Dilemma-JPhil resubmission


Modular architectures and informational encapsulation: A dilemma 

Dustin Stokes (University of Toronto) 
Vincent Bergeron (University of Ottawa) 

 

Abstract 

Amongst philosophers and cognitive scientists, modularity remains a popular choice for an 

architecture of the human mind, primarily because of the supposed explanatory value of this 

approach. Modular architectures can vary both with respect to the strength of the notion of 

modularity and the scope of the modularity of mind. We propose a dilemma for modular 

architectures, no matter how these architectures vary along these two dimensions. First, if a 

modular architecture commits to the informational encapsulation of modules, as it is the case 

for modularity theories of perception, then modules are on this account impenetrable. 

However, we argue that there are genuine cases of the cognitive penetrability of perception 

and that these cases challenge any strong, encapsulated modular architecture of perception. 

Second, many recent massive modularity theories weaken the strength of the notion of 

module, while broadening the scope of modularity. These theories do not require any robust 

informational encapsulation, and thus avoid the incompatibility with cognitive penetrability. 

However, the weakened commitment to informational encapsulation significantly weakens 

the explanatory force of the theory and, ultimately, is conceptually at odds with the core of 

modularity.	
  We then propose a non-modular notion of functionally independent system that, 
we argue, achieves the explanatory force sought by modularity theorists.  

 

 

Amongst philosophers and cognitive scientists, modularity remains a popular choice for 

an architecture of the human mind. Jerry Fodor (1983), who was influential in 

establishing the concept of a module, writes: “One day . . . Merrill Garrett made what 

seems to me to be the deepest remark that I have yet heard about the psychological 

mechanisms that mediate the perception of speech. ‘What you have to remember about 

parsing is that basically it’s a reflex.’” (Dedication). Reflexes are quick, inflexible, 



	
   2	
  

involuntary responses to stimuli, and Fodorian modules are like reflexes. In its most 

recent and general form, the modularity hypothesis consists in viewing the human 

mind, or at least part of it, as a configuration of quick specialized mental mechanisms, or 

subsystems, that are functionally independent of one another, and that typically operate 

over a distinct domain of information. 

There are compelling theoretical and empirical motivations for this approach. 

Theoretically, modularity nicely accommodates adaptationist and other evolutionary 

explanations of mental phenomena. It also provides materials for a simple explanation of 

important empirical data, including a wide range of behavioural dissociations, as well as 

the speed and robustness of processing enjoyed by the human mind. Most broadly, 

modularity provides an intuitive framework for characterizing the structure-function 

correspondence between brain structures and particular perceptual and cognitive 

functions.  

Although it is sometimes misrepresented as doing precisely this, Fodor’s pioneering 

discussion of the concept did not involve a definition of ‘module’. (Fodor 1983; see also 

Coltheart 1999). Fodor did, however, provide a list of properties symptomatic of 

modules. Fodorian modules are typically domain specific, hardwired, computationally 

autonomous, informationally encapsulated, fast, and their operation is mandatory. It is 

noteworthy how much of this characterization follows the reflex metaphor. Domain-

specificity parallels the singularity of the stimulus that sets off a reflex; autonomy, 

mandatoriness, hardwiring, and encapsulation mirror the standard reflex-arc model. 

Fodor maintains that “The notion of modularity ought to admit of degrees” (Fodor 1983: 

37), and that “if a psychological system has most of the modularity properties, then it is 



	
   3	
  

very likely to have all of them” (Fodor 1983: 137). Importantly, Fodor claimed only that 

input systems are modular. His primary subject matter was perceptual systems, but he 

also made the case for systems devoted to low-level linguistic decoding. Higher-level 

conceptual or cognitive systems, then, are not modular on Fodor’s general architecture. 

Commitments with respect to Fodor’s original analysis of modularity vary. Several 

modularity theorists take domain-specificity to be definitive of modularity (Coltheart 

1999). Fewer require innate specificity, even if related explanations and arguments often 

invoke evolutionary considerations. Others maintain that modules are informationally 

encapsulated and computationally autonomous (Farah 1994, Sperber 1996; 2001.)1 

Recent theorists have extended the modularity thesis beyond Fodor’s input systems. It 

is common among evolutionary psychologists to endorse some version of what Dan 

Sperber (1994) has called the massive modularity thesis. The general hypothesis states 

that all, or nearly all, of the mind is modular, and modules have been postulated to 

account for cognitive capacities as diverse as theory of mind, face recognition, cheating 

detection, reading, and a variety of social understanding abilities.  

Our suggestion is that informational encapsulation is essential to a distinctive, non-

trivial modularity theory. As it will be understood here, if a module m is 

informationally encapsulated then m cannot, during the course of its processing, access 

or compute over information found in other components of the overall system. As such, 

an encapsulated module m is impenetrable with respect to the other components of the 

system, since the processing of m is insensitive to (and so does not compute over) the 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
1 In at least two places, Fodor himself explicitly states that “informational encapsulation is an essential 
property of modular systems” (Fodor 1985: 3; see also 1983: 71). Elsewhere, however, he is less clear on 
his commitment regarding the same claim.  



	
   4	
  

information available elsewhere in the system (Pylyshyn 1980). This basic analysis of 

modularity is important to any substantive modular account of the mind because it 

constitutes the foundation of modularity in so far as modules are, in a sense to be 

explained later, functionally independent systems.  

In this respect, the modularity theorist faces a dilemma that hinges on the 

commitment to informational encapsulation. On the one hand, a commitment to 

informational encapsulation, as made by modularity theories of perception, is 

inconsistent with the cognitive penetration of perceptual experience. And, we argue, 

there are genuine cases (old and new) of the cognitive penetrability of perception. On 

the other hand, as recent modularity theorists have done, one might weaken the notion 

of module so as not to require informational encapsulation. The result, however, is an 

account that is inconsistent with one of the central motivations for modular 

architectures and, more fundamentally, with the conceptual core of the very notion of 

modularity.  

The first horn challenges strong, encapsulated modularity: any modularity theory that 

includes a commitment to informational encapsulation. The second horn challenges 

massive modularity, which broadens the scope of modularity but weakens the notion of 

modularity so as not to require informational encapsulation. Either way, modular 

architectures appear to be significantly challenged. 

In the concluding section, we argue that a non-modular notion of functionally 

independent systems can achieve much of the explanatory force originally sought by 

modularity theorists. This notion characterizes the specialization of the architectural 

units of the mind in terms of their cognitive workings rather than the cognitive or 



	
   5	
  

psychological uses to which they are put. This approach avoids the challenges to 

modular architectures, while allowing for a robust functional analysis of cognitive 

systems.  

 

I. Informationally encapsulated modules: Cognitive penetrability and the challenge for 

encapsulated modularity 

  

Both encapsulated and unencapsulated modularity theorists take perceptual systems 

to be modular. If perceptual modules are informationally encapsulated, then at the very 

least, they are not penetrable by the information or processing of higher-level cognitive 

systems. Most theorists seem to take the concepts ‘informational encapsulation’ and 

‘cognitive impenetrability’ to be co-extensive, if not equivalent—Fodor in fact 

originally argued for the encapsulation of modular input systems by arguing against 

claims about the cognitive penetrability of those systems (Fodor 1983: 73-86). The 

following discussion requires only the assumption that informational encapsulation of 

perceptual modules entails cognitive impenetrability.  

On this account, then, perceptual processing is not influenced by cognitive states like 

beliefs or desires. Evidence of this influence—that is, of the cognitive penetration of 

perception—thus threatens any modularity theory that includes a commitment to 

informational encapsulation of perceptual modules.  

It will be useful here to offer some clarifications. First, distinguish perceptual 

experience from higher-level cognitive and affective states and processes like belief, 

judgement, desire, emotion, and so on. The latter possess some kind of linguistic or 



	
   6	
  

propositional content, and play an immediate role in practical reasoning. The former is, 

whatever else one says about it, characterized by phenomenal character or content and 

depends non-trivially on one or more sensory organ. Philosophers debate how to draw 

the line between perception and cognition. The only point that need be granted here is 

that there are clear cases of perceptual states and clear cases of cognitive states. So there 

are visual experiences, auditory experiences, olfactory experiences, and so on; and these 

can be distinguished from states like belief and processes like decision making. 

Second, distinguish the cognitive penetration of perceptual experience from the 

cognitive penetration of perceptual processing. The former concerns some difference in 

the phenomenal content or character of a perceptual experience, where this difference 

depends non-trivially upon some cognitive state or processing in the system. The latter 

only concerns some cognitive effect on perception at the level of processing. The fact 

that perceptual processing at some stage is cognitively penetrated does not, by itself, 

entail the cognitive penetration of experience. Experience may depend on a wider class 

of processing and, in principle, the cognitive influences on perceptual processing (at 

some particular stage or other) may not ultimately influence conscious experience. 

Moreover, some aspects of perceptual processing may not give rise to a conscious 

experience but rather, for example, the sub-personal guidance of motor performance 

(Goodale and Milner 1992).  

However, cognitive penetration of perceptual experience does entail cognitive 

penetration of perceptual processing at some level. There is much to be said here. The 

only assumption we need regarding the relation between perceptual experience and 

perceptual processing is this. Whether one takes experience to be identified with, 



	
   7	
  

constituted by, supervenient upon, or the output of perceptual processes, a difference 

in perceptual experience implies a difference in perceptual process. This will be 

generally true—albeit for different reasons—no matter how one’s metaphysics of mind 

varies according to these alternatives. So if experience is penetrated, then information 

relevant to cognitive systems, or the processing of cognitive systems, directly influences 

the processing of perceptual systems. It is this entailment relation that is important for 

the criticism we offer below. 

One final point: Although the cognitive penetration of experience entails the 

penetration of processing at some stage, the cognitive penetration of experience is 

compatible with the cognitive impenetrability of processing at some stage or even most 

(but not all) stages. This point is instructive: one cannot argue from the purported fact 

that some particular perceptual module is impenetrable to the claim that perception 

broadly or perceptual experience (in that modality) is impenetrable. Zenon Pylyshyn, 

for instance, argues that “early vision” is not penetrable by cognition (Pylyshyn 1999). 

Pylyshyn’s empirical claim about early visual processing, even if true, is insufficient to 

support the thesis that perception is cognitively impenetrable. Indeed, Pylyshyn admits 

that the output of this component of the visual system, as he and most theorists 

understand it, does not (alone) determine perceptual experience. His defence of 

cognitive impenetrability is thus consistent with the cognitive penetrability of 

perception; one might accept that the computations performed by the early visual 

system are impenetrable by cognitive states but maintain that perceptual processing is 



	
   8	
  

penetrated elsewhere such that the resulting perceptual experience is causally 

dependent upon cognition2 

Although there is ample evidence for “top-down” processing in the brain—where 

information is exchanged between various areas of the cortex, including those areas 

believed to process higher-level or conceptual information—current neuroscience lacks 

an uncontroversial mapping from conceptual mental states (like belief) onto brain 

structures. And some such mapping would be necessary for neuroscience to provide a 

verdict on the actuality of cognitive penetration.3 Consequently, empirical evidence for 

cognitive penetration must be obtained at the behavioural or psychological level, rather 

than merely the neurological level. Predictably, there are a number of possible 

alternative interpretations of this data, and so the inference structure is abductive. 

Critics of cognitive penetrability appeal to these alternative interpretations as better 

explaining alleged cases of cognitive penetration. We identify four such general 

skeptical strategies. With these strategies in hand, a working definition of cognitive 

penetration can be devised, in hopes of isolating the target phenomenon in a way 

agreeable to both sides of the debate.  

First, for some cases, critics claim that what is affected by the subject’s cognitive 

states is the subject’s memory rather than her perceptual experience. Subjects recall the 

stimulus to be some way (i.e. a way different than if the relevant cognitive states are 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
2 A number of critics have questioned Pylyshyn’s conclusions in this general way (Bermudez 1999; 
Macpherson 2012; Moore 1999; Noë and Thompson 1999). It is also worth noting that Pylyshyn’s 
empirical claim can be challenged (see Boynton 2005; Kamitani and Tong 2005).  
3 Some have argued that evidence for reentrant neural pathways is evidence for cognitive penetration 
(Churchland 1988; 1989). Others have argued against this line of reasoning (Fodor 1988; Gilman 1991; 
Raftopoulos 2001). For purposes of this discussion, we simply assume that the neurological evidence is 
presently insufficient to count in either direction.   



	
   9	
  

controlled for) as a result of some other cognitive state, and report a memory of the 

stimulus rather than a perceptual experience. This evidences cognitive penetration of 

cognition. And this is uncontroversial: memories can be faulty, and in ways influenced 

by what we believe, desire, or otherwise think. Call this the memory interpretation.4  

A second strategy is the attention-shift interpretation. This interpretation maintains 

that in the cases in question, cognitive states of the experimental subjects cause a shift 

in attention, generally involving some overt action, which then results in the change in 

perceptual experience. Thus the link between cognition and perception is mediated by 

an external action. This is no different in kind, critics urge, from an ordinary perceptual 

scenario where one, for example, has some belief about one’s environment (e.g. that 

there is an irritating noise somewhere around here) and this belief causes some action 

(e.g. going to look for the thing making the irritating noise), which in turn results in a 

changed perceptual experience (e.g. seeing the squeaky faucet). This familiar cognitive-

behavioural dynamic is important to everyday life, but unless cognitive penetration is 

trivially rampant, this is not cognitive penetration. Cases involving shifts of attention, 

this interpretation suggests, are to be treated similarly: these scenarios lack an 

appropriate internal connection, and so there is nothing in this causal chain to properly 

call ‘penetration’ (see Pylyshyn 1999: 343).5 

Third, critics have suggested that experimental subjects are not reporting a 

cognitively affected perceptual experience, but instead a judgement of the perceived 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
4 For one example, see McCurdy 1956. Note also that if memory is factive as some have argued 
(Williamson 2002), then the memory-interpretation amounts to something like a quasi-memory 
interpretation.  
5 Fodor also appeals to this general response in his debate with Paul Churchland on the theory-ladenness 
of perception/observation (see Fodor 1988; Churchland 1988; see also Fodor 1983). 



	
   10	
  

stimulus. So the perceptual experience of the stimulus remains unaffected. At most, the 

subject judges or evaluates the stimulus in a way she would not if she lacked some 

background cognitive state/s. This difference manifests in the different reports of the 

experimental subjects versus the control subjects in particular studies. Call this the 

judgement interpretation.6 

Finally, an intra-perceptual interpretation claims that the effect is not a cognitive one 

but instead occurs as an adjustment or adaptation within the perceptual system. There 

are many ways to develop this alternative. One might claim that certain types of stimuli 

(for example, naturally occurring stimuli like ripe fruit) are such that the human 

perceptual system is appropriately “tuned” so as to represent these objects more quickly 

or in some enhanced way. These effects would not be learned, instead perhaps artefacts 

of the evolution of human sensory systems. Or one might claim that a subject acquires a 

non-cognitive association with the stimulus type, and this association affects how 

perceptual information regarding tokens of this type are processed and/or how one acts 

in response to the stimulus. Or one may argue that sensory systems are sufficiently 

plastic to “learn” new ways to process sensory information, perhaps for some adaptive 

advantage.7  Differences to one side, the thread common to these interpretations is that 

some cases may be better explained by changes in the sensory system that do not 

depend upon background cognitive states.  

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
6 For additional discussion of these and other strategies for the cognitive impenetrability theorist, see 
Macpherson 2012; AUTHOR MASK B-XXXX. 
7 Against Churchland’s appeal to subjects’ adjustment to inverting lens as evidence for diachronic 
cognitive penetration, Fodor appeals to an intra-perceptual interpretation (see Fodor 1988: 193). For a 
more recent use of this kind of interpretation, see Deroy (forthcoming), who analyzes some of the research 
also discussed below.  
 



	
   11	
  

Grant that if any alleged case of cognitive penetration can be interpreted in one of 

these alternative ways, then the critics are correct: it is not a genuine case of cognitive 

penetration of experience. We can then define cognitive penetration so as to rule out 

these interpretations, and ask if any case plausibly meets the definition. If the answer is 

‘yes’, then the critics must secure some alternative interpretation to deflect the case/s. 

Here is such a definition: 

 

(CP) A perceptual experience E is cognitively penetrated if and only if (1) E is 

causally dependent upon some cognitive state C and (2) the causal link between E 

and C is internal and mental. 

   

The definition requires a few qualifications. First, clause (2) says that if an 

unscreened internal cause involves a cognitive state—that is, the causal chain runs from 

experience back to a belief, desire, or some other cognitive state without deviating from 

internal mental processes—then the perception depends (internally) upon a cognitive 

state. Counterfactually, had C not been present in the process, E would not be had by 

the subject. C is thus a necessary causal condition for E. Understood probabilistically, C 

is not a strictly necessary causal element, but one that is highly relevant to the 

probability of that perceptual experience; E is more likely to be had when C is present, 

and less when not present. The preferred notion of causation is of little matter so long as 

the internal causal dependence is maintained.8  

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
8 One should also note that C is a non-sufficient cause of E. There are other relevant causal factors.  



	
   12	
  

Second, (CP) excludes obviously non-genuine cases of cognitive penetration.  For 

example, a desire to see the symphony, coupled with a true belief about the location of 

the symphony, may result in a perceptual experience of the symphony (several 

experiences in fact). But this should not count as an instance of cognitive penetration of 

experience, else the concept ‘cognitive penetration’ becomes trivial. (CP) delivers the 

appropriate result: such a case fails to satisfy our definition, since it fails to satisfy clause 

(2). In cases like this, a cognitive state (or some cognitive states) motivates an action (or 

set of actions) which eventually results in the relevant experience.  The perceptual 

experience thus causally depends upon the relevant cognitive state/s. Clause (2) ensures, 

however, that this is not an instance of cognitive penetration, since the cognitive state 

(or states) is screened from being internally, causally efficacious: the cognitive state 

causes an (external) action which eventually results in the experience.9  

Importantly, a perceptual experience that satisfies this definition cannot be 

interpreted in any of the four ways described above. Clause (1) of (CP) rules out the 

memory, judgement, and intra-perceptual interpretation, since it requires a cognitive 

influence on perception, rather than just an influence on some other cognitive state in the 

system or an intra-sensory adjustment. Clause (2) of (CP) rules out the attention-shift 

interpretation, since it requires a non-externally mediated causal link between the 

cognitive state and the perceptual experience. The question now becomes: are there any 

experimental cases that satisfy (CP)? We now consider three sets of studies that strongly 

suggest that the answer is ‘yes’.  

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
9 In this way, CP is consistent with other recently proposed definitions of cognitive penetrability: 
Macpherson 2012; Siegel 2011; AUTHOR MASK B; C.  



	
   13	
  

The first study was performed by Jerome Bruner and C.C. Goodman (1947). It is a 

pioneering study in at least two ways, it employed a testing methodology—involving 

online perceptual tasks, as will be explained below—which best isolates cognitive effects 

on perception. Second, it initiated a movement in psychology, the New Look. New look 

psychology was important both for its explicit opposition to its behaviourist 

predecessors—contra behaviourism, the New Look argued that the proper explananda 

of psychology include internal mental states and processes—and for the way it made 

that opposition—the New Look theorized perceptual experience as an active 

construction of representations of the environment. The thesis was essentially a 

universally quantified cognitive penetrability claim: (human) perceptual representations 

are always (or nearly always) constructed in a way informed by the perceiver’s “mental 

set”—her expectations, needs, values, desires, and other higher-level mental states. 

Criticism, and eventual dismissal, of New Look targeted the strength of this claim: 

critics constructed a range of experiments where perception appeared unaffected by 

cognitive states. But a complete dismissal of New Look, like the universal claim upon 

which it was predicated, is overstated. Many of the New Look studies provide good 

evidence for an existentially quantified cognitive penetrability claim. And the claim of 

mere existential strength is the subject of today’s debate.10  

Bruner and Goodman’s 1947 studies ran as follows. Three groups (10 persons per 

group) of 10 year old children, two experimental and one control, were put before a 

wooden box with a glass screen on its face. In the centre of the screen was a small patch 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
10 See Balcetis and Dunning 2006 and van Ulzen 2008 for brief historical discussions of the rise and fall of 
the New Look movement, as well as new studies in the New Look spirit. 



	
   14	
  

of light, nearly circular in shape, the circumference of which could be adjusted by a 

small knob located on the bottom right corner of the box. The two experimental groups 

of children were presented with ordinary American coins of varying values. As they 

looked at the coins, placed flat in the palm of the left hand, positioned at the same 

height and six inches to the left of the adjustable patch of light, they were asked to 

adjust the patch to match the size of the presented coin. The subjects could take as 

much time as they liked to complete the task. The control group was instead presented 

with cardboard discs of sizes identical to the relevant coins, and asked to perform the 

same task. In the experimental group, perceptual experience of the coins was 

“accentuated.” The experimental subjects systematically overestimated the size of the 

coin, and sometimes by a difference as high as 30% as compared with control subjects, 

who report the size of the cardboard analogues with near perfect accuracy.11  

The second experimental variation divided experimental groups into subgroups 

comprising “rich” and “poor” children. The task was the same, except only real coins 

were used. Here, rich children, as the previous results would suggest, still overestimate 

the size of the coins, but at percentages significantly lower than the poor children. 

Indeed, poor children systematically overestimate the size of coins, by as much as 50%, 

and by differences as high as 30% as compared to rich children.12 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
11 For example, experimental subjects overestimated the size of a dime by an average of 29%; controls 
underestimated the size of the cardboard analogue of a dime by -1%. 
12 A number of theorists were critical of particular details and the broad scope of the New Look approach 
(Klein, Schlesigner, and Meister 1951; Carter and Schooler 1949; Lysak and Gilchrist 1955). These critics 
were right to challenge the universally quantified New Look claim, and by simply acquiring evidence for 
cases where cognition apparently fails to affect perception. But, as suggested above, this evidence fails to 
undermine the more modest implication that cognition sometimes influences perception in the relevant 
ways. Moreover, the Bruner and Goodman 1947 results have been broadly replicated by a number of 
similar studies at least insofar as these studies all evidence some higher-level effect on perceptual 



	
   15	
  

This case prima facie satisfies (CP). The experimental subjects have a perceptual 

experience, the character or content of which causally depends on a cognitive state, in 

this case, a desire or value. And the causal link between the experience and cognitive 

state is internal and mental. 

Nonetheless, critics may, and have, resisted this as a genuine case of cognitive 

penetration, and by appeal to one of the strategies outlined above. But this rejection is a 

mistake premised on a failure to carefully consider Bruner and Goodman’s experimental 

procedures. Importantly, the procedure for each experiment involved an online 

matching task. Subjects took as much time as they needed to adjust the light patch to 

match the size of the coins. The coins were presented at the same time as, at the same 

horizontal level as, and six inches to the left of, the adjustable light patch. Subjects did 

not visually inspect the coin and then shift to a distinct visual field, adjusting the light 

patch by memory. Instead, their task was to adjust the patch of light to match what they 

were seeing, while they were seeing it. In no relevant sense were they forced to base 

their report just on memory. The memory-interpretation thus fails.  

For the same reason, the attention-shift interpretation fails: subjects did not attend to 

one stimulus (the disc or coin) and then shift attention to a distinct visual field where 

the second stimulus (the adjustable light patch) was located. The subjects would have 

shifted their gaze from disc/coin to light patch, but this slight shift would not have 

differed across control and experimental subjects (and so would fail to explain the 

relevant differences between the two groups). 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
experience. See Bruner and Postman 1948; Postman, Bruner, and McGinnies 1948; Bruner, Postman, and 
Rodrigues 1951; Dukes and Bevan 1952; Bruner and Rodrigues 1953; Bruner and Minturn 1955; Blum 
1957; Holzkamp and Perlwitz 1966.  



	
   16	
  

The intra-perceptual interpretation, no matter how it is developed, simply does not 

apply. The experimental stimuli in question are artefactual and so are not stimuli for 

which we have an evolved perceptual sensitivity, nor a kind with which we are likely 

to form non-cognitive associations.  

The judgement interpretation is most commonly used to deflect cases like the Bruner 

and Goodman case. The only version of this interpretation that is inconsistent with 

cognitive penetrability is one that claims that the perceptual experiences of the subjects 

are accurate across control and experimental subjects alike, while the experimental 

subjects make a misjudgement of the size of the coins. This interpretation is less 

plausible than the interpretation it opposes, since it requires attributing a judgement or 

belief to the subject that does not correspond to the perceptual experience that she has 

simultaneously with that judgement or belief. It requires that the subject, while 

inspecting the visual stimulus—which is, again, at a location six inches to the immediate 

left of the adjustable light patch—consistently makes erroneous judgements about what 

she is seeing. The reporting method here is key: since subjects report by adjusting the 

light patch, the judgement interpretation requires that subjects see the coin and light 

patch accurately (e.g. they see the light patch as significantly bigger than the dime at 

the time of concluding with a report, since this is the resulting data) but then judge, 

with both stimuli present, that the target and light patch are the same size. 

Here, the cognitive impenetrability theorist might respond by invoking instances 

where perception and judgment do come apart in just this way. So, for example, 

although one sees the Müller-Lyer lines as being of different lengths, one believes (if 

one knows the illusion) that the lines are of the same length. And indeed one cannot 



	
   17	
  

manage to see them accurately in spite of this background knowledge (Fodor 1983, 

1985, 1988; Pylyshyn 1999). So, the critic would argue, a consistent mismatch between 

simultaneous experience and judgment is not so uncommon, and perhaps Bruner and 

Goodman’s subjects can be explained similarly. However, the subjects in the Bruner and 

Goodman experiments are importantly different from standard perceivers of the Müller-

Lyer and other such illusions. When one judges and reports that the Müller-Lyer lines 

are of the same length, one bases this report not on current perceptual experience, but 

on knowledge of the illusion. Bruner and Goodman’s subjects are different in this 

regard: they intend for their report to be one of what they presently see (Bruner and 

Goodman take careful measures to adequately instruct the subjects). If asked, the 

subjects would certainly report that they are matching the light patch to what they are 

seeing, and moreover would confirm that their report—that the light patch matches the 

coin—is based on what they see. To treat these subjects like perceivers of the Müller-

Lyer illusion requires that they are systematically mistaken about this: the subjects are 

not reporting on the basis of what they see. 

The judgement interpretation, then, must maintain that these subjects are continually 

ignoring, remaining unconscious of, or somehow otherwise failing to accurately report 

their perceptual experience. Imagining the phenomenology of such a situation further 

reveals its implausibility. In the first experiment, the judgement interpretation requires 

that a subject have a veridical experience of, for example, a dime. And, simultaneously 

and upon inspection, she adjusts the matching patch of light to 129% of the 

(veridically) perceived coin (and, moreover, this light patch is also perceived 

veridically). And in the second experiment: a quarter is perceived (by “poor” subjects) 



	
   18	
  

veridically but, simultaneously and upon inspection, the light patch is over-adjusted to 

150% of the (veridically) perceived coin. These are significant differences in size and so, 

given the online nature of the task, maintaining the judgement interpretation requires 

attributing significant cognitive error to the subjects. Moreover, what explains this 

error? The judgement interpretation must provide some answer here but it is far from 

clear what form it would take. The cognitive penetration interpretation, by contrast, 

only requires attribution of perceptual error (the effects of background cognitive states) 

that, in turn, explains the subjects’ reports. Given the experimental circumstances, the 

latter interpretation is far more plausible; the judgement interpretation fails.  

Crucial to disarming the alternative skeptical interpretations is the online nature of 

Bruner and Goodman’s experimental procedure. Many recent studies in psychology are 

suggestive of cognitive penetration13, but the most convincing studies follow some of 

Bruner and Goodman’s basic methods. We now briefly present two such studies—the 

first on colour perception of natural and artificial objects, the second on the influence of 

racial categories on visual experience. Both studies involve relevantly controlled online 

perceptual tasks.  

In a recent study, Thorsten Hansen and colleagues tested colour perception of objects 

with high “colour diagnositicity”, objects the concepts of which are partly constituted 

by a distinctive colour concept: YELLOW for bananas, RED for strawberries, ORANGE 

for carrots, and so on (Hansen et al 2006). The procedure involved the presentation, on a 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
13 For example, each of the following studies present data that may be plausibly explained in terms of 
cognitive penetration: Balcetis and Dunning 2006, 2010; Payne 2001, 2005; Stefanuci and Proffitt 2008, 
2009; Witt and Dorsch 2009. However, the experimental controls in these studies are such that the results 
could also be plausibly explained in terms of one (or more) of the mentioned alternative interpretations.  



	
   19	
  

computer monitor, of digital photographs of natural fruits/vegetables, presented in their 

typical colour, set against a uniformly grey background. The subject’s task was to 

adjust the fruit image to what she judged to be a neutral (achromatic) grey. What in fact 

happens is that subjects adjust the image past achromatic grey and into the opponent 

colour range (e.g. adjusting a banana image past grey into the bluish hue). The 

researchers describe this as the memory colour effect. The researchers quantify this effect 

with a memory colour index (MCI), which in simplest terms provides a measure of the 

achromatic adjustment, towards the colour typical of the stimulus object (negative 

index) or away from it into the opponent hue range (positive index), relative to the 

typical colourfulness of the object. So for example, for a banana, the MCI is the ratio of 

the distance of shift past the perceiver’s grey point into the bluish hue range to the 

distance of the shift from the typical yellow of the banana to the grey point, (with both 

of these distances measured along the same axis of typical adjustment for subjects).14 For 

all of the experimental conditions, the MCI ranges from +4 to +13%, with a mean effect 

of +8.23%. As the researchers clarify, this quantification corresponds to an effect that is 

approximately three to five times above the threshold of discrimination. As a control, 

subjects perform the same task with uniformly coloured discs, and there is no memory 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
14 More specifically, the researchers clarify the calculation of the MCI as it is used in all three of the colour 
perception studies discussed here, as follows. “For the MCI the achromatic adjustments are projected on 
the axis of the typical adjustments that leads through the subjective grey point. The distance of this 
projection from the subjective grey point measures how strong the shift along this axis was. For the MCI 
this measure is divided by the length, i.e. the saturation, of the typical adjustment. In this way, the MCI 
represents the ratio of achromatic shift relative to the colourfulness of the typical colour. The sign (+/–) of 
the MCI reflects the direction in which the adjustment is shifted away from the subjective grey point. A 
positive MCI indicates an achromatic adjustment opposite to the typical adjustment. A negative MCI 
implies, contrary to the memory colour effect, that there is a shift of the achromatic adjustments towards 
the same direction as the typical adjustments. The MCI has been calculated separately for each participant 
using their subjective grey point” (Witzel et al 2011: 37). 



	
   20	
  

colour effect: subjects adjust the discs to achromatic grey with perfect accuracy. We 

should emphasize that in this study (and those discussed immediately below), the task 

was clearly perceptual and online. Subjects took as much time as they felt necessary to 

make the adjustments, thus making adjustments to the perceptual stimuli in real time. 

This case plausibly meets definition (CP). As the researchers hypothesize, a 

fruit/vegetable image, say a banana, still appears yellow to the subject at the point of 

achromatic grey. This hypothesis explains the fact that the subject adjusts the image 

into the bluish range, to compensate for the residual yellow, and then reports the fruit 

to be grey (when in fact it is slightly blue). This colour experience seems to depend, in a 

direct way, upon beliefs or conceptual associations with the relevant fruit/vegetable 

objects. Because the testing procedure involves online adjustment of the target stimuli 

itself, the memory interpretation is not appropriate. For similar reasons, the attention-

shift interpretation fails: there is no plausible explanation whereby subjects, in 

experimental but not control conditions, execute overt (or covert) attention to get the 

relevant effects.15 And the judgement interpretation would require that, as the subject 

visually inspects and adjusts the target stimulus, she veridically perceives the stimulus 

(e.g. a banana image as slightly blue) but then reports a judgement that it is perfectly 

grey. So she sees the stimulus accurately but reports it erroneously. And this error has 

to be explained in a way that current (veridical) perception is bypassed or ignored as 

informing the subject’s report, in spite of the task being an explicitly perceptual one. 

This looks much less plausible than an explanation where a non-veridical experience, 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
15 This is by contrast, for example, with Fodor’s favoured explanation of the way one can shift, by 
attentional changes, one’s experience of the Necker cube or the duck-rabbit. See Fodor 1988: 190. 



	
   21	
  

itself causally dependent on background cognitive states, causes a judgement and report 

that the target is perfect grey (when it, the banana image for example, is in fact 

objectively, slightly blue). Here the report is erroneous—as the data make clear—but 

the error in report is explained by perceptual error, and the perceptual error is 

explained by cognitive penetration.  

However, what about the intra-perceptual interpretation? One might worry, that 

since the target stimuli are all natural objects, the memory colour effect is symptomatic 

of hard-wired sensitivities of the human perceptual system. An enhanced perceptual 

sensitivity to ripe fruit and vegetables would plausibly be an evolutionary advantage 

for humans. And so granting that subjects still see the banana image as slightly yellow 

even when it is objectively grey, one might argue that this is best explained by facts 

about human perceptual processing and how it has evolved, without any needed appeal 

to cognition. 

This interpretation may appear even more plausible in the light of a second study 

performed by some of the same researchers, where the memory colour effect was most 

pronounced for realistic images of fruits/vegetables (e.g. those depicting texture) and 

mostly absent for mere fruit/vegetable outline shapes (Olkonnen et al 2008). However, 

this interpretation is easily dispelled by a very recent study (Witzel et al 2011). These 

studies involve artificial, human-made objects as stimuli. In a preliminary study, the 

researchers identify artificial objects with maximal colour diagnosticity, the blue Smurf, 

the Pink Panther, the red Coca-Cola logo, a green ping pong table, and so on. Images of 

these objects are then included in an experiment where the task is the same as the above 

two studies, plus a few additional controls. Target objects are initially presented in a 



	
   22	
  

random colour (e.g. a fire extinguisher might appear as blue rather than its typical red 

colour) against a uniformly grey background, and subjects then adjust the object to 

what they perceive to be achromatic grey. Additionally, control objects that typically 

vary in colour (e.g. a sock) and control objects that are typically achromatic (e.g. a golf 

ball) are presented in a random colour where the task is the same. Under these 

conditions, there is no evident effect for control objects, and a significant effect for 

colour diagnostic objects. The mean MCI for fourteen colour diagnostic stimuli was 

+3.31%, with a high of +10.3% (for the blue Nivea tin). Just as in the earlier studies, 

the results provide strong evidence for a cognitive effect on perceptual experience. And 

importantly, the intra-perceptual interpretation is not applicable to this most recent 

study: there is no story to be told about the evolution or plasticity of perceptual systems 

for the perception of cartoon icons or soda logos.16 

Finally, consider a recent study on racial stereotypes and face perception (Levin and 

Banaji 2006). In Experiment 1, subjects were presented, on a computer monitor, with 

realistic greyscale images of male faces with features stereotypical of either black or 

white persons (with hair removed). The task was to match the luminance of an 

adjustable greyscale face to the target face (in some conditions the adjustable face was of 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
16 It might be worth adding that this effect is plausibly culture-sensitive. In their initial study to identify 
colour diagnosticity for artificial objects, which was performed in Germany, Witzel et al (2011) found that 
some stereotypically German images were highly colour diagnostic (as measured by reaction time and 
accuracy of typical colour identification)—for example the orange Die Maus (a German television 
character), the yellow German mailbox, the yellow (German-made) UHU glue tube. But some non-German 
objects were not sufficiently colour diagnostic (relative to German subjects)—for example, the yellow 
Ferrari symbol and the red Soviet flag. These researchers did not run the study using these non-colour 
diagnostic (relative to German subjects) objects, but presumably if they had, any memory colour effect 
would have been insignificant at best. This factor lends additional plausibility to the effect being a 
cognitive one: subjects learn (and sometimes differently in different places) what colours are typical of 
artificial objects, and these beliefs or conceptual associations affect colour experience of (images of) such 
objects.   



	
   23	
  

the same racial prototype as the target, and in others of the opposite racial prototype). 

Although the luminance of the two target prototypes was (objectively) identical, 

subjects consistently adjusted to a lighter grey for the stereotypical white faces and to a 

darker grey for the stereotypical black faces (in both mixed-race and same-race 

conditions).17 Now, one might worry that in this experiment, the results are mere effects 

of optics. For example, in spite of both prototypes being identically greyscale, the 

contours of typical black facial features versus typical white facial features vary such 

that, one might conjecture, the first type of face naturally looks darker than the second 

type of face. Or one might attempt to invoke some variation of the intra-perceptual 

interpretation to deny the cognitive influence of racial stereotypes. These 

interpretations, however, fail to apply to a second variation of the study, one that in fact 

yields even more striking results. 

In Experiment 2, Levin and Banaji first, in a preliminary study, created a racially 

ambiguous face by morphing a range of prototypical black and white-face features, and 

then confirmed the ambiguity of the face by appeal to racial classification results across 

15 subjects. On an instruction screen, the ambiguous face (call this ‘BW’) was then 

paired with either an unambiguously white face (call this ‘W’) or an unambiguously 

black face (call this ‘B’). And in each condition, both faces were labelled, either ‘Black’ 

or ‘White’. So for example, when paired with an unambiguously white face, the 

ambiguous face (BW) was labelled ‘Black’ and the unambiguous white face (W), labelled 

‘White’. Taking this condition as our example—Levin and and Banaji call this the 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
17 More specifically, for example, with a black-face prototype as target, subjects adjusted a white-face 
prototype to 4.65 levels darker (out of 256 possible greyscale levels for the computer monitor) than a 
white-face prototype target (where, again, both targets are of identical luminance levels).  



	
   24	
  

“BW/W condition”—the task phase proceeded as follows. Subjects were presented with 

a series of trials, where each trial involved either the ambiguous face (i.e. the one 

labelled ‘Black’ in the instruction phase of the BW/W condition) or the unambiguous 

white face (labelled ‘White’ in the instruction phase), both of identical luminance, 

coupled with an adjustable rectangular region of uniform grey. The task in each trial 

was to adjust the grey report patch to match the face simultaneously perceived. Result: 

the racially ambiguous face is reported in a way that strongly correlates with the 

semantic labelling prime. So, in the BW/W condition, the lightness report for the 

ambiguous (‘black’-labelled BW) face was .465 levels darker (than the objective 

luminance of the target) and 17.85 levels lighter for the unambiguous (‘white’-labelled 

W) white face. And here is perhaps the most striking result: when the same ambiguous 

face BW is labelled ‘White’ (in the opposite “B/BW condition”) the report for BW is 

15.95 levels lighter. So: present a face identical both with respect to luminance and 

facial features, but change the label from ‘Black’ to ‘White’, and the reported match 

goes from .465 levels darker to 15.95 levels lighter (than the objective luminance of the 

ambiguous face)(2006: 505-6).18 This is not an effect explained just by optics or (intra-) 

perceptual features; the linguistic label is clearly playing an operative role in the 

subject’s perceptual experience.  

To conclude discussion of this final set of studies, consider the remaining alternative 

interpretations. The Levin and Banaji results are not well-explained by memory since 

this study involves online perceptual matching tasks. Nor is there any reason to think 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
18 See footnote 17 for clarification regarding the grey measures. 



	
   25	
  

that overt shifts in attention are explanatory of the effects.19 Thus both the memory and 

attention-shift interpretations fail to apply. What about the judgement interpretation? 

Considering Experiment 2, this interpretation would require that a subject, upon 

confirming her report and moving to another trial, has veridical experiences of the 

target face and the report region of grey—for example, where the grey report region 

would appear as (on average) 15.95 levels lighter than the simultaneously perceived 

(‘white’-labelled) ambiguous face. But then somehow the subject, in spite of perceiving 

this difference, judges and reports the face and report region as matching. This is far 

less plausible than the opposing cognitive penetrability thesis. The best explanation, 

here and above, is that the subject is having a non-veridical experience. She sees the 

prototypical white and prototypical black face as lighter and darker, respectively, and 

in Experiment 2, this effect is exaggerated by a linguistic labelling prime. The non-

veridical experience is a result of penetrating cognitive states, in this case, racial 

stereotypes or beliefs.20  

 Summarizing, there are crucial methodological features common to the three sets 

of studies discussed above. First, in all of these studies, subjects must perform online 

perceptual tasks. Thus the target stimulus—a coin, a Smurf image, a greyscale face—is 

present and perceivable while the subject makes her report. Use of this methodology 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
19 In fact, the researchers devise a third experiment explicitly devoted to discounting an explanation 
where attention is drawn to facial contours (e.g. of the stereotypical black face) in a way that explains the 
perceptual differences that appear in the results. They construct greyscale line drawings—with either 
white lines or black lines providing the facial outlines, but with no other shading of facial features—of 
the white and black prototypes. The results are relevantly the same and statistically significant: subjects 
choose darker samples for the black prototype faces and lighter samples for the white prototype faces. See 
Levin and Banaji 2006: 506-8. 
20 Macpherson 2012 briefly discusses both Hansen et al 2006 and Levin and Banaji 2006. She also provides 
a detailed analysis of an earlier study on colour perception, Delk and Filenbaum 1965.  



	
   26	
  

disarms the memory and attention-shift interpretations. Second, the method for 

reporting involves some direct kind of manipulation, either of the target stimulus itself 

or of some match disc or region. It is this methodology that, coupled with the online 

methodology, disarms the judgment interpretation. (Compare: Many other experiments 

use verbal reports of some kind. And a task involving a verbal report—for example, 

providing a numerical estimate of the distance of a perceived object or the height of an 

incline—opens space for judgement about perception and, in turn, encourages the 

judgement interpretation.) Finally, the stimuli used in these studies are all ones about 

which we learn and form beliefs, desires, and other cognitive states. It is this 

methodology that, at least partly, disarms the intra-perceptual interpretation. As a point 

of methodology, we suggest that any experimental attempts to hone in on cognitive 

penetration should employ, at minimum, this combination of features. And this 

approach can be traced back to Bruner and Goodman’s important work of over 60 years 

ago.    

Now, what does all of this imply for a strong modularity theory, any theory that 

commits to informationally encapsulated perceptual modules? If the above discussion is 

successful, then the standard alternative strategies fail to deflect the discussed cases as 

genuine evidence for the cognitive penetration of perception. In each case, whether it is 

a desire, value, belief, or some other higher-level mental state, there is evidently some 

cognitive state (internally) influencing experience. These cases are best described as 

meeting the conditions of (CP). And therefore, as we will now argue, perceptual systems 

are not informationally encapsulated.  



	
   27	
  

Recall that perceptual systems are paradigms for modular systems. And recall further 

that if one is an encapsulated modularity theorist, then one commits to the informational 

encapsulation of modules. The informational encapsulation of perceptual modules 

entails cognitive impenetrability. Finally, the cognitive penetration of perceptual 

experience entails, at some level, the cognitive penetration of perceptual processing. 

Thus any legitimate case of the cognitive penetration of experience undermines the 

alleged informational encapsulation of the relevant perceptual systems, and in turn 

challenges any theoretical architecture of those systems that commits to informational 

encapsulation as necessary for modules. 

Here, finally, is the first horn of the dilemma for modular architectures of the mind. 

There are legitimate cases of the cognitive penetration of experience. We have defended 

three sets of studies against the relevant alternative interpretations. And so perceptual 

systems—in these cases vision—are not informationally encapsulated. Any modularity 

hypothesis about the architecture of perceptual systems that commits to the necessity of 

informational encapsulation (and by implication: cognitive impenetrability) for 

modularity is therefore threatened.  

To clarify our critique, it will be useful to briefly consider a hypothetical defence for 

the encapsulated modularity theorist in response to this first horn of the dilemma. Our 

suggestion is not that perception (or vision, more specifically) is, as it were, 

unencapsulated through-and-through. As discussed above, the entailment relations 

between the cognitive penetration of perceptual experience and the cognitive 

penetration of perceptual processing would not support this last inference. So, the 

modularity theorist might retort, the penetration of experience is compatible with the 



	
   28	
  

impenetrability (and thus encapsulation) of some (but not all) components or systems in 

perceptual processing, which means that some components of perceptual systems may 

be strongly modular.21  

To this, the modularity theorist might add that the modularity approach is a 

conceptual framework for modelling parts of the mind. As such the framework is 

valuable if and to the degree that it usefully explains some features of the human mind. 

And, the above critique notwithstanding, hypothesizing encapsulated modules does 

successfully explain significant aspects of perceptual processing. Consequently, cases of 

cognitive penetration of perceptual experience do not pose a serious threat to 

modularity understood in this way.22 

 Reply: the fact that some aspects of perceptual processing can be explained by 

encapsulated modularity does nothing to save a modular architecture of perception. For 

example, feature detecting components like groups of simple and complex cells in the 

primary visual cortex are likely encapsulated, as are many other neural circuits and low-

level components in the overall visual system. In fact, it may be that certain sub-systems 

in vision—for example, Pylyshyn’s early vision—are encapsulated in spite of the 

penetration of visual experience. This would be to maintain the commitment to 

informational encapsulation and thus a strong notion of ‘module’. But note that the scope 

of modularity on such a view, that is, the kinds of systems to which the conceptual 

framework can be successfully applied, is significantly weakened. Such a modularity 

theorist can only claim that some of the visual system is modular and, importantly, 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
21 We thank XXXX for pressing us to consider this reply for modularity theory. 
22 We thank an anonymous referee for pressing us to consider this response. 



	
   29	
  

cannot claim that vision is, generally, modular. This last claim is inconsistent with 

genuine cases of cognitive penetration. 

It is important, moreover, not to overemphasize a characterization of (perceptual) 

modularity as concerning only perceptual processing. This is because an interest in 

mental architecture is guided not merely by goals of psychological modelling. Another 

crucial issue of relevance is epistemic: a modularity theory of perception promises a 

preferable epistemology, one where perceptual systems rapidly deliver perceptual 

representations in a way not prone (or less prone) to errors introduced by the cognitive 

agent. As Fodor puts the point, the “function of perception is to deliver to thought a 

representation of the world.” And since here the goal is to represent “[n]ot the distant 

past, not the distant future and not…what is very far away…it is understandable that 

perception should be performed by fast, mandatory, encapsulated, etc. systems…” 

(Fodor 1985: 5; emphasis added). The systems in question are sub-personal modules, but 

the representations they provide or give rise to are personal-level experiences. Given 

the epistemic role that such representations are supposed to serve, and the supposed 

epistemic advantage of modular perceptual systems, the modularity theorist should be 

no happier with evidence for penetrated experience than he is with evidence for 

unencapsulated perceptual processing.23      

Where does this leave the view? The claim that some individual low-level circuits are 

encapsulated and thus strongly modular is largely uncontroversial among cognitive 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
23 Fodor makes similar suggestions elsewhere; see, for example, his discussion of “perceptual 
identifications” (1983: 68-71). And Pylyshyn (1980) makes similar commitments, claiming that the 
reliability of perception requires cognitive impenetrability. For further discussion of the epistemic 
consequences of cognitive penetrability, see Lyons 2011; Siegel forthcoming; AUTHOR MASK D. 



	
   30	
  

scientists. And the claim that some sub-systems in perception are strongly modular is 

insufficient to support the claim that the general structure of perception (or, more 

specifically, vision) is strongly modular. In turn, these weakened claims are insufficient 

to secure the putative epistemic benefit of modular perceptual systems. In short, one 

cannot save a modular architecture of perception by appeal to encapsulated perceptual 

components or sub-systems. To do so would be to opt for strength of modules over 

scope, in turn undermining the modularity hypothesis as an architecture of perceptual 

systems.       

 
 
II. Informationally unencapsulated modules: A challenge for the massive modularity 
hypothesis 
 

A number of recent theorists have weakened the notion of modularity with respect to 

Fodor’s original characterization and, in particular, with respect to informational 

encapsulation. This change in the notion of modularity tends to accompany a 

broadening of the scope of modular theories. Thus, massive modularity theorists take 

much if not the whole of the human mind to be modular, including higher level 

conceptual and cognitive systems. If, as we have argued in the previous section, 

informational encapsulation is too strict a requirement on the modularity of perception, 

then it makes sense not to require it of higher-level cognitive systems. Weakening 

modularity in this way, however, comes with significant costs to any modular account 

of cognition. First, it weakens the explanatory value of modular architectures. Second, it 

threatens the internal coherence of modularity theories. 

Peter Carruthers, a massive modularity theorist, argues that  



	
   31	
  

 

if a thesis of massive mental modularity is to be even remotely plausible, then by 

‘module’ we cannot mean ‘Fodor-module’. In particular, the properties of having 

proprietary transducers, shallow outputs, fast processing, significant innateness 

or innate channelling, and encapsulation will very likely have to be struck out. 

(Carruthers 2006: 12; emphasis added.)  

 

According to Carruthers, massive modularists should expect most (if not all) central 

cognitive modules to be unencapsulated. He writes: 

 

…even where a system has been designed to focus on and process a particular 

domain of inputs, one might expect that in the course of its normal processing it 

might need to query a range of other systems for information of other sorts. 

(Carruthers 2006: 10).  

 

In other words, an unencapsulated module, in order to perform its task, will often need 

to compute over information that is made available by other systems. For example, the 

mind-reading system “may need to query a whole range of other systems for 

information relevant to solving the task in hand” (Carruthers 2006: 11).  

Evolutionary psychologists, many of whom subscribe to the massive modularity 

hypothesis, also tend to argue for (or assume) the compatibility of modularity with 

unencapsulation. Hagen (2005) explicitly states what is often implicitly assumed in this 

field: 



	
   32	
  

 

Why, except when processing speed or perhaps robustness is exceptionally 

important, should modules not have access to data in other modules? Most 

modules should communicate readily with numerous (though by no means all) 

other modules when performing their functions, including querying the 

databases of selected modules (163).   

 

Any such modularity theorist thus claims that systems, like the mind-reading system, 

can be modular in spite of being informationally unencapsulated. As Carruthers 

suggests, this might be a necessary adjustment of a general modular architecture for the 

simple reason that anything stronger is implausible.  

One main theoretical advantage of, and indeed motivation for, postulating modular 

architectures is that they explain dissociations between cognitive functions. A cognitive 

task A is said to be dissociated from cognitive task B when some individuals are 

observed who show a significant deficit with respect to A in the absence of a 

corresponding deficit in B. A and B are said to be doubly dissociated when, in addition, 

we observe individuals in whom B is significantly impaired without a corresponding 

deficit in A. Cognitive scientists generally hold that dissociations are signs of functional 

independence, and will often postulate the existence of cognitive modules on the basis 

of these behavioural patterns. If A is observed to fail when B does not, then one may 

infer that A involves a mechanism, or module, M that B does not recruit. When M is 

obstructed, it is argued, A fails and B does not. In the case of a double dissociation, the 



	
   33	
  

inference is stronger, namely that A and B are each performed by a distinct mechanism, 

or module, that the other does not recruit.  

A classic, although by no means uncontroversial, case of the stronger version of the 

inference concerns the face recognition module hypothesis. Many theorists postulate the 

existence of a face recognition module on the basis of a double dissociation between face 

recognition and other cognitive-perceptual functions such as visual object recognition. 

On the one hand, there are well-documented cases in the neuropsychological literature 

of patients who have a face recognition deficit (prosopagnosia) but who appear to retain 

all other cognitive and perceptual capacities, including the ability to respond to and 

identify an array of physical objects (i.e. non-faces) (Rossion et al. 2003). On the other 

hand, there are reports of patients with impaired visual object recognition who appear 

to retain all other cognitive and perceptual capacities, including the ability to respond 

to and identify familiar faces (Rumiati & Humphreys 1997). This is, ostensibly, a double 

dissociation between face and object recognition and it has motivated some theorists to 

posit two corresponding, separate modules: a face recognition module and a visual 

object recognition module (Coltheart 1999: 119). 

This reasoning from dissociation data to modularity—call it the functional modularity 

inference—has been central to the development of modern neuropsychology. In the last 

thirty years, philosophers and cognitive scientists have refined concepts of dissociation 

and narrowed the scope of the inference, and there is an emerging consensus that the 

inference should be understood as an inference to the best explanation, where one infers 

that a cognitive system is modular on the grounds that this hypothesis best explains a 

set of dissociation data, given the fact that modular systems would produce dissociations 



	
   34	
  

if damaged in different ways (Shallice 1988, Coltheart 2001).24 The status of the inference 

as a central methodological tool, however, is very much a matter of debate. Various 

authors have argued, on both theoretical and empirical grounds, that the existence of a 

double dissociation between subjects’ performance on two different cognitive tasks does 

not necessarily constitute strong evidence for the existence of separate cognitive 

functions or modules (Dunn & Kirsner 2003; Juola, & Plunkett 2000; Plaut 1995; Van 

Orden, Pennington, & Stone 2001). We take no side in this debate. Instead, we question 

whether an inference to the best explanation is supported when modules are assumed to 

be unencapsulated.   

Let us suppose, then, as the weakened modularity theory we’re considering does, 

that modules are not encapsulated. Suppose, for example, that the alleged face and 

object recognition modules are not encapsulated, that they both often need to compute 

over information made available by other systems in order to perform their tasks. This 

means that a double dissociation between face and object recognition could occur even if 

both alleged modules remained intact (i.e. were not damaged). This would occur, for 

instance, if both modules need to compute over information normally made available by 

other systems and damage to these other systems (or damage to some pathway between 

them and the two modules) prevents the availability of the needed information. In this 

case, the double dissociation could no longer be taken as a sign that the two modules are 

functionally independent, as both of them could fail to perform their tasks for reasons 

that have nothing to do with a failure of their respective mechanisms. But since it is the 

assumed functional independence of cognitive modules, in the functional modularity 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
24 See Shallice 1988 for a detailed discussion of this methodology. 



	
   35	
  

inference, that is supposed to explain the existence of dissociation data, it is therefore 

difficult to see how the face and object recognition module hypotheses could, in this 

case, best explain the observed double dissociation between face and object 

recognition.25 

By contrast, a double dissociation between face and object recognition is adequately 

explained by encapsulated modularity. Importantly, because the face and object 

recognition modules would in this case be informationally encapsulated, their normal 

functioning would thereby not depend on information made available by other systems. 

A double dissociation between face and object recognition would thus suggest that the 

modules themselves have been separately damaged.26  

This point about the explanatory weakness of unencapsulated modularity is worth 

further emphasis, since it suggests the important distinction between dissociation at the 

level of behaviour (behavioural dissociations) and dissociation at the level of 

mechanisms (mechanistic dissociations). As explained above, unencapsulated modules 

can give rise to behavioural dissociations when there is damage to the systems that 

provide these modules with information required for task performance. But the fact that 

unencapsulated modules can give rise to behavioural dissociations does not imply that 

the mechanisms underlying their function are dissociable.27 In fact, there is conceptual 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
25 The putative theory of mind module is another case, one where unencapsulated modularity would fail 
to explain behavioural dissociations between mind reading and other cognitive capacities. See Gerrans 
and Stone (2008) for an extended discussion of this case. 
26 This is not to say, however, that such an encapsulated modular architecture is the only plausible 
explanation of the double dissociation between face and object recognition. Even with encapsulated 
modularity, the functional modularity inference remains abductive.  
27 The distinction between behavioural dissociation and mechanistic dissociation mirrors, though 
imperfectly, the distinction between perceptual experience and perceptual processing. And just as one 
must be careful in making inferences from (alleged) facts about experience to (alleged) facts about 



	
   36	
  

tension between the notions of unencapsulation and dissociation at the level of 

mechanisms.      

 To see this, consider a minimal conception of modularity. As Carruthers suggests, a 

module must be, at the very least, a “dissociable functional component” (Carruthers 

2006: 2). This minimal conception of modularity is what gives the functional modularity 

inference its theoretical force. Behavioural dissociations are explained by modular 

architectures—and are thus signs of functional independence between cognitive 

mechanisms—since dissociability (at the level of mechanisms) implies functional 

independence and modules are dissociable systems. We agree that a minimal notion of 

modularity should include dissociability, since without it a modular architecture would 

reduce to functional decomposition. And functional decomposition—understanding the 

mind in terms of functional components and sub-components—is uncontroversial as an 

approach, except perhaps in some connectionist quarters.    

 On the one hand, therefore, a cognitive system S is considered functionally 

independent from another system O if S and O’s function can be dissociated. This means 

that S’s function is not affected by what happens to O, and that S can be modified (or 

damaged) without affecting O’s function.  

 On the other hand, S is considered unencapsulated relative to O if S needs to compute 

over information made available by O in order to perform its task. This means that S 

depends on O for its normal functioning. S is thus functionally dependent on O, and is 

therefore not dissociable from O (since dissociability implies functional independence). 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
processing, and vice versa, one must be careful about making inferences from facts about behaviour to 
facts about mechanism, and vice versa. 



	
   37	
  

In sum, if functional independence is understood in terms of dissociation, as the minimal 

conception of modularity suggests, then S cannot both depend on information provided 

by O and be dissociable from O. 

 An illustration may help. Carruthers explains the minimal conception of modularity 

with the following analogy. 

 

The hi-fi is modular if one can purchase the speakers independently of the tape-

deck, say, or substitute one set of speakers for another for use with the same tape-

deck. Moreover, it counts towards the modularity of the system if one doesn’t have 

to buy a tape-deck at all—just purchasing a CD player along with the rest—or if the 

tape-deck can be broken while the remainder of the system continues to operate 

normally. (Carruthers 2006: 2; emphasis added). 

 

Carruthers goes on to suggest that although operationally distinct in the above ways, 

the components of the hi-fi, once conjoined as a system, do depend upon one another in 

other ways: the CD player requires the amplifier to distribute sound, the speakers 

require input from the amplifier to make sounds, etc. Indeed, some of these dependence 

relations will be asymmetric: the CD player needs the amplifier to distribute sound, but 

not vice versa. The important point to note for present purposes is that in spite of these 

dependence relations, the hi-fi components are (relevantly) computationally 

autonomous: the CD player may require the amp to deliver its output, but it does not 

need to compute over information made available by the amp in reading data off of a 

CD. In other words, in performing its task, it is encapsulated from the amp, the 



	
   38	
  

speakers, and so on. Likewise for other components in the system: the tape-deck reads 

data, the tuner acquires a radio signal, the speakers deliver a range of sounds, and so on, 

all independently.  

 Thus the hi-fi analogy is a useful one, at least for modularity as traditionally 

understood. The trouble is that unencapsulated cognitive modules are relevantly 

disanalogous to hi-fi components. Like the hi-fi modules, the cognitive modules 

envisaged by massive modularity work together, exchanging input and output, and 

often asymmetrically. But unlike the hi-fi components, an unencapsulated module M, as 

per the massive modularity theorist, will often depend for its normal operation on other 

components in the system. And this means that M will not, in these cases, be dissociable 

from these other components. Here again M cannot both depend on information 

provided by other systems and be dissociable from them.  

 This, finally, is the second horn of our proposed dilemma, which challenges 

modularity theorists that expand the scope of modularity by weakening the strength of 

modules so as not to require informational encapsulation. Weakening modularity to this 

degree weakens the explanatory value of modular architectures, which in turn weakens 

the functional modularity inference. Second, the very notion of an unencapsulated 

module appears to be at odds with the core of modularity: conceptualizing modules as 

dissociable functional components. Recall further that this weakened modularity may be 

partly motivated by—in addition to broadened scope—acknowledgement of the 

apparent failure of encapsulated modularity to explain various perceptual phenomena. 

This was the first horn of our dilemma: perceptual systems are not encapsulated 

modules if perception is cognitively penetrated. And there is compelling empirical 



	
   39	
  

evidence for phenomena best explained by cognitive penetration. This concludes our 

proposed dilemma for modular architectures of the mind. Narrow scope-encapsulated 

modularity theories are challenged by the first horn, broad scope-unencapsulated 

modularity theories, by the second horn.   

 

III. Functional independence without modularity 

 

Over the past century and a half, a large body of neuropsychological data, primarily in 

the form of dissociation data, indicates that there are specialized neural circuits in the 

brain and that there are stable relations between these circuits and particular cognitive 

functions.28 In fact, modular and non-modular theorists alike see specialization within 

the brain as an undisputed fact. Both sides, therefore, would agree in substance with 

Norman Geschwind’s account of the general architecture of the brain as “more or less 

specialized groups of cells connected by relatively discrete pathways” (Geschwind 

1965). What is at issue is how best to characterize this specialization. 

 Fodor, as discussed, characterizes the specialization of perceptual systems within the 

strong (encapsulated) modularity framework. According to this approach, individual 

brain areas can be ascribed specific perceptual functions when brain areas constitute 

“domain-specific computational systems characterized by informational encapsulation, 

high speed, restricted access, neural specificity, and the rest.” (Fodor 1983: 101). It is 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
28 More recently, functional neuroimaging data (e.g. fMRI, PET), in the form of selective activations of 
brain areas for certain tasks, also point to the wide range of specialized neural circuits in the brain, 
although the methodology in this case differs from the standard behavioral dissociation logic in 
neuropsychology. 



	
   40	
  

when brain areas can be characterized in this way that, according to Fodor, we should 

expect to find stable (i.e. lawful) relations between structure and function. Or, to put it 

differently, we should expect to find stable relations between particular brain areas and 

specific cognitive functions when brain areas can perform their computational functions 

independently of other brain areas (by virtue of being encapsulated).  

 By contrast, Fodor was much less optimistic about the prospect of finding stable 

structure/function correspondence in the case of unencapsulated computational systems.  

 

Consider, by contrast, [unencapsulated] systems, where more or less any subsystem 

may want to talk to any other at more or less any time. In this case, you’d expect 

the neuroanatomy to be relatively diffuse. At the limit, you might as well have a 

random net, with each computational subsystem connected, directly or indirectly, 

with every other; a kind of wiring in which you get a minimum of stable 

correspondence between neuroanatomical form and psychological function. (Fodor, 

1983: 118).    

 

 On Fodor’s model, you get neural specificity, and thus stable structure/function 

correspondence, only when cognitive systems perform their computations 

autonomously and locally. In the case of unencapsulated systems, the computational and 

informational resources needed to perform the task at hand are distributed across a wide 

range of systems, which is why we should not expect to find stable correspondence 

between unencapsulated computational systems and specific cognitive functions.  

 In the previous section, we argued that there is tension between the minimal 



	
   41	
  

conception of modules as dissociable functional components and the idea that modules 

can be unencapsulated. The argument was that unencapsulated systems are not 

functionally independent. And since dissociability implies functional independence, 

unencapsulated systems cannot be dissociable. 

 These claims – Fodor’s claim that stable structure/function correspondence requires 

encapsulation and our claim that modules cannot be both unencapsulated and 

dissociable – are intimately related. Both stable structure/function correspondence and 

dissociablibity, it seems, require functional independence, but unencapsulated systems, 

as we saw, are not functionally independent.  

 We now argue that functional independence can be defined in such a way that 

unencapsulated systems are functionally independent. We then argue that the resulting 

notion of functionally independent system, while considerably weaker than the minimal 

conception of a module, is sufficiently strong to characterize stable structure/function 

correspondence.  

 To see how a system S can be functionally independent from another system O 

without being encapsulated from O, consider the case of Broca’s area (BA). This area is 

involved in language processing (production and perception of speech), and in order to 

contribute to this cognitive capacity it needs to compute over information that is 

processed and made available by other areas, one of which is the superior temporal 

sulcus (STS) which processes and stores phonological representations (Hickok & Poeppel 

2007). BA is thus unencapsulated relative to the STS. (Notice also that since BA’s 

contribution to language processing would be affected if the STS were damaged, it is 

therefore not dissociable from the STS.) Nevertheless, there is a sense in which BA is 



	
   42	
  

functionally independent from the STS. 

 To see this, let us first introduce a distinction between the low-level computational 

operations, or “workings”, performed by BA, and the higher-level cognitive “uses” to 

which it is put (AUTHOR MASK A-XXXX, Anderson, 2010). We know that BA is put to 

a number of linguistic and non-linguistic uses—for example, it is involved in both 

musical and linguistic syntactic processing, in object manipulation, and in action 

sequencing and action perception (Nishitani et al. 2005; Patel 2003; Schubotz & Fiebach 

2006). This, in turn, has been interpreted as evidence that BA’s contribution to these 

various cognitive uses could be performed by a “reusable” set of low-level 

computational operations, or workings—e.g. sequencing operations in the form of 

“detection, extraction, and/or representation of regular, rule-based patterns in 

temporally extended events” (Schubotz & Fiebach 2006: 501).29 ‘BA’s function’, 

therefore, can be interpreted in two different ways depending on whether one is 

referring to its local workings or to its higher-level cognitive uses.  

 In the light of this distinction, we can now specify the sense in which BA is 

functionally independent from the STS. BA is functionally independent from the STS 

with respect to sequencing operations in the sense that BA performs these operations and 

has the capacity to perform them even if the STS failed to compute anything. 

Generalizing:  

 

(FI) A system S is functionally independent from another system O with respect to 

working W, iff S performs W and has the capacity to perform W even if O failed 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
29 See Fiebach and Schubotz (2006) and Fedorenko et al. (2009) for other proposals of BA’s workings. 



	
   43	
  

to compute anything.30 

 

And accordingly: 

  

(FI system) A system S is an FI system with respect to working W, iff S performs 

W and has the capacity to perform W even if no other systems computed 

anything. 

 

 A few points are in order. First, to say that S has the capacity to perform W even if 

another system O failed to compute anything is to say that S possesses the right kind of 

machinery to perform W given that it is provided with the right kind of information. 

The second conjunct of the condition for functional independence must thus be read 

counterfactually.  

 In the case in which S performs W over information provided by O, this means that S 

has the right kind of machinery to perform W over the kind of information that O 

normally provides. For example, if we assume that BA does in fact perform sequencing 

operations over phonological representations made available by the STS, then to say 

that BA has the capacity to perform sequencing operations even if the STS failed to 

compute anything (and thus failed to make available the relevant information) is to say 

that BA has the right kind of machinery to perform these operations over the kind of 

information that the STS normally provides. (In this case, it is easy to see that BA does 

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
30 This formulation is adapted from the analysis of isolability provided by Lyons (2001): “A substrate S is 
isolable with respect to task T iff S performs task T and could do so even if nothing else computed any 
(cognitive) functions”(289). 



	
   44	
  

have this capacity since, as noted, BA’s workings appear to be put to a wide range of 

cognitive uses in which the STS does not participate, e.g. action sequencing and object 

manipulation.)    

 Second, the notion of FI system is considerably weaker than the minimal notion of a 

module as a dissociable functional component. FI does not require systems to be 

dissociable (as explained above). In fact, the basic cognitive architecture suggested by 

FI—the idea that the mind is composed of FI systems—cannot, in any non-trivial way, 

count as modular. To see this, suppose that modules are no more than FI systems or 

networks of FI systems. Suppose, for example, that BA plus the STS, plus some other 

systems constitute the speech production module. This is to assemble a module out of FI 

systems: decomposing the mind into cognitive capacities and attributing these capacities 

to an assemblage of systems. By this method, any identifiable cognitive capacity could 

be turned, trivially, into a module, since any identifiable cognitive capacity could, 

under this architecture, be implemented by an FI system or a network of FI systems. 

What is trivial is not that our proposed cognitive architecture could potentially apply to 

any identifiable cognitive capacity: this is exactly what a basic functional architecture 

should do. What’s trivial, instead, is calling any FI system or network of FI systems 

capable of implementing a cognitive capacity a ‘module’; labeling each cognitive 

capacity a ‘module’ adds nothing of theoretical import to the functional decomposition 

approach to studying the mind. 

 There is, in addition, another aspect of FI that makes it weaker than the minimal 

conception of modularity. In order for a system S to be functionally independent from 

another system O (according to FI), the requisite condition applies only to one of the two 



	
   45	
  

levels of functional specification—i.e. it applies to workings (at the local level), but not 

to cognitive uses (at the systemic level).  

 For example, assuming again that BA does in fact perform sequencing operations 

over phonological representations made available by the STS, BA is functionally 

independent from the STS with respect to its sequencing operations (local workings), but 

it is not functionally independent from the STS with respect to speech production since 

both areas are jointly put to this cognitive use. To see this, consider a modification of 

(FI) where the working/use distinction is not made. Replacing ‘working’ with the more 

general term ‘function’ we have: 

 

(FI*) A system S is functionally independent from another system O with respect to 

function F, iff S has F and is capable of Fing even if O failed to compute 

anything. 

 

Substituting ‘BA’ for S and ‘STS’ for O we have:  

 

BA is functionally independent from the STS with respect to function F, iff BA has F 

and is capable of Fing even if the STS failed to compute anything. 

 

In this case, the condition is satisfied if F refers to sequencing operations, but it is not 

satisfied if F refers to speech production. 

 Importantly, the distinction between these two levels of functional specification is 

rarely operative in modular theorizing. For example, neuropsychologists typically 



	
   46	
  

attribute cognitive uses (as opposed to workings) to brain areas on the basis of 

dissociation data. Thus, Broca’s area has been characterized as a speech production 

module; the temporoparietal junction has been characterized as a theory of mind (or 

belief attribution) module; the fusiform face area (FFA) has been characterized as a face 

recognition module, and so on.  

 This, in fact, should not be surprising since the functional modularity inference rests 

on behavioral data that are derived from an analysis of the performance of brain-

damaged patients on various cognitive tasks. As such, these data consist of the 

specification of the behavioural consequences of the (mal)workings of various cognitive 

components, which means that they will naturally be expressed in ways that capture 

one or more of the cognitive uses of these components within the larger cognitive 

economy.  

 Given that this is how neuropsychological functional specification has typically been 

carried out (i.e. characterizing the specialization of a particular brain area in terms of the 

its cognitive uses), Fodor’s pessimism about the prospect of finding stable 

correspondence between unencapsulated systems and specific cognitive functions is 

understandable. Indeed, this form of functional specification amounts to characterizing 

structure/use correspondence, but unencapsulated systems, as Fodor pointed out, will 

tend to have a wide range of cognitive uses across different domains. The picture is 

quite different, however, if we conceive of unencapsulated systems as FI systems. FI 

systems are characterized as structure/working correspondences—they characterize the 

specialization of brain areas in terms of their cognitive workings. And since the 

workings of an FI system are both functionally independent (in FI’s sense) and stable 



	
   47	
  

across its different cognitive uses, this form of functional specification allows for the 

characterization of stable correspondence between unencapsulated systems and their 

workings. For example, BA could be characterized as a sequential processor, a 

structure/working specification that remains stable across BA’s wide range of cognitive 

uses.  

 Unencapsulation, therefore, is not an obstacle to the identification of stable 

structure/function correspondence (contra Fodor), as long as the target is 

structure/working, as opposed to structure/use, correspondence. It is the functional 

independence of a system (in FI’s sense), not encapsulation, which determines whether a 

stable relationship can be established between that system and a particular function. So, 

rigorous functional analysis can be performed equally on both encapsulated (Fodorian) 

and unencapsulated (FI) architectures. 

 This is not to say, however, that the functional specification of Fodorian modules and 

unencapsulated FI systems can be approached in the same way. Besides encapsulation, 

these two kinds of systems will typically differ in another important way. Fodorian 

modules, because they are encapsulated, are constrained on the range of information 

they can access in the course of their processing. They perform their computations over 

a restricted class of inputs, which in the case of perceptual systems is a narrow range of 

distal properties (e.g. colors, faces). The specification of Fodorian modules will thus 

typically involve domain-specific (or content-specific) structure/function 

correspondence (e.g. colour-processing and face-recognition modules). The specification 

of unencapsulated FI systems, by contrast, will tend to involve domain-neutral (or 

content-neutral) structure/function correspondence, since the workings of these systems 



	
   48	
  

will typically contribute to a wide range of cognitive uses across different informational 

domains. 

 Modular theorists of all stripes have considered domain-specific functional 

specification to be one of the major strengths of modular theorizing. Knowing, for 

instance, that a particular area of the brain specializes in face-recognition, or belief 

attribution, would tell us something quite specific about how these capacities are 

effected within the larger cognitive economy.  

 Now, what our discussion of FI systems indicates is that domain-neutral specification 

can be equally informative about the organization of cognition, but in a different way. 

The specification of domain-neutral structure/working correspondence can account for 

the contribution of a given brain structure to a wide range of (sometimes seemingly 

unrelated) cognitive capacities. And by the same token, this kind of functional 

specification can expose common structural and computational principles underlying 

seemingly unrelated capacities. For example, there is something quite informative about 

the fact that there appears to be common cognitive mechanisms for object manipulation, 

action sequencing, musical processing, and speech, and that these mechanisms are 

performed by a set of reusable neural structures (in Broca’s area). In fact, the reuse of 

neural circuitry for various cognitive purposes appears to be a central organizational 

principle of the brain31, which means that the specification of reusable FI systems may 

have wide-ranging implications for the study of cognitive architecture.  

	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  	
  
31 Anderson (2010) provides an integrated review of recent neural reuse theories. In particular, see 
Vittorio Gallese’s “neural exploitation hypothesis” (2008), Susan Hurley’s “shared circuits model” (2008), 
Stanislas Dehaene's “neuronal recycling” theory (2005) and Michael Anderson's “massive redeployment” 
hypothesis 
(2007). 



	
   49	
  

 We therefore agree with both Fodor and massive modularity theorists with respect to 

what needs to be explained: specialization within the mind (and the brain) in the form 

of stable relations between brain structures and particular cognitive functions. We have 

argued, however, that both encapsulated (Fodorian) and unencapsulated modularity are 

inadequate as a basis for cognitive architectures.  

 Encapsulated modularity appears incapable of explaining well-evidenced cognitive-

perceptual relations, namely, cognitive effects on perceptual experience—and thus 

processing—supposed to be incompatible with encapsulated perceptual systems.  

Unencapsulated modularity is both too weak to adequately explain what it is supposed 

to explain best, and is at odds with the core conception of modularity. This is the 

dilemma for modularity architectures of the mind. Cognitive specialization, we suggest, 

would be better and more generally explained by a different framework, one based on 

non-modular functionally independent systems. 

 

Acknowledgements 

  

  

Bibliography 

Anderson, M.L. (2007). The massive redeployment hypothesis and the functional 

topography of the brain. Philosophical Psychology, 21(2): 143-174. 

Anderson, M. (2010) ‘Neural reuse: A fundamental organizational principle  

of the brain,’ Behavioural and Brain Sciences: 33: 245-266. 

Balcetis, E. and Dunning, D. (2006) ‘See what you want to see: Motivational influences  



	
   50	
  

on visual perception,’ Journal of Personality and Social Psychology 91(4): 612-25. 

_____(2010) ‘Wishful Seeing: Desired objects are seen as closer’ Psychological  

Science: 21: 147-52. 

Bermudez, J. (1999) ‘Cognitive impenetrability, phenomenology, and nonconceptual  

content,’ Behavioural and Brain Sciences 22(3): 367-8. 

Blum, A. (1957) ‘The value factor in children’s size perception, Child Development 28:  

14–18. 

Boynton, G. M. (2005) ‘Imagining orientation selectivity: Decoding conscious  

perception in V1,’ Nature Neuroscience, 8: 541–542. 

Brewer, W. F., & Lambert, B. L. (2001) ‘The Theory-Ladenness of Observation and the  

Theory-Ladenness of the Rest of the Scientific Process,’ Philosophy of Science 68: 

176-86. 

Bruner, J.S., & Goodman C.C. (1947) ‘Value and need as organizing factors in 

 perception,’ Journal of Abnormal and Social Psychology 42: 33-44. 

Bruner, J. S., & Minturn, A. L. (1955) ‘Perceptual identification and perceptual  

organization,’ Journal of General Psychology, 21-28. 

Bruner J. S., & Postman, L. (1948) ‘Symbolic value as an organizing factor in  

perception,’ Journal of Social  Psychology 27: 203-208. 

Bruner, J.S., Postman, L., and Rodrigues, J. (1951) ‘Expectation and the Perception of  

Color’ American Journal of Psychology 64: 216-227. 

Bruner, J.S., & Rodrigues, J. S. (1953) ‘Some determinants of apparent size,’ Journal of  

Abnormal and Social Psychology 48: 17-24. 

Carruthers, P. (2006) The Architecture of the Mind, Oxford: Oxford University Press. 



	
   51	
  

Carter, L. F., & Schooler, K. (1949) ‘Value need and other factors in perception,’  

Psychological Review 56,: 200-207. 

Churchland, P. M. (1988) ‘Perceptual Plasticity and Theoretical Neutrality: A Reply to  

Jerry Fodor,’ Philosophy of Science 55: 167-187.  

Coltheart, M. (1999) ‘Modularity and Cognition,’ Trends in Cognitive Science 3: 115-20. 

Cowie, F. (2008) ‘Us, Them and It: Modules, Genes, Environments and Evolution,’  

Mind and Language 23: 284-92. 

Dehaene, S. (2005). Evolution of human cortical circuits for reading and arithmetic: The 

‘neuronal recycling’ hypothesis. In: S. Dehaene, J. Duhamel, M.D. Hauser and G. 

Rizzolatti (eds). From Monkey Brain to Human Brain. Cambridge, MA: MIT Press. 

Delk, J. L. and Fillenbaum, S. (1965) ‘Differences in Perceived Colour as a Function  

of Characteristic Color,’ The American Journal of Psychology, 78: 290-93. 

Deroy, O. (forthcoming) ‘Phenomenal contrast without the cognitive penetrability of  

perception,’ Philosophical Studies. 

Dukes, W. F., & Bevan, W. (1952) ‘Size estimation and monetary value: a correlation,’  

Journal of Psychology 34: 43-53. 

Dunn, J. C. & Kirsner, K. (2003). What can we infer from double dissociations? Cortex,  

39(1), 1-7. 

Farah, M. (1994) ‘Neuropsychological inference with an interactive brain: A critique of  

the locality assumption,’ Behavioural and Brain Sciences 17: 43-104.  

Fedorenko, E., Patel, A., Casasanto, D., Winawer, J. and Gibson, T. (2009) Structural 

intergration in language and music: Evidence for a shared system. Memory and 

Cognition 37(1): 1-9. 



	
   52	
  

Fiebach, C. J. & Schubotz, R. I. (2006). Dynamic anticipatory processing of hierarchical 

sequential events: a common role for Broca's area and ventral premotor cortex 

across domains? Cortex, 42(4), 499-502. 

Fodor, J. (1983) Modularity of Mind, Cambridge, MA: MIT Press. 

_____(1985) ‘Precis of The Modularity of Mind’, The Behavioural and Brain Sciences 8: 1- 

5 

_____(1988) ‘A reply to Churchland's “Perceptual plasticity and theoretical  

neutrality,”’ Philosophy of Science 55: 188-19 

Gallese, V. (2008). Mirror neurons and the social nature of language: The neural 

exploitation hypothesis. Social Neuroscience, 3(3-4): 317-33. 

Gentaz, E. and Rossetti, Y. (1999) ‘Is haptic perception continuous with cognition?’  

Behavioural and Brain Sciences 22(3): 378-9.  

Gerrans, P. and Stone, V. E. (2008) ‘Generous or parsimonious cognitive architecture? 

Cognitive neuroscience and theory of mind’. British Journal for the Philosophy of 

Science 59: 121-41.  

Geschwind, N. (1965). Disconnexion syndromes in animals and man. Brain, 88, 237- 

294, 585-644. 

Goodale, M. A. and Milner, D. (1992) Separate visual pathways for perception and 

action. Trends in Neurosciences 15(1): 20-25. 

Hagen, Edward H. (2005). Controversial issues in evolutionary psychology. In D. M.  

Buss (Ed.), The handbook of evolutionary psychology. (pp. 5-67). Hoboken: John 

Wiley & Sons. 



	
   53	
  

Hickok, G. and Poeppel, D. (2007). The cortical organization of speech processing. 

Nature Reviews Neuroscience, 8, 393-402. 

Holzkamp, K. and Perlwitz, E. (1966) ‘Absolute oder relative Größenakzentuierung?  

Eine experimentelle Studie zur sozialen Wahrnehmung,’ Zeitschrift für 

experimentelle und angewandte Psychologie 13: 390-405. 

Hurley, S. (2008). The shared circuits model (SCM): How control, mirroring, and 

simulation can enable imitation, deliberation, and mindreading. Behavioral and 

Brain Sciences, 31: 1-58. 

Juola, P. & Plunkett, K. (2000). Why double dissociations don't mean much. In G.  

Cohen, R. A. Johnston, & K. Plunkett (Eds.), Exploring cognition: damaged brains 

and neural networks. (pp. 319-327). Psychology Press. 

Kamitani, Y., & Tong, F. (2005) ‘Decoding the visual and subjective contents of the  

human brain,’ Nature Neuroscience 8: 679–685. 

Klein, G. S., Schlesinger, H. J., & Meister, D. E. (1951) ‘The effect of personal values on  

perception—an experimental critique,’  Psychological Review 58, 96-112. 

 

Lyons, J. (2001) ‘Carving the Mind at its (Not Necessarily Modular) Joints,’ British  

Journal for the Philosophy of Science 52 (2):277-302. 

_____(2011) ‘Circularity, Reliability, and Cognitive Penetrability of Perception’,  

Philosophical Issues 21(1): 289-311. 

Lysak, W., & Gilchrist, J. C. (1955) ‘Value, equivocality, and goal availability as  

determinants of size judgments,’ Journal of Personality 23: 500-501. 

Machery, E. (2008), ‘Massive Modularity and the Flexibility of Human Cognition,’  



	
   54	
  

Mind and Language 23: 263-72. 

Macpherson, F. (forthcoming) ‘Cognitive Penetration of Colour Experience:  

Rethinking the Issue in Light of an Indirect Mechanism,’ Philosophy and 

Phenomenological Research. 

McCurdy, H. G. (1956) ‘Coin perception studies and the concept of schemata,’  

 Psychological Review 63: 160–168. 

Moore, C. (1999) ‘Cognitive impenetrability of early vision does not imply cognitive  

impenetrability of perception,’ Behavioural and Brain Sciences 22(3): 385-6. 

Nishitani, N., Schürmann, M., Amunts K. and Hari, R. (2005) Broca’s region: From 

action to language. Physiology 20:60–69. 

Noë, A. and Thompson, E. (1999) ‘Seeing beyond the modules toward the subject of  

perception,’ Behavioural and Brain Sciences 22(3): 386-7. 

Patel, A. D. (2003). Language, music, syntax and the brain. Nat Neurosci, 6(7), 674- 

681. 

Payne, K. (2001). ‘Prejudice and Perception: The role of automatic and controlled 

processes in misperceiving a weapon.’ Journal of Personality and Social 

Psychology 8: 181-92. 

Payne, K., Shimizu, Y., Jacoby, L. (2005) ‘Mental control and visual illusions: Toward  

explaining race-biased weapon misidentifications.’ Journal of Experimental  

Social Psychology 41: 36–47. 

Plaut, D. C. (1995). Double dissociation without modularity: evidence from  



	
   55	
  

connectionist neuropsychology. Journal of Clinical & Experimental 

Neuropsychology, 17(2), 291-321. 

Postman, L., Bruner, J. S., & McGinnies, E. (1948) ‘Personal values as selective factors  

in perception,’ Journal of Abnormal and Social Psychology 43: 142–154. 

Pylyshyn, Z. (1980) ‘Computation and cognition: issues in the foundations of  

cognitive science,’ Behavioural and Brain Sciences 3:111-132. 

_____(1999) ‘Is vision continuous with cognition? The case for cognitive  

impenetrability of visual perception,’ Behavioural and Brain Sciences 22 (3):341-

365. 

Rossion, B., Caldara, R., Seghier, M., Schuller, A. M., Lazeyras, F., & Mayer, E. (2003).  

A network of occipito-temporal face-facesensitive areas besides the right middle 

fusiform gyrus is necessary for normal face processing. Brain, 126 (Pt 11), 2381-

2395. 

Rumiati, R. I. & Humphreys, G. W. (1997). Visual object agnosia without alexia or  

prosopagnosia. Visual Cognition, 4, 207-217. 

Shallice, T. (1988) From Neuropsychology to Mental Structure, Cambridge: Cambridge  

University Press. 

Siegel, S. (2011) ‘Cognitive Penetrability and Perceptual Justification,’ Nous 46 (2). 

_____(forthcoming ) ‘The Epistemic Impact of Reasoning in the Basement,’  

Philosophical Studies 

Sperber, D. (1994) ‘The Modularity of Thought and the Epidemiology of  



	
   56	
  

Representations,’ in L.A. Hirschfeld and S.A. Gelman (eds), Mapping the Mind: 

Domain Specificity in Cognition and Culture, New York: Cambridge University 

Press. 

_____(1996) Explaining Culture: A Naturalistic Approach, Oxford: Blackwell. 

_____(2001) ‘Defending massive modularity,’ In E. Dupoux (ed.) Language, Brain and  

Cognitive Development: Essays in Honor of Jacques Mehler, Cambridge, MA: MIT 

Press 

Stefanuci, J.K., and D.R. Proffitt. (2008) ‘Skating down a steeper slope: Fear influences  

the perception of geographical slant.’ Perception 37: 321-323. 

_____(2009) ‘The Roles of Altitude and Fear in the Perception of Height.’ Journal of  

Experimental Psychology 35: 424–438. 

Van Orden, G. C., Pennington, B. F., & Stone, G. O. (2001). What do double  

dissociation prove? Cognitive Science, 25, 111-172. 

van Ulzen, N.R., Semin, G.R., Oudejans, R., Beek, P. (2008) ‘Affective stimulus  

properties influence size perception and the Ebbinghaus illusion’,  Psychological 

Research 72: 304–310. 

Wilson, R. (2008), ‘The Drink You Have When You’re Not Having a Drink,’ Mind and  

Language 23: 273-83. 

Witt, J.K. and T.E. Dorsch. (2009) ‘Kicking to bigger uprights: Field goal kicking  

performance influences perceived size.’ Perception 38: 1328-1340. 

Young, A. W. (1996). Face recognition. In J. G. Beaumont, P. M. Kenealy, & M. J. C.  

Rogers (Eds.), The Blackwell dictionary of neuropsychology. Blackwell. 

AUTHOR MASK A, B, C, D-XXXX