) and recently Glennan (2018) also capture mechanisms in this sense, as what 

acts to produce a phenomenon. Krickel calls these “Acting Entities Approaches”: they 

“assume that mechanisms are not objects but process-like, in the sense that they 

consist of actual manifestations of causal activities of various entities that interact” 

(Krickel, 2019, p. 25). Nicholson calls this the “causal mechanism” sense: “A step-

 
16 According to Glennan (2007), there can even be one-off mechanisms. 



 

 

38 
 

by-step explanation of the mode of operation of a causal process that gives rise to a 

phenomenon of interest” (Nicholson, 2012, p. 153). 

According to MDC, mechanisms have productive continuity leading to a 

phenomenon. Mechanisms themselves are composed of activities, entities and their 

spatio-temporal organization within the context of a mechanism-as-causing. Entities 

(i.e. the things that engage in activities17 which have a robustness apart from their 

place in a mechanism) and activities (i.e. producers of change, specific “causings”) 

are components of mechanisms (i.e. what produces or underlies a phenomenon), and 

they cannot be reduced to one another. MDC is explicitly and irremediably dualistic: 

“both entities and activities constitute mechanisms. There are no activities without 

entities, and entities do not do anything without activities.” (MDC, 2000, p. 8). 

We saw that modules or mechanisms-as-systems perdure, even when they are not 

acting.  Craver and Tabery make very clear that mechanisms-as-causings do not do 

so: “[e]ntities (or objects) are not mechanisms. Mechanisms do things. If an object is 

not doing anything (i.e., if there is no phenomenon), then it is not a mechanism” 

(Craver & Tabery, 2019). That a mechanism is only such when it’s producing a 

phenomenon doesn’t mean that components of a mechanism cannot persist even 

when they aren’t acting. Consider the mechanism of the action potential in neurons 

discussed in MDC 2000. The potassium channels, the axon, all of these continue to be 

 
17 It has been said that MDC individuates entities on the basis of their physical properties; that entities 
are concrete physical objects that exist in space and time (Krickel, 2019, p. 117). One might thus think 
that such account might have trouble accommodating explanations couched in intentional terms, but in 
fact it might turn out that dropping the requirement that entities are individuated on the basis of 
physical properties is a minor extension of the theory.  



 

 

39 
 

there, having “a kind of robustness and reality apart from their place within that 

mechanism” (Glennan, 1996, p. 53). 

In summary: mechanistic explanation tells us how things work, by identifying 

entities, activities and their spatio-temporal organization, from initial or set-up 

conditions to finish or termination conditions. Mechanisms-as- causings exist insofar 

as they are acting to produce phenomena. 

While these accounts were originally developed for other scientific disciplines, what 

seems clear is that cognitive scientists have provided numerous examples of 

mechanisms-as- causings: for example, the sequence of causally interacting parts that 

gives rise to instances of the decoy effect, McGurk effect, Thatcher effect, change 

blindness, or the well-known placebo effect. In the next section, I will discuss the 

latter as an example of a mechanism-as-a-sort-of-causing. 

 

5. An illustration of a mechanism-as-a-sort-of-causing: placebo 
analgesia 

 
 
 
The “placebo effect” is an illustration of the sort of case that captures nicely a 

mechanism-as-a-sort-of-causing. Placebos are sham treatments that are provided to 

patients, who believe them to be clinically effective. Placebos are well known to 

improve health outcomes: they account for as much as 50% of the effectiveness of 

analgesics, and they are also known to improve the immune response, motor 

outcomes in the case of Parkinson disease, depression, and so on (Barrett et al., 



 

 

40 
 

2006). The phenomenon of interest is the improvement in the patient’s condition 

following the administration of placebo.  

Placebo effects are so robust that when a new medical intervention (for a condition 

for which there isn’t a standard of care) is tested in a two armed, randomized clinical 

trial, a placebo is always given to the control group. Importantly, the improvement 

patients experience with the placebo cannot be attributed to the medically inert 

substance or intervention they receive, nor, in double blind studies, to expectations of 

the clinical staff. Instead, it results from the patient’s anticipation that the intervention 

will help (Wager & Atlas, 2015). In other words, the improvement in patients’ 

conditions when given placebos is not caused by the physical or chemical properties 

of the sham substance. Ingesting such a substance in another context, for instance, 

one in which the patient knows s/he is taking a sugar pill, has no effect. The placebo 

effect is caused by a mind/brain mechanism that has as a component the belief or 

expectation that the substance received will help.  

There are a multiplicity of mechanisms that may come under the label of “placebo 

effects”. Among these, the mechanisms behind placebo analgesia are among the most 

researched (Kong et al., 2007). In this section I will introduce some of the several 

mechanisms posited to underlie such pain relief. 

Let’s start by taking a step back and looking at normal pain processing in people 

without prior expectations. The brain contains several pain-responsive regions, some 

of which process the sensory components of pain (including the posterior insula, SS1 

and SS2), others its badness (e,g, the anterior insula, thalamus and the anterior 



 

 

41 
 

cingulate cortex).  When signals from across the body transporting information about 

noxious stimuli arrive, these areas are aroused, and the pain-responsive regions 

contribute to the generation of feelings of pain and the derived motor responses. 

Some even talk about a “neurologic pain signature” using fMRI on the basis of 

patterns of activation of those brain regions (see Wager et al., 2013; Reddan & 

Wager, 2018) 

Placebos work by causing a reduction of the activity in these regions. First, the person 

needs to be given the placebo intervention in a way that mimics the treatment ritual 

(Sanders et al., 2020). The contextual information in the ritual —whether it is via 

verbal suggestions, retrieval of previous therapeutic experiences in similar settings or 

with similar interventions, observations of others getting treated, etc.; see Colloca 

(2019)— generates expectations of pain reduction. When someone has been primed 

to expect an improvement of their health state, their brain patterns are different from 

those of someone who is in a neutral state. Believing and/or expecting to receive an 

improvement in health leads to an increase in the activation certain areas of the 

prefrontal cortex and nucleus accumbens, which are believed to be responsible for 

appraisals of context and its meaning. Their activation leads to several changes. First, 

a partial suppression of the pain signals coming from the medulla carrying 

information about the type and intensity of the pain. The amplitude of event-related 

potentials produced in response to the painful stimuli is reduced, resulting in a 

decrease in experienced pain level (Wager & Atlas, 2015). 

 



 

 

42 
 

In addition, the activation of those areas in the prefrontal cortex results in the 

production and release of endorphins, a type of endogenous peptides that bind chiefly 

to opiate receptors in the brain, which results in a decrease in experienced pain. That 

some modulatory effects are mediated by endogenous opioids is well-stablished since 

a classic study (Levine et al., 1978) blocked placebo anesthesia effects by 

administering patients naxolone, an opioid receptor antagonist. Other neurochemicals 

mediating modulatory effects are dopamine, endocannabinoids, cholecystokinin 

(CCK) or serotonin, which also open possibilities for intervention (Frisaldi et al., 

2020). 

There are good reasons to believe that the analgesic effect of a placebo is a 

mechanism in the sort-of-causing sense. The placebo effect (and its sibling, the 

nocebo effect) seem to be a by-product of the more general evaluative learning 

system, which modulates expectations of goodness or badness based on previous 

cues. The formation of placebo responses has been accounted for as a general form of 

reward learning (de la Fuente-Fernández, 2009). It seems that the evaluative learning 

system is an adaptation, which at least in its associative (as opposed to cognitive) 

form we share with other animals. There is evidence of placebo effects mediated by 

Pavlovian conditioning occurring in e.g. rats (Herrstein, 1962). As I discussed in 

section 2, particular applications of a module’s function do not qualify as proper 

functions themselves, unless they have been selected for in another way —e.g. via 

learning processes. But the learning involved in placebo is not selection of a function 

(i.e. it is not preceded by selective retention of a specific function to be performed in 

contexts related to pain and medical settings), but rather just another learning of cues 



 

 

43 
 

and their likely values — something to be expected given that module’s proper 

function.  

Consistent with this, the interventions that can be made on the effect size of placebos 

have to do with modification of the cues or their linked rewards, not with placebos 

qua such. For instance, both the effect size and the duration of a placebo response are 

often equivalent to the active treatment being studied (Tuttle et al., 2015). This seems 

to conform to a Woodwardian intervention, in that changing the magnitude of a 

component of the mechanism produces effects on the magnitude of the phenomenon 

(Woodward, 1997). As mentioned, it is also possible to block analgesic effects by 

administering the relevant receptor antagonists, which also have been found to reduce 

reward-learning based on e.g. food (Galaj & Ranaldi, 2021).  

 

6. The relationship between mechanisms-as-systems and as-
causings 

 
 
 
Thus far, I argued that the mind-brain involves at least both mechanisms-as-systems 

or “modules”, and mechanisms-as- causings. A complete cognitive science would 

probably involve modelling of both “sorts”18. Despite the superficial similarities, the 

two notions involve different metaphysical commitments. Modules have a continued 

 
18 I don’t want to exclude in principle that there may be more “sorts” of mechanisms 
besides these two. Glennan (2017), for instance, allows for “one-off” mechanisms. 
Although I contemplate the possibility of one-off mechanisms, cognitive science is 
mostly interested in general explanations for general phenomena, so I won’t discuss 
unique cases here. 



 

 

44 
 

existence, even when they aren’t actively working. On the other hand, mechanisms-

as-causings are temporally tied to the production a given phenomenon. Confusing the 

two is making a category mistake—it is confusing a continuant with a causal 

occurrence.  

There are additional differences between modules and mechanisms-as- causings. I 

illustrated some with the examples discussed in 3 and 5, but let me elaborate a bit 

further on some other differences we can expect.   

One such difference will be whether it’s possible to change the range of 

circumstances in which the mechanism can operate.  Modularists like Carruthers 

define a module’s domain by the range of inputs suitable to turn the system on, not by 

module-independent input kinds19 (2006). For example, the proprietary domain of the 

face recognition system are face-like objects that are processed by the face 

recognition system. A sometimes-overlooked fact is that token modules can change 

their proprietary domains through time. For example, there is evidence that during a 

child’s development, the phoneme recognition system becomes increasingly 

specialized, resulting in increased sensitivity to phonemes of one’s language, at the 

expense of increased insensitivity to those of other languages that cross-cut the 

phonetic space differently. For another example, the exact number system slowly 

expands its proprietary domain during development, as children learn to use number 

 
19 In contrast, for Fodor “domain specificity has to do with the range of questions for 
which a device provides answers (the range of inputs for which it computes 
analyses)” (1983, p. 103). While these notions are conceptually different, 
extensionally they point I make above remains. 
 



 

 

45 
 

words to represent number concepts (Libertus et al. 2016, p. 208).  The domain of 

some modules can also change during adulthood: the face recognition system, for 

example, can be trained to also process Greebles (Gauthier & Tarr, 1997). In contrast, 

it is impossible that the range of circumstances in which a particular mechanisms-as-

causing can operate changes. Once the explanandum phenomenon is fixed, the world 

fixes what are the background and set-up or starting conditions of the mechanism-as-

causing (Craver & Tabery, 2019). 

Relatedly, modules usually have a dedicated brain network, whether that is scattered 

across the brain or localized primarily in one area. In contrast, cognitive mechanisms-

as-causings may have a lot more variation in their constituents. To mention a few: a 

mechanism-as-causing may be composed of a mixture of brain networks that aren’t 

regularly that causally interconnected (as in the mechanism for spontaneously 

producing a joke); or by just a very small part of the brain in a way that doesn’t 

follow natural boundaries (as in the mechanism of action potential); or by things 

outside the skull (as in the mechanism by which Otto came to know where Moma 

was, which includes his notebook, to use Clark and Chalmers’ 1998 example). Part of 

the reason for this large variability in what realizes a mechanism-as-causing is that 

the selection of phenomena to be explained for the mechanist is, to a certain extent, 

interest-dependent.  

Let me articulate what I mean by that. “Phenomena”, for MDC and the theorist of 

mechanisms-as- causings, is an epistemic category, not a metaphysical one. Anything 

can be a phenomenon of interest; it is the scientists’ choice to delineate it—including 



 

 

46 
 

making a decision about its range of application and degree of detail. While the 

researcher has a certain freedom in choosing the phenomenon, ultimately it is 

objective facts (about what produces that phenomenon) that establish what the 

underlying mechanism(s) for it is (are). Once a phenomenon is chosen, the 

explanatory work is to find the mechanism in the world causally responsible for it. In 

the best of cases, there is a single mechanism-type producing the phenomenon-type 

that the researcher is interested in. In other cases, however, lumping or splitting may 

be necessary. In any case, the mechanism is stablished by what it does. In contrast, 

the modularist and the theorist of mechanism-as-systems speak of a system’s 

“proprietary function” (see discussion in section 2). The notion of “proprietary 

function” applies to a more restricted set of cases than that of “phenomena”. A 

system’s function is the task it is meant to perform; not everything a system does 

when it is operating qualifies. In contrast, a mechanism is delineated by what it does.  

Despite these differences between modules and mechanisms-as-causings, there is a 

sense in which they overlap. Suppose that we fix the phenomenon in such a way that 

it overlaps with a module’s proprietary function. As I mentioned before, once the 

phenomenon is fixed, the mechanism responsible for it is also delineated. Then, in 

that case, when a token module is performing its proprietary function (let us designate 

that function with the letter X), then that module is a token mechanism-as -causing 

for phenomena X, while it is doing X. That is, during the time in which the module is 

actively functioning as it is meant to, that system is a mechanism-as- causing for 

phenomenon X. In that case, it may be unnecessary to distinguish which of the two 

notions of mechanism we are using.  



 

 

47 
 

Nonetheless, it is important to keep mechanisms-as-systems and mechanisms-as- -

causings distinct. Doing so allows for more clarity on quantification of mechanisms, 

as well as how to treat broken ones.  

Regarding quantification, mechanisms-as-systems and mechanisms-as -causings have 

different ways to count tokens of a kind. If we are counting mechanisms-as- causings, 

then we would look at the phenomena being produced. A mechanism occurs only so 

when it’s producing a phenomenon. Every time a mechanism is causally responsible 

for a phenomenon, then that is a token mechanism. When it stops acting, then we no 

longer have a mechanism. On the other hand, mechanisms-as-systems are objects that 

persist through time regardless of whether they are in operation or not. Here, 

sameness of mechanism seems to be determined by sameness of parts (although the 

exact conditions for identity are not clear; think of ship of Theseus-style problems). In 

any case, mechanisms-as -causings greatly outnumber mechanisms-as-systems.  

Lastly, I want to highlight that the notion of “broken normal” or “damaged 

mechanism” only makes sense under modular approaches that take mechanisms to be 

systems. A module has a proprietary function, whether or not it carries it out. But 

mechanisms-as-causings exist only so insofar they are producing phenomena. If they 

are not producing anything, they don’t.  

In a nutshell: the set of modules and the set of mechanisms-as-causings involved in 

cognition are not co-extensive; while modules can be regarded as mechanisms-as-

causings while they are performing its proprietary function, there are mechanisms-as-



 

 

48 
 

causings (sometimes recurrent mechanism-types, like the placebo one) that aren’t 

modules.  

 

7. How to treat work in progress 
 
In sections 2-6, I reviewed how mechanisms-as-systems and mechanisms-as -

causings are different. However, it is worth taking a step back to focus not so much 

on complete explanations —cognitive science has few of these anyway— but on 

research in progress. Most research involves fragmentary models, in which we know 

some but not all the parts of the target mechanism. If both sorts of mechanisms 

(system-like and cause-like) exist in cognitive science, how do researchers figure out 

into which category their target mechanism fits? How should they be treating their 

target mechanism in the meantime? 

Eric Hochstein has recently argued that working scientists should be “adopting 

deliberate metaphysical positions when studying mechanisms that go beyond what is 

empirically justified regarding the nature of the phenomenon being studied, the 

conditions of its occurrence, and its boundaries.” (Hochstein, 2019, p. 579). In his 

opinion, it is acceptable for scientists to make metaphysical commitments at the onset 

of experimental investigation, provided that we are willing to revise these 

commitments on light of the evidence: “our metaphysical commitments might 

eventually need to be revised after a great deal of empirical work has taken place” 

(Hochstein, 2019, 588). However, in my opinion, making a metaphysical 

determination whether the target mechanism is a system or a causing before the 



 

 

49 
 

evidence is conclusive would only make scientists talk past each other. This can be 

seen, I believe, in discussions regarding the boundaries of the mind/brain20 as well as 

mechanisms hypothesized by evolutionary psychologists21. The metaphysical 

commitments of our theories must correspond to what is warranted by the evidence, 

not to a philosophical demand to keep our metaphysics tidy.  

Instead, I think here (at early research stages) is where we should employ the minimal 

notion of mechanism, formulated by Glennan and Illari:  

“A mechanism for a phenomenon consists of entities (or parts) whose activities and 

interactions are organized so as to be responsible for the phenomenon” (Glennan & 

Illari, 2017, p. 2). 

What makes this notion minimal? To start, it is meant to capture the main dimensions 

of the mechanistic framework in general, without tying it to more specific accounts. 

The different sorts of accounts can be seen as varieties of the minimal notion of 

mechanism, which add to it further details (and further commitments) of the account 

in question. Importantly, the minimal notion is not metaphysically committed to 

mechanisms being systems or causings, or constitutive or etiological, or persistent or 

one-off. 

Back to my proposal. We should understand hypotheses or proposals in the early 

stages—when scientists speak of “mechanisms”, “systems”, “modules”, “causal 

 
20 Adams and Aizawa (2001, 2012) and Rupert (2009) make similar points. 
21 See, for example, Atran (2001). 



 

 

50 
 

factors”, “contributing factors”, etc. behind a phenomenon—as non-committal as yet 

about whether the phenomenon is underlain by a mechanism-as-system or 

mechanism-as-sort-of-causing. At early, hypothesis-generating stages, researchers are 

not saying anything metaphysically “thick” about what is behind the phenomenon in 

question. It would be an error to assume, for instance, that a range of behaviors that 

would have been adaptive are always produced by a dedicated, innately-channeled  

“system” in Carruthers’ (2006)  sense. Even if that ended up being the case, it 

wouldn’t be epistemically justifiable to hold that assumption until there is more 

evidence for the claim. But, they are definitely produced by (some) mechanism(s) or 

another. 

The New Mechanistic philosophy provides us with a toolbox to aid discovery. Which 

account to employ depends on our epistemic state regarding the target mechanism, as 

well as its nature —whether it is a persisting functional component or a sequence of 

causally interacting parts. At beginning research stages, the minimal notion is the one 

that should be employed; every other account of mechanisms (and modules) is 

thicker. Each makes concrete proposals as to which properties would count as 

evidence that, in reality, there is a mechanism-as-system or a mechanism-as-a-sort-of-

causing producing the phenomenon. For example, for MDC, it would be changes in a 

phenomenon when you intervene on a part;  for modules, it would be developmental 

and pathological evidence. Each account offers heuristics to help guide discovery, 

precisely because they are thicker. 



 

 

51 
 

If I am right, cognitive science must make use of more than one sort of account of 

mechanism, because those accounts capture the variety of mechanisms and of stages 

of knowledge about them. None of the accounts, in isolation, is adequate to the 

different stages of epistemic progress in discovering a mechanism, nor can they do 

justice to the diversity of mechanisms we find when it comes to cognition. Part of the 

literature has treated mechanistic accounts as each having its own proprietary 

scientific field. I am proposing that within the same field (for instance, cognitive 

science), there may be more than one sort of mechanism. If more than one account of 

mechanism applies to cognitive science, maybe the same is true for other scientific 

fields as well.   

This in turn puts an interesting twist as to how to understand the plurality of accounts 

of mechanisms available. If I am right, there is no winner-takes-all, even for a 

particular scientific discipline. An account may be appropriate for some mechanisms 

but not others; this is why we should view the New Mechanistic Philosophy as a 

toolbox of accounts. The criteria for choosing a given account of mechanism should 

not be merely a matter of scientific field, but a matter of the nature of the particular 

mechanism out there in the world one is investigating. 

  



 

 

52 
 

 

 

Chapter 3: Can the New Mechanistic Philosophy of 
Science find a role for representations? 
 

 

 

1. Introduction 
 
 
Within the philosophy of science there have been competing views on what 

explanation is. For example, the classic deductive-nomological model considered that 

to explain a phenomenon we must describe the initial conditions C, and the law-like 

generalizations L, from which to deduce the phenomena to be explained. However, it 

was soon evident that this universalist proposal posed some requirements on 

explanation that may be unachievable in the special sciences22, where exceptionless 

generalizations are rare and individual variation is to be expected. Instead, 

explanations in those sciences take a different form: scientists often appeal to 

mechanisms behind phenomena, and see their work as that of discovering these 

mechanisms. For instance, while discussing research go