Why reading minds is not like reading words


Written by Brent Strickland and Pierre Jacob.

In a recent review paper in Science (2014. 344-6190) entitled “The cultural evolution of mind reading,” Cecilia Heyes and Chris Frith argue that human children learn to read minds much like they learn to read words, via explicit verbal instruction from knowledgeable adults. On their view, both abilities are inherited culturally as opposed to genetically. Their argument for this thought-provoking analogy rests on three basic claims:

(1) Mindreading exhibits as much cultural diversity as reading words.

(2) The case of word reading shows that dedicated neural structure is not ipso facto evidence for genetic inheritance.

(3) The putative evidence for mindreading in preverbal infants shows the presence of something they call implicit mindreading, which is not genuine mindreading.

Cross-cultural variation

Isn’t there a clear asymmetry between reading words and mindreading? While writing systems were invented some 5,000 years ago, there is no evidence in any culture of a population of healthy adults lacking the basic mindreading ability to ascribe beliefs, intentions and desires to others.

In support of developmental cross-cultural variation, Heyes & Frith mention studies (by Shahaeian et al., 2011) showing different developmental trajectories in the way children from different cultures learn the meanings of some words for psychological states. For example, in heavily verbal tasks, children from cultures with high ‘collectivist’ values (e.g. China and Iran) display an explicit understanding of access to knowledge before an explicit understanding of belief diversity. On the other hand, children from ‘individualist’ cultures (e.g. Australia and the US) display the reverse pattern. This suggests that different languages and cultures highlight different aspects of theory-of-mind. But it would take more than this to undermine the fundamental assumption that people from different cultures have the same basic abilities to ascribe beliefs, intentions and desires to others. After all, this assumption is also not undermined by the fact that neuroscientists and Evangelical preachers speaking the same language are likely to hold different explicit theories about how the mind works or by the fact—if it is a fact—that CIA officers trained in verbal techniques of lie detection are better at detecting deception than untrained people.

Neural specificity

Recent work has shown that a particular area in the human visual cortex, known as the visual word form area, is dedicated to the visual recognition of written words. Literacy, of course, is not genetically inherited. Heyes & Frith’s suggestion is that mindreading is like literacy: the ability to mindread depends on dedicated neural structures, but it is nevertheless culturally (not genetically) transmitted. However, there is a fundamental aspect in which the analogy breaks down. According to Dehaene and colleagues, dedicated neural resources that evolved for object and face recognition have been recycled to allow word reading (Dehaene and Cohen, 2007). But there is no equivalent evidence that the specialized neural resources that underlie mindreading originally evolved for something different from mindreading and were later recycled. Instead, the best current working hypothesis is that those neural areas actually evolved to support mindreading.

Conflicting developmental data

Heyes & Frith’s analogy between learning to read minds and learning to read words faces an obvious developmental objection. While there is evidence that preverbal infants have basic mindreading abilities, children do not typically learn to read fluently until they are at least 5 years old. Furthermore, this latter achievement depends on extensive, laborious and explicit training by knowledgeable adults. However, there is no evidence that children are explicitly and laboriously taught by knowledgeable adults to ascribe intentions, desires and beliefs to others. Nor could children learn to ascribe psychological states to others from hearing stories (e.g. Little Red Riding Hood) read to them by adults (as hypothesized by e.g. Gallagher and Hutto, 2008), since understanding and enjoying such stories clearly presuppose mindreading abilities.

In order to deflate this obvious asymmetry, Heyes & Frith address the conflicting developmental findings, many of which have focused on the ability to ascribe false beliefs to others. On the one hand, the evidence shows that when they are tested on the basis of verbal measures, most children fail false-belief tasks until they are 4,5 years old (cf. Wellman et al., 2001). On the other hand, when they are tested on the basis of non-verbal measures (e.g. looking time or looking behavior), the evidence shows that preverbal infants expect agents to act in accordance with the contents of their true or false beliefs (cf. Baillargeon et al., 2010).

Problems with Heyes & Frith’s solution

Two main strategies have been proposed in the developmental literature to reconcile the discrepant findings: a cultural constructivist strategy and a nativist strategy. Heyes & Frith opt for the cultural constructivist resolution of the discrepant developmental evidence based on a sharp dichotomy between implicit and explicit mindreading. They assume that only findings based on verbal tests can be evidence of genuine, i.e. explicit mindreading and argue that findings about preverbal infants based on non-verbal tests are evidence of merely implicit mindreading.

However, this sharp dichotomy is unconvincing. First, it seems to confuse measures with abilities: one and the same psychological mechanism can be probed with two different measures. Moreover, Heyes & Frith argue that the infant findings based on non-verbal tests are merely evidence of implicit (not explicit) mindreading. So the essential challenge for Heyes & Frith’s strategy is to offer low-level explanations of the findings about preverbal infants consistent with their interpretation of implicit mindreading, but which do not appeal to full blown mindreading. But actually, as they recognize it, Heyes & Frith do not agree about the nature of this implicit mindreading.

According to Heyes, implicit mindreading is not mindreading at all: it depends on what she calls ‘submentalizing’ mechanisms, which are purely associative (cf. Heyes, 2014). On the other hand, Frith maps the distinction between implicit and explicit mindreading onto the distinction between an early developing and a later developing system, both of which are dedicated to mindreading others’ psychological states using different resources (cf. Apperly and Butterfill, 2009). The challenge for Heyes & Frith is to account for the full range of growing data providing evidence that preverbal infants do track the contents of others’ false beliefs. This challenge is made harder by the fact that they do not agree on what psychological mechanisms are really measured by non-verbal tests. On Heyes’ view, all mindreading depends on cultural learning. But on Frith’s two-systems view, the early developing system of mindreading does not seem to depend on cultural learning.

By contrast with the cultural constructivist strategy, advocates of the nativist view deny that only findings based on verbal tests are evidence for genuine mindreading abilities. They take the results about infants at face value and argue that these findings show that a single system of mindreading is operative in development. On the nativist account, children’s cognitive development enriches their mindreading abilities: as their own belief forming mechanism matures and they themselves acquire novel beliefs about richer and more complex subject matters (in particular, via verbal communication with knowledgeable speakers), mindreaders also become able to ascribe to others new beliefs with richer and more complex contents. But the basic mindreading system at work is one and the same.

The burden of the nativist one-system strategy is to explain why false-belief tasks based on verbal tests are so challenging for 3-year-olds. In a typical such false-belief task, participants, who have seen a toy being moved from one location to another in the absence of its owner, are asked to predict where the owner, when she comes back, will look for her toy. As advocates of the nativist one-system account have argued, if young children understand the linguistic meaning of the question, there are at least two basic reasons why they may incorrectly point to the toy’s actual location instead of correctly pointing to where the owner of the toy believes it to be. They may lack the executive resources needed to inhibit the content of their own knowledge of the true location of the toy from intruding into their answer to the question. Or they may be misled by various pragmatic factors into focusing on the toy’s actual location (Helming, Strickland, & Jacob, 2014).


Clearly, learning to read words heavily depends on cultural learning, i.e. on the student’s ability to make sense of the teacher’s verbal and non-verbal communicative acts. But as shown by the pragmatic investigation of human ostensive communication in the past forty years or so (Grice, 1989; Sperber and Wilson, 1986), the recipient of a verbal and non-verbal communicative act could not learn anything from the agent’s ostensive communicative behavior unless he recognized that the agent intended to make some relevant information manifest to him. In short, the student could not learn from the teacher without reading her mind.

Furthermore, much recent evidence shows that preverbal human infants are uniquely sensitive to ostensive signals. For example, human infants selectively respond to another’s direct gaze, to being addressed in motherese and to others’ contingent responses (Csibra and Gergely, 2009). The evidence also shows that early word learning depends on young children’s ability to ascribe communicative intentions to competent speakers (Parise and Csibra, 2012). Indeed, there is compelling evidence that language-acquisition and cultural learning require mindreading abilities. So, while learning to read words wholly depends on cultural learning, mindreading cannot be similarly acquired through cultural learning, since cultural learning itself depends on mindreading.

In summary, while Heyes & Frith’s comparison between word reading and mindreading is original and worthy of consideration, careful reflection shows that only learning to read words, not learning to read minds, can rest on cultural learning. Furthermore, it also shows that unlike word reading, it is most likely that mindreading has a genetic basis and is part of human core cognition.



Apperly, I. and Butterfill, S. (2009) Do humans have two systems to track beliefs and belief-like states? Psychological Review, 116, 4, 953–970.

Baillargeon, R., Scott, R.M. and He, Z. (2010) False-belief understanding in infants. Trends in Cognitive Sciences, 14, 3, 110-118.

Csibra, G. and Gergely, G. (2009) Natural pedagogy. Trends in Cognitive Sciences, 13, 4, 148-153.

Dehaene, S. and Cohen, L. (2007) Cultural recycling of cortical maps. Neuron, 56, 384–398.

Gallagher, S. and Hutto, D. (2008) Understanding others through primary interaction and narrative practice. In Zlatev, J., Racine, T., Sinha, C. and Itkonen, E. (eds) The Shared Mind: Perspectives on Intersubjectivity.Amsterdam: John Benjamins (pp. 17-38).

Grice, P. (1989) Studies in the Way of Words. Cambridge, MA: Harvard University Press.

Helming, K.A., Strickland, B., and Jacob, P. (2014) Making sense of early false-belief understanding. Trends in Cognitive Sciences, 18, 4, 167-70.

Heyes, C. (2014) Submentalizing: I am not really reading your mind. Perspectives on Psychological Science, 9, 131–143.

Heyes, C. and Frith, C. (2014) The cultural evolution of mind reading. Science, 344, 1357-1361.

Parise, E. and Csibra, G. (2010) Electrophysiological Evidence for the Understanding of Maternal Speech by 9-Month-Old Infants. Psychological Science,
23, 7, 728–733.

Perner, J., & Ruffman, T. (2005) Infants’ insight in to the mind: How deep? Science, 308, 214-216.

Shahaeian, A., Peterson, C.C., Slaughter, V. and Wellman, H.M. (2011) Culture and the sequence of steps in theory of mind development. Developmental Psychology, 47, 1239–1247.

Sperber, D. and Wilson, D. (1986) Relevance, Communication and Cognition. Cambridge, MA: Harvard University Press.

Wellman, H.M., Cross, D. and Watson, J. (2001) Meta-analysis of theory of mind development: the truth about false belief. Child Development, 72, 655-684.


  • Dan Sperber
    Dan Sperber 22 January 2015 (23:51)

    Yes, the analogy between reading minds and reading words put forward by Cecilia Heyes and Chris Frith has the lightness of a glass of Prosecco: pleasurable but with a short finish. Let me add to the disanalogies that Brent and Pierre convincingly outline one that may not be as easy to recognise as such but that I see as quite fundamental.

    Learning to read is a matter of categorising shapes as letters and combinations of letters: varieties of “a,” “a” and “A” as tokens of the letter A, and so on. It is not a task radically different from that the task of learning to categorize shapes as instances of categories such as geometric forms (circles, triangles, and so on), as conventional depictions of stars, arrows, and so on, or as patterns typical of specific minerals, plants, or animals. Reading recruits visual discrimination abilities that are already there and already deployed for a variety of categorisation tasks.

    Mindreading requires task-specific resources. Imagine a species that, to start with, doesn’t have representations in its ontology, or any way to metarepresent representations in terms of their content. No way to represent: ‘Leila believes that the cat is on the mat,’ that is, no way to re-use with a novel representational function whatever mental resources are used to represent that the cat is on the mat in order to metarepresent the content of Leila’s belief as such. If you don’t see the problem, envisage the possibility that there may be — I believe that there is — a problem, a big problem that you don’t see.

    I have never seen any account of how you could learn to metarepresent the content of representations. The common idea that you might teach an organism that doesn’t have metarepresentational capacities how to metarepresent contents is hardly less preposterous than the idea that you could teach an organism that sees in black and white and doesn’t have the wherewithal to perceive colours how to see colours. In both cases, there is a gap in representational resources. The gap may well be lesser for metarepresentations than it is for colours, but it is a gap all the same, and again, there is no story on offer as to how this gap can be filled via teaching and learning. Teaching and learning how to read doesn’t have any such gap, not even a small one, to overcome. This, I suggest, is another and very serious disanalogy between reading minds and reading words.

  • Pierre Jacob
    Pierre Jacob 23 January 2015 (17:40)

    Yes, Dan, thanks for your illuminating comments. One point though in mitigation of your powerful argument against the learnability/teachability of the ability to metarepresent the content of representations for creatures lacking representations in their ontology.

    Data (by Senju et al.) showing that Asperger patients who fail to gaze accurately at the location where a mistaken agent is likely to act (in anticipation of the agent’s action) might be taken as evidence that to some extent they to lack (or lacked) mental representations in their ontology at some point at which healthy developing individuals did not. Nonetheless the evidence also suggests that they can pass elicited-response FB tasks, which may be interpreted as evidence that language may have helped them to some extent to learn to metarepresent the contents of others’ mental representations.

  • Pierre Jacob
    Pierre Jacob 23 January 2015 (18:03)

    Right Brent. Interesting, but it’s not entirely clear to me how the view that Asperger patients lack social motivation can apply to the relevant dissociation here.

    I don’t think that if you understand the content of non-verbal stimuli in an anticipatory gaze False Belief scenario, then gazing at the likely location of the mistaken agent’s action in anticipation of the act is under the control of the participants’ motivation. It would if anticipatory gaze was a voluntary action, which it is not, it seems to me.

    You suggest that being asked a question by an experimenter might motivate them. But on the one hand, I would have thought that Asperger patients are less interested in verbal communication with others than healthy controls. On the other hand, if they lack the motivation to attend to the relevant FB scenario (or stimuli), then they are likely to fail the elicited-response FB task in any case. If so, then being motivated to answer the question would not be sufficient to pass the task if they did not attend to the stimuli for lack of motivation.

  • Dan Sperber
    Dan Sperber 23 January 2015 (19:37)

    I agree with Brent. I don’t think that there is any strong, let alone conclusive evidence to show that Aspergers or for that matter people with regular autism altogether lack metaprepresentational abilities. That they make a less frequent and less reliable use of these abilities, at least in mindreading, is another, orthogonal matter.

    Anyhow, look at the bigger picture. If you have no account of how an organism without metarepresentational abilities could learn to metarepresent representations, then what hypothesis the experimental evidence mentioned by Pierre might be corroborating? That the acquisition of a metarepresentational ability language in people with Asperger who had no such ability to start with is actually taking place through learning language even if we don’t know how? For this you would need evidence on another scale and of another force.

    In any case, your excellent “circularity argument” applies to language learning: you need metarepresentational abilities to acquire language, hence your acquisition of metarepresentational abilities cannot be an outcome of language acquisition.

    Mind you, it is quite plausible, on the other hand, that language acquisition considerably boosts your metarepresentational abilities, but this is another story.

  • Chris Frith 8 February 2015 (18:18)

    I am most grateful to Brent Strickland & Pierre Jacob for their interest in our ideas about mind reading and their scholarly critique of these ideas. I am pleased to have the opportunity to refine my thoughts on the matter. Specifically, I want to discuss the relationship between implicit and explicit mind reading. While I agree with Cecilia that this is an important distinction, we disagree about the precise nature of the implicit version.

    Here I shall argue against two claims made by Strickland & Jacob. First, that there is no sharp distinction between implicit and explicit mind reading. Second, that there are no cognitive processes that are recycled, or built on, to enable mind reading, since the relevant neural processes ‘evolved to support mind reading’.

    Implicit & explicit mind reading are distinct

    I believe that there is compelling evidence that implicit and explicit mind reading depend on independent mechanisms that have different properties (Apperly and Butterfill, 2009). First, there is evidence from studies of autism. High functioning adults with autism can perform normally on a battery of explicit mind reading tasks, while failing to show anticipatory eye gaze in an implicit task (Senju et al., 2009). Second, there is evidence from brain imaging. While right temporal-parietal junction robustly activates during ascription of false beliefs in explicit tasks, this region shows no such distinction in implicit tasks (Schneider et al., 2014). Third, there is evidence that taking account of the mental states of others implicitly involves fast acting processes that are automatic in the sense that they are unaffected by cognitive load (Qureshi et al., 2010). In contrast, explicit mind reading tasks are affected by cognitive load (Apperly et al., 2006).

    Extensive study of the differences between explicit and implicit processes on the performance of skills other than mind reading suggest that explicit versions are not simply implcit versions in a slightly different format. In many cases, the new skill is supported by explicit processes at the beginning of practice, and by implicit processes at the end of practice. Changes associated with practice can be observed in the brain during performance of such tasks. For example, activity in prefrontal cortex, much in evidence at the beginning, has disappeared by the end of practice (e.g. Jueptner et al., 1997).

    Although it has no effect after sufficient practice, cognitive load impairs performance at the beginning of the learning of a motor skill. However, it has no effect on the learning of the skill itself (Eysenck and Thompson, 1966), suggesting that the load has no effect on the implicit component of the task. For classification tasks cognitive load disrupts acquisition of explicit knowledge, but has no effect on acquisition of implicit knowledge (Foerde et al., 2007). These results suggest that the implicit skill does not depend on some automatic version of the process underlying the explicit skill. Instead different processes are involved. In the case of mind reading this distinction seems even more likely given that children start with an implicit version of the skill and subsequently acquire an additional, explicit version.

    Mind reading has evolutionary roots

    Precursors for the many components of mind reading have been proposed. These include, for example, recognition of emotional expressions and the ability to follow eye gaze (Seyfarth and Cheney, 2013). Here I will consider just two precursors for which the associated computational mechanisms are beginning to be understood. In both cases the basic computation involves predictive coding, whereby inferences about the state of the world are continuously updated on the basis of prediction errors (Rao and Ballard, 1999).

    First, there is the basic value learning mechanism that enables us to decide which objects to approach and which to avoid (e.g. Lak et al., 2014) involving the striatum and medial prefrontal cortex (Kable and Glimcher, 2007). We use the same neural systems to learn about the value of objects from observing whether other people approach or avoid them (e.g. Campbell-Meiklejohn et al., 2010 in humans; Kashtelyan et al., 2014 in rats). This system, in particular medial prefrontal cortex, becomes relevant for mind reading when it is used to track the values that other people assign to objects (Garvert et al., 2015). Representing the preferences of others is an efficient means for predicting what they will do since their preferences are likely to apply across many different situations (Robalino and Robson, 2012).

    Second, there are the motor control mechanisms that enable us to choose the appropriate movements to achieve our current, immediate goal. These mechanisms include a prediction of the kinematics of the movement we are about to make (the forward model Miall and Wolpert, 1996). Forward modelling is relevant for mind reading since it enables us to make inferences about the intentions of others from observing the kinematics of their actions (Kilner et al., 2007; Wolpert et al., 2003). On the basis of an estimate of peoples’ intentions we can predict the timing and trajectory of their movements using forward models based on our own motor system (Patel et al., 2012). The resultant prediction errors allow us continuously to refine our estimates of other peoples’ intentions.

    I suggest that these are cases where existing neural mechanisms have been recycled in the service of mind reading. Whether infants are specifically taught to attend to the preferences of others and the kinematics of their movements, is a question to be explored further.

    Implicit and explicit mind reading have different functions

    The major function of implicit mind reading is to predict, from moment to moment, what a particular person is likely to do. Such continuous prediction is important for interactions with others, whether these interactions are collaborative or competitive.

    In contrast, the major function of explicit mind reading is to create and justify explanations of actions, whether our own or those of others. Whether such justifications are convincing will necessarily depend on the beliefs of the culture in which the justifier is embedded. For example, there is the belief, widespread among the people of the Pacific, that it is impossible, or at least extremely difficult, to know what other people think or feel (Robbins and Rumsey, 2008). In such cultures we might expect children to have problems with explicit mind reading tasks (Mayer and Träuble, 2013), but not with implicit ones (Barrett, 2013 #2828).

    Children in Japan are delayed relative to Western children in passing explicit mind reading tasks. When asked to explain why an actor searched in the wrong place for his chocolate, Japanese children talked about the physical situation (where the chocolate was) or social rules (he promised to do so) rather than mental states {Naito, 2006 #3221}. In contrast, Japanese children did not show a delay on a non-verbal version of the task where they pointed to indicate what the actor would do (Moriguchi et al., 2010). These studies suggest that there are much stronger cultural effects on explicit mind reading tasks that involve talking about theories of behaviour than there are on implicit mind reading tasks that involve predicting what people will do, when there is no requirement for justifying the prediction.

    I conclude that there are many good reasons for distinguishing between mechanisms underlying performance on implicit and explicit theory of mind tasks.

    Chris Frith

    Wellcome Trust Centre for Neuroimaging at University College London


    Apperly, I.A., et al., 2006. Is belief reasoning automatic? Psychol Sci. 17, 841-4.

    Apperly, I.A., Butterfill, S.A., 2009. Do humans have two systems to track beliefs and belief-like states? Psychol Rev. 116, 953-70.

    Campbell-Meiklejohn, D.K., et al., 2010. How the opinion of others affects our valuation of objects. Curr Biol. 20, 1165-70.

    Eysenck, H.J., Thompson, W., 1966. Effects of distraction on pursuit rotor learning performance and reminiscence. British Journal of Psychology. 57, 99-106.

    Foerde, K., Poldrack, R.A., Knowlton, B.J., 2007. Secondary-task effects on classification learning. Memory & Cognition. 35, 864-874.

    Garvert, M.M., et al., 2015. Learning-Induced Plasticity in Medial Prefrontal Cortex Predicts Preference Malleability. Neuron. 85, 418-428.

    Jueptner, M., et al., 1997. Anatomy of motor learning .1. Frontal cortex and attention to action. Journal of Neurophysiology. 77, 1313-1324.

    Kable, J.W., Glimcher, P.W., 2007. The neural correlates of subjective value during intertemporal choice. Nat Neurosci. 10, 1625-1633.

    Kashtelyan, V., et al., 2014. Observation of Reward Delivery to a Conspecific Modulates Dopamine Release in Ventral Striatum. Current Biology. 24, 2564-2568.

    Kilner, J.M., Friston, K.J., Frith, C.D., 2007. Predictive coding: an account of the mirror neuron system. Cogn Process. 8, 159-66.

    Lak, A., Stauffer, W.R., Schultz, W., 2014. Dopamine prediction error responses integrate subjective value from different reward dimensions. Proceedings of the National Academy of Sciences of the United States of America. 111, 2343-2348.

    Mayer, A., Träuble, B.E., 2013. Synchrony in the onset of mental state understanding across cultures? A study among children in Samoa. International Journal of Behavioral Development. 37, 21-28.

    Miall, R.C., Wolpert, D.M., 1996. Forward Models for Physiological Motor Control. Neural Networks. 9, 1265-1279.

    Moriguchi, Y., et al., 2010. Japanese children’s difficulty with false belief understanding: is it real or apparent? Psychologia. 53, 36-43.

    Patel, D., Fleming, S.M., Kilner, J.M., 2012. Inferring subjective states through the observation of actions. Proc Biol Sci. 279, 4853-60.

    Qureshi, A.W., Apperly, I.A., Samson, D., 2010. Executive function is necessary for perspective selection, not Level-1 visual perspective calculation: Evidence from a dual-task study of adults. Cognition. 117, 230-236.

    Rao, R.P.N., Ballard, D.H., 1999. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature neuroscience. 2, 79-87.

    Robalino, N., Robson, A., 2012. The economic approach to ‘theory of mind’. Philos Trans R Soc Lond B Biol Sci. 367, 2224-33.

    Robbins, J., Rumsey, A., 2008. Introduction – Cultural and linguistic anthropology and the opacity of other minds. Anthropological Quarterly. 81, 407-420.

    Schneider, D., et al., 2014. Implicit false-belief processing in the human brain. NeuroImage. 101, 268-275.

    Senju, A., et al., 2009. Mindblind eyes: an absence of spontaneous theory of mind in Asperger syndrome. Science. 325, 883-5.

    Seyfarth, R.M., Cheney, D.L., 2013. Affiliation, empathy, and the origins of theory of mind. Proceedings of the National Academy of Sciences. 110, 10349-10356.

    Wolpert, D.M., Doya, K., Kawato, M., 2003. A unifying computational framework for motor control and social interaction. Philos Trans R Soc Lond B Biol Sci. 358, 593-602.

  • Hugo Mercier
    Hugo Mercier 9 February 2015 (10:38)

    This is a fascinating debate, and I’d like to bounce back on one of Chris’ suggestions, namely, that “the major function of explicit mind reading is to create and justify explanations of actions.” I completely agree with this. And yet I don’t think that it follows from this that “whether such justifications are convincing will necessarily depend on the beliefs of the culture in which the justifier is embedded.” A strong version of Chris’ view would be that culture entirely determines what is a good explanation, which would make explanations arbitrary. Not only would this predict much starker cultural variation than that observed, but it would also make explanations altogether moot. Let’s take each point in turn.

    Cultural variation. In spite of the interesting cultural variations pointed out by Chris and Celia, the big picture seems to be one of overwhelming similarity. We won’t find a culture in which people explain eating an apple by hating apples, or, worse, by wanting it to be Monday. Even when a culture’s explanations seem weird for the members of another culture, it is generally because of missing implicit premises. Once these are made explicit, even apparently weird explanations can start to make sense—otherwise, their would be no way to make heads or tails of the behavior from people belonging to other cultures (this is obviously a very naïve take on a long standing discussion in anthropology, but it’s probably close enough to the truth for present purposes).

    The function of explanations. It seems likely that the main function of explaining behavior is to affect others’ evaluations of one’s (or others’) behavior. This means that people must be able to evaluate these explanations, in order to evaluate others’ behavior appropriately. This means that explanations cannot be entirely arbitrary. Arbitrary explanations could only be good explanations if the behavior itself was arbitrary—and, plainly, human behavior is, by and large, not arbitrary. Even in games, possible the most arbitrary context, one must make reference, even if implicitly, to non-arbitrary elements, such as a desire to win.

    Instead of arbitrary rules, people rely on their intuitions to evaluate explanations. For instance, when you want tea you intuitively put the kettle on. You can rely on this intuition (among other things) to infer that someone who puts the kettle on might be doing it because she wants tea, or that someone who wants tea is likely to put the kettle on. You can also use this intuition to evaluate someone’s explanation that “I put the kettle on because I wanted to make some tea.” People don’t think that “I put the kettle on because I wanted to make some tea” is a good explanation because it being widespread in a culture. It becomes widespread in a culture because people intuitively find it to be a good explanation.

    This is obviously an extremely simple description of the complex relations between implicit and explicit mindreading, but the main point should stand: explicit mindreading (i.e. producing and evaluating explanations of behavior) has to extensively rely on implicit mindreading. It cannot be an entirely separate mechanism that would mostly rely on arbitrary, culturally determined rules. In turn, this suggests that implicit mindreading is full-blown mindreading (with the ability to attribute false beliefs, etc.), otherwise it could not be relied on for full-blown explicit mindreading.

  • Pierre Jacob
    Pierre Jacob 13 February 2015 (14:09)

    I wish to make two points in response to Chris’s very interesting response. First, I wish to resist endorsing his distinction between implicit and explicit mindreading. Secondly, I wish to comment on the functions respectively assigned by Chris to implicit and explicit mindreading.

    It seems to me that there are two main areas of recent experimental psychological research that are highly relevant to the discussion of what Celia and Chris are inclined to call implicit mindreading: on the one hand, there are developmental findings jointly based on change-of-location false-belief tasks and non-verbal spontaneous-response measures (looking time, looking behavior and helping) suggesting that preverbal human infants can track the contents of others’ false beliefs. On the other hand, there are findings suggesting that human adults can take another’s visual perspective or track the contents of others’ false beliefs automatically, where a computation is taken to be automatic if its output is irrelevant to the task at hand. For example, when asked to judge how many dots either they or an avatar could see, participants were slower and made more errors when they and the avatar could not see the same number of dots than when they could (Samson et al., 2010). This is evidence of automatic visual perspective-taking (in Level 1 perspective-taking tasks) because what the avatar can see is irrelevant to how many dots participants can see. When asked to track the motion of a ball and to detect its presence behind an occluder, adult participants were faster when they expected the ball to be there than when they did not expect it to be there. But they also were faster when an avatar falsely expected it to be there than when neither the avatar nor participants expected it to be there (Kovacs et al., 2010). This last finding suggests that participants automatically tracked the content of the avatar’s false belief because what the avatar believes is irrelevant to whether the ball is behind the occluder. By contrast, there is also evidence from adults that Level 2 visual perspective-taking tasks are not achieved automatically (Surtees et al., 2012).

    While Celia and Chris want, I think, to claim that both cases are instances of implicit mindreading, they disagree about whether they are instances of genuine mindreading. Celia thinks that only explicit mindreading is genuine mindreading. But Chris believes that adult automatic computations of either an avatar’s visual perspective or an avatar’s false belief about a ball’s location are instances of genuine but implicit mindreading. He also believes, I think, that appropriate change-of-location false-belief tasks are evidence of genuine but implicit mindreading by preverbal human infants.

    Chris mentions two kinds of reasons for thinking that “implicit and explicit mindreading depend on independent mechanisms that have different properties”: the dissociation displayed by high functioning adults with autism and the evidence showing that only automatic mindreading is unaffected by cognitive load. I believe that both kinds of evidence can be accommodated by a one-system view of mindreading. First, as Brent and Dan have suggested in earlier discussion, Asperger patients may need to be explicitly prompted by a question in order to use their mindreading abilities. Secondly, there is evidence that while Level 1 perspective-taking in healthy adults is automatic (Samson et al., 2010), Level 2 perspective-taking in healthy adults is not (Surtees, et al., 2012). Far from reflecting the difference between implicit and explicit mindreading, this difference may reflect the difference between tasks whose solution requires, and tasks whose solution doesn’t require, further executive resources.

    Another reason why Chris thinks that implicit and explicit mindreading are different skills is that the “extensive study of the differences between explicit and implicit processes on the performance of skills other than mindreading suggests that explicit versions are not simply implicit versions in a slightly different format.” However, for two complementary reasons, I think that this cannot really support Chris’s distinction because most of the examples he mentions in his response are cases of learning new motor skills where the earlier stage is an explicit process and the final outcome is an implicit one. First, as he of course recognizes, if the implicit/explicit distinction is to apply to mindreading, then the earlier ontogenetic stage must be implicit, while the later ontogenetic stage can either be implicit or explicit. Secondly, the transition from the initial explicit stage to the implicit outcome of the process of learning a motor skill is continuous and graded. But tracking the content of an agent’s false-belief in a change-of-location task is not graded: either one can track the content of the agent’s false belief or one cannot. Arguably, preverbal human infants can but non-human primates cannot.

    As I understand it, the fundamental reason why Chris thinks that implicit and explicit mindreading are different skills is that he subscribes to Apperly and Butterfill’s (2009) two-systems model of mindreading. In fact, Chris wants to map the implicit/explicit distinction onto the two-systems model. Now this model itself is committed to a pair of strongly related and contentious assumptions, one of which is that it disallows any role for learning in the early-developing, efficient and inflexible system, and the other of which is that the early-developing system remains operative alongside the later-developing flexible and inefficient system in older children and adults. In other words, this model is committed to the view that the early-developing system has strong arbitrary signature limits: in particular, Apperly and Butterfill (2009) have argued that tracking the content of another’s false belief about object-identity is such a signature limit. I believe and have argued elsewhere that so far there is no compelling experimental evidence showing such signature limits (Jacob, 2012; 2014). Chris may well disagree with me on this score.

    Not only does Chris think that implicit and explicit mindreading are different skills, but he also argues that they have different functions: while the function of implicit mindreading is to predict, from moment to moment, an agent’s action, the function of explicit mindreading is to create and justify explanations of actions. I don’t find Chris’s argument for the distinction between implicit and explicit mindreading sufficiently compelling to give up a one-system view of mindreading. I am further reluctant to accept Chris’s idea that the function of explicit mindreading is to justify explanations of actions, for two complementary reasons.

    First of all, having the mindreading ability to ascribe psychological states to an agent of an instrumental action might be a necessary but not a sufficient condition for explanation, let alone for seeking and/or offering justifications. Arguably, to offer an explanation of an agent’s behavior is to answer a particular kind of challenge, namely a why-question. To make sense of a why-question requires the pragmatic ability to exclude many possible irrelevant alternatives. Furthermore, the ability to justify is just the other side of the ability to seek justifications for explanations, i.e. to improve and deepen the quality of explanations, in answer to further challenges. The ability to offer explanations of behavior and to justify explanations must rest on the ability to explicitly construe the description of an agent’s action as the conclusion of a reasoning that rests on premises describing the agent’s motivation and epistemic state.

    Secondly, most 4,5-olds who know the location of a toy have been shown to be able to correctly predict where a mistaken agent will look for her toy. This is widely considered to be evidence that 4,5-year-olds have the explicit mindreading ability to ascribe false beliefs to an agent, even if the agent turns out to be a puppet. However, this mindreading ability is clearly not sufficient for having the further reflective ability to recognize that after all the prediction does not make sense because a puppet cannot have beliefs and desires and hence cannot be really looking for her toy. One may attend to the fact that the toy is in the box without further attending to the content of an agent’s false belief that it is in the basket. One may also attend to the content of the agent’s false belief without attending to the further facts that the agent is a puppet and that puppets have no beliefs. The contents of assumptions (which are generated by cognitive mechanisms) can be (more or less) explicitly or implicitly represented (or entertained), but the mechanisms themselves are neither explicit nor implicit.


    Apperly, I. and Butterfill, S. (2009) Do humans have two systems to track beliefs and belief-like states? Psychol Rev. 116, 953-70.

    Jacob, P. (2012) Do we use different tools to mindread a defendant and a goalkeeper? Culture and Cognition Blog (http://www.cognitionandculture.net/home/blog/44-pierre-jacobs-blog/2455- do-we-use-different-tools-to-mindread-a-defendant-and-a-goalkeeper).

    Jacob, P. (2014) Another look at the two-systems model of mindreading. Culture and Cognition Blog (http://www.cognitionandculture.net/home/blog/44-pierre-jacobs-blog).

    Kovacs, A. et al. (2010) The social sense: susceptibility to others’
beliefs in human infants and adults. Science 330: 1830–1834.

    Samson, D. et al. (2010) Seeing it their way: Evidence for rapid and involuntary computation of what other people see. J Exp Psychol Hum 36: 1255–1266.

    Surtees, et al. (2012) Direct and indirect measures of Level-2 perspective-taking in children and adults. Br J of Dev Psy30: 75-86.

  • Pierre Jacob
    Pierre Jacob 8 February 2016 (09:47)

    We are very grateful to Celia Heyes for her gracious reply to our critical discussion of her provocative Science paper with Chris Frith. While the topics under discussion are complex and contentious, Celia Heyes’s reply clearly illustrates the depth of our disagreements. Although we do not agree with her ‘submentalizing’ interpretation of the developmental findings, we do agree with her that in several recent papers, she has indeed offered new arguments that challenge the view (to which we subscribe) that recent developmental findings based on spontaneous-response tasks show that preverbal infants can track the contents of others’ false beliefs.

    Neural recycling which Celia discusses under ‘neural specificity’ is a hypothesis about the neural bases of cognitive mechanisms. We agree with her that Dehaene and colleagues’ work on reading words shows that neural recycling is entirely compatible with the view that learning to read words rests on cultural learning. Whether or not some precise version of the recycling hypothesis will be implemented in the case of mind reading, our main point of disagreement with Celia lies in the implications of our circularity argument for learning to read minds, whose three steps we rehearse here: (i) learning to read words rests on cultural learning. (ii) Cultural learning rests on mindreading. (iii) Therefore, the ability to read minds cannot rest on cultural learning.

    As the conclusion of the circularity argument clashes with Heyes and Frith’s main thesis (that both reading words and reading minds rest on cultural learning), Heyes and Frith cannot accept the conclusion. In accordance with much of her work, the paragraph on ‘conflicting developmental data’ in her response shows that she rejects our second premise: as she puts it, she is more optimistic than we are “about the possibility of learning-from-teaching – even language learning-from-teaching – before mindreading has got off the ground.” Assuming for the sake of the argument that language-learning and cultural learning are independent of mind reading (which we doubt), we would like to make a couple of remarks about what she calls ‘epistemic engineering’, whose purpose is to help children learn to compute others’ mental states by pointing to evidence relevant to the computation of ‘easy’ mental states first.

    The question is: what could be the basis for drawing a distinction between others’ mental states which are respectively easy and hard to compute? As Heyes and Frith argue in their paper, we can draw a distinction between words according to whether learning to read them is easy or hard. But as Dan Sperber has observed, it is highly unlikely that the visual task of learning to read words, which are already part of a speaker’s linguistic vocabulary, can serve as a model for learning to ascribe psychological states to others from scratch. In their paper, Heyes and Frith suggest two entirely different criteria for the distinction between mental states which are easy or hard to learn to ascribe to others. One is that it is easier for young children to learn to ascribe to others desires than beliefs because “children are constantly attempting to fulfill their [own] desires.” The other is that it is easier for children to learn to ascribe to others emotions than beliefs because “emotions are often reflected in distinctive facial expressions.” While the former criterion rests on some first-person priority according to which one learns to recognize the contents of others’ psychological states by their similarity to one’s own, the latter rests on some behaviorist assumption according to which understanding others’ mental states rests on the perception of their behavior.

    While we are not convinced that the two criteria are entirely consistent with each other, we further question the view that others’ motivations (desires or emotions) are in all cases easier to compute than others’ epistemic states. In order to ascribe a desire or an emotion to someone else, one must perform a backward computation from the perception of an agent’s behavior to the agent’s motivation, which is the cause of the perceived behavior. But in order to ascribe a perceptual belief to an agent, one must perform a forward computation from the observed fact to the agent’s belief which is caused by the fact. Finally, in most cases, an agent’s behavior (e.g. reaching for a target) could not be construed as a means of fulfilling her motivation unless the agent had antecedently formed a belief about the target’s location. In a nutshell, you cannot deny the second premise of the circularity argument for free.