I am grateful for Tania Lombrozo’s, Gualtiero Piccinini’s, and Pierre Poirier and Guillaume Beaulac’s very insightful and challenging commentaries. In this article, I examine their remarks, and I reply to their objections. In section 1, I examine the relation between the philosophy and the psychology of concepts before turning toward the characterization of concepts proposed in Doing without Concepts (henceforth DwC) in section 2. I examine whether exemplars are concepts in section 3; whether prototypes and theories are distinct concepts in section 4; whether it is possible to unify prototypes, exemplars, and theories in section 5; and whether two kinds of theories need to be distinguished in section 6. In section 7, I reiterate an objection against the view developed by Piccinini and by Poirier and Beaulac, while I criticize the dual-process theories in section 8. In section 9, I briefly examine the relation between language and concepts before focusing on definitions in section 10. Finally, in section 11 I turn to my proposal to eliminate “concept.”Footnote 1
1. Concepts in Philosophy and in Psychology
I begin with examining Piccinini’s and Lombrozo’s discussions of the relation between the philosophy and the psychology of concepts.
Piccinini objects to my claim that philosophers and psychologists are typically theorizing about different things when they theorize about concepts, but the considerations he puts forward in support of his criticism are not convincing. He asserts that I did not clarify adequately what philosophical and psychological theories of concepts are. However, while I did not define “philosophical” and “psychological,” I fleshed out the distinction between philosophical and psychological theories of concepts by giving paradigmatic examples of both types of theories: Peacocke (1992), Fodor (1998), and Rey (1983, 1985) on the one hand; Rosch and Mervis (1975), Hampton (1979), Murphy and Medin (1985), and Nosofsky (1992) on the other. So the claim put forward in DwC is that the theories developed by the former and by the latter group of researchers are not about the same things.
Piccinini also mentions that some philosophers (e.g., Fodor, 2008) take themselves to be theorizing about what psychologists are theorizing (he also could have noted that some psychologists—see, e.g., Lombrozo’s commentary—take themselves to be theorizing about what philosophers are theorizing), but this carries little weight since the point of the arguments made in DwC is precisely to correct Fodor’s and others’ misconceptions about the respective objects of the psychology and philosophy of concepts. He also notes that philosophers working on concepts are interested in the study of cognition, as are psychologists working on concepts, but this obviously does not mean that the latter and the former are theorizing about the same things. Psychologists working on reasoning and psychologists working on episodic memory are not working on the same things, despite the fact that both groups of psychologists are working on cognition.
Lombrozo holds that, while psychologists and philosophers have different research questions, they are theorizing about the same things—a particular kind of mental representations—and are merely interested in different aspects of these representations (see also Margolis and Laurence, 2010). Note first that Lombrozo agrees with me that philosophers are mistaken when they object to psychological theories that fail to explain what philosophers strive to explain (and vice versa). Second, I doubt that philosophers and psychologists merely are studying different aspects of the same things. If this were correct, then when prototype theorists say that concepts are prototypes or when exemplar theorists say that concepts are sets of exemplars, what they say would be literally mistaken. The principle of charity suggests that those psychologists are better understood as theorizing about different things than philosophers.
2. The Notion of Concept
Piccinini and Lombrozo are not satisfied with the characterization of the notion of concept provided in DwC. I discuss their concerns in this section.
Piccinini (see his note 3 and the beginning of his commentary) provides two alternative characterizations of the notion of concept: “building blocks of thoughts” and “representations posited to explain certain cognitive phenomena.” However, these two characterizations are not equivalent: It is far from obvious that the representations posited to explain “recognition, naming, inference,” etc., are the building blocks of thought (if Piccinini uses this expression in its usual sense). Second, Piccinini’s as well as Lombrozo’s appeal to the metaphor “building blocks of thoughts” is unfortunate since, as noted in section 1.4.3 of DwC (p. 26), this metaphor is hopelessly vague (for further discussion, see Machery, 2010).
Piccinini’s criticisms of my characterization of the notion of concept (see C in the précis) are unconvincing (note 3). Because no citation is offered to support the claim that psychologists are willing to call “concepts” “structures that are not stored in long-term memory,” this claim is hard to evaluate. Although the notion of a higher cognitive competence is indeed vague (see discussion pp. 8-9 of DwC), it is fleshed out by means of examples throughout the book. Finally, Piccinini is right to be concerned about the use of the notion of judgment in C since some of the processes underlying higher cognitive competences might produce outputs that are not well characterized as judgments. However, the use of the notion of judgment in C is not really essential: “judgment” can be replaced with “output.”
Lombrozo questions whether concepts can be identified with bodies of knowledge since these bodies of knowledge have themselves to be represented (“articulated” as she puts it), and since these representations are likely to bottom out in representations that are not bodies of knowledge (see also Rupert, forthcoming). For instance, suppose that one identifies water with the prototype that water is typically transparent, liquid, tasteless, etc. Then, this prototype itself must be represented, including by means of representations of the property of being transparent, the property of being liquid, and so forth. And if these are bodies of knowledge, then these bodies of knowledge require further representations until we reach representations that are not bodies of knowledge.
Lombrozo’s concern is serious indeed. However, one should be cautious in inferring the structure of the vehicle of concepts (how they are represented or articulated) from a description of their content (what knowledge they consist in). When a psychologist proposes a model of a concept, e.g., a prototype model of bird, she describes the content of this prototype. The features described in this model (e.g., is feathered, has a beak, etc.) make explicit the information people have in mind about birds. But one should not automatically infer from this that the prototype is a complex representation made of component representations corresponding to these features. For instance, from the fact that the content of the prototype of birds includes the information that birds have feathers, one should not ipso facto conclude that the vehicle of this prototype is made of a representation of feathers.
To drive the point home, one can consider the classic connectionist models of concepts developed in the 1980s and 1990s (Clark, 1993). In these models, a tokened prototype is a pattern of activation in the hidden layer. This prototype can be said to represent birds as having feathers, although none of its parts is a representation of feathers. One can also consider the network models of concepts developed in the 1970s (e.g., Collins and Quillian, 1969). In these models, the vehicle of a concept is an atomic node whose content (the body of knowledge it consists in) comes from its inferential or perhaps its causal connections to other atoms. So, in these models, the prototype of birds is an atomic node that stores the knowledge that birds have feathers because it is inferentially or causally connected to the atom for feathers.
The upshot of this discussion is that hypothesizing that concepts are bodies of knowledge (in line with many psychologists’ views of concepts, as Lombrozo acknowledges) does not require postulating representations that are not bodies of knowledge.
Second, Lombrozo proposes that concepts should be viewed as explanantia instead of as explananda. Phenomena such as systematicity, categorization judgments, etc., need to be explained, and we appeal to concepts to explain them. I do not think that Lombrozo and I disagree about this question. The point of developing a theory of concepts is indeed to explain various characteristics of our higher cognition (why we categorize the way we categorize, etc.). However, to be able to do so, one needs to understand the properties of concepts (what knowledge they consist in, what kind of processes is defined over them, what their vehicles are, etc.). For instance, some psychologists (Johnson-Laird, Paivio, Schwanenflugel, etc.) have attempted to explain various characteristics of our understanding of abstract and concrete words (e.g., the concreteness effect) by developing hypotheses about the vehicle of concepts (see, e.g., Paivio’s dual-code theory). Lombrozo’s analogy with atoms is in fact revealing. The properties of atoms (the fact that they are made of electrons rotating around a nucleus, itself made of various particles) explain numerous phenomena atoms are meant to explain.
3. Are Exemplars Concepts?
In contrast to Lombrozo and to Poirier and Beaulac, Piccinini argues that sets of exemplars cannot be concepts on two grounds: (1) For most categories, substances, etc., we have no exemplars, and (2) for some categories the number of exemplars is so large that using them as a concept would not be computationally feasible. None of these arguments, however, should be allowed to undermine the extensive literature by exemplar theorists (which is not discussed by Piccinini).
Before rebutting these two arguments, it is useful to clarify the meaning of “exemplar.” An exemplar is not just any singular representation of an individual object (e.g., my dog Fido), of a particular sample (my ring as a sample of gold), or of a particular event (e.g., my last visit to the dentist). To be an exemplar, a singular representation must also have some specific functional properties. Alone or in combination with other exemplars, it must be used in the processes that underlie our higher cognitive competences. So, if I cannot use the singular memory of my aunt to reason about aunts, categorize something as an aunt, draw inductions about aunts, this memory is not an exemplar of an aunt.
This clarification being made, Piccinini’s first argument suffers from two problems. First, it is unclear why he denies having any exemplar of lily, iguana, etc. I do not think that Piccinini is simply appealing to introspection since, I suspect, he agrees with me that not all concepts are introspectively accessible (DwC pp. 86-8). Indeed, his own views about implicit concepts entail that we do not have introspective access to all concepts. It would seem that he denies having an exemplar of iguana because he does not have “a stable representation of it that distinguishes it from other members of the same category” (my emphasis). This is, however, a bad reason to deny having this exemplar since it is not part of the notion of an exemplar that to have an exemplar of an individual object one must be able to distinguish this individual from other individuals of the same category. Indeed, we can have representations of individuals even if we are not able to distinguish them from members of the same categories. Consider for example the new chair on the porch of my house in Pittsburgh. I would be unable to distinguish it from another chair of the same kind if the latter were substituted from the former, but I do have a singular representation of it all the same (whether it is an exemplar depends on its functional properties).
Second, supposing that it is true that some categories are not represented by exemplars, it follows that exemplars are not concepts only if the natural kind assumption is accepted. If all concepts are of the same kind (the natural kind assumption) and if some categories are represented by concepts, but not by exemplars, then it follows that concepts are not exemplars. But because Piccinini agrees with me in rejecting the natural kind assumption, the conclusion does not follow. Indeed, it might well be that exemplars are concepts, but that some categories are not represented by this kind of concepts.
Piccinini’s second argument does not fare better on at least two grounds. First, it takes any singular representation to be an exemplar, a mistake as we have seen. So, from the fact that I have many singular representations of humans, it does not follow that I have many exemplars of humans. Second, while it is common to appeal to considerations of computational feasibility in philosophy of mind (see, e.g., the discussions of the massive modularity hypothesis), these considerations are typically purely rhetorical since nobody knows what is computationally feasible by the brain (except when an algorithm is computationally intractable).
4. A Prototype-Theory Hybrid?
Piccinini proposes that statistical information about typical properties and causal information are part of the same concept. In effect, he proposes that prototypes and theories form a hybrid concept (on hybrid concepts, see section 3.3 of DwC). Unfortunately, he provides no argument or empirical evidence in support of this claim, appealing instead to parsimony. However, given the poor track record of parsimony in cognitive science (DwC p. 126), I find this consideration unconvincing.
In fact, there is some good evidence that prototypes and theories form different concepts given the individuation criteria proposed in DwC and in the précis. Rehder (2006) has shown that when one draws an induction about a category one uses either some statistical information about the typical properties (giving rise to characteristic typicality effects) or some causal information, but not both (for review and discussion of this experiment, see pp. 205-6 of DwC). This suggests that statistical information and causal information for at least some categories are not connected: Using one of them does not prime the cognizer to use the other. This suggests that prototypes and theories do not form hybrids.
5. Unifying Exemplars, Prototypes, and Theories?
Lombrozo proposes to unify exemplars and prototypes. She neither denies the existence of exemplars and prototypes nor proposes that they are part of exemplar-prototype hybrids. Rather, she proposes that people have an abstraction mechanism that forms more or less abstract representations of categories (exemplars being less abstract representations, prototypes being more abstract representations). This hypothesis is supported by Smith and Minda’s (1998) empirical findings—for at least some category structures subjects acquire first exemplars of the category members before forming prototypes—and by the modeling work of Griffiths and colleagues (2008).
Lombrozo’s proposal is a serious alternative to the heterogeneity hypothesis, and more evidence is needed to decide between these alternatives. I want to emphasize, however, the evidence that speaks in favor of the heterogeneity hypothesis as a counterpoint to the evidence mentioned by Lombrozo. Using visually presented artificial categories, Allen and Brooks (1991) have shown that people can simultaneously acquire and also simultaneously use two distinct concepts—a set of exemplars and something like a definition (discussed pp. 195-6 of DwC; replicated by Smith et al. [1998] in a brain imagery experiment). In an experiment that used words as stimuli, Malt’s (1989) protocol analysis has shown that many participants use both prototypes and exemplars. Studying the bodies of knowledge used by expert physicians to categorize diseases and to reason about them, Brooks and colleagues have argued that expert physicians simultaneously have distinct bodies of knowledge about diseases: a theory, a prototype, and exemplars (for review, see Norman et al., 2006). This latter body of findings is particularly important because it examines people’s concepts in the real world, not the concepts they acquire in the highly artificial conditions of concept learning experiments.
How are we to explain this disagreement between the findings and models presented by Lombrozo and the findings summarized here? The kind of abstraction models presented by Lombrozo account well for the findings in category learning experiments, where participants have to form a concept of a category on the basis of the presentation of category members (for the history of this experimental design, see Machery, 2007). The reason, I speculate, is that, if the explicit goal of an experiment is to learn how to classify artificial stimuli, nothing promotes the acquisition of several concepts (except for particular category structures; see, e.g., Allen and Brooks, 1991): Acquiring a single concept is sufficient to fulfill the experiment.
Lombrozo also argues that theories are connected to prototypes and exemplars in that they determine which properties are represented by prototypes and exemplars (following Murphy and Medin, 1985; see DwC pp. 102-3). On the basis of her remarkable work on explanation (e.g., Lombrozo, 2009), she notes that people’s capacity to explain phenomena—plausibly using theories—influences their similarity computation in subsequent categorization tasks, and thus influences either their prototypes or their exemplars. The idea that theories influence the acquisition and use of prototypes has been implemented in several exemplar models (Carmichael and Hayes, 2001; Rehder and Murphy, 2003; see the discussion in Hayes and Kearney, 2010), but, while these models are interesting, they underestimate the role of theories in higher cognition. The role of theories is not merely to determine which prototypes or exemplars are acquired and used (Lombrozo agrees with this, I think); rather, they are directly used to categorize and to draw inductions, and, as Rehder’s work discussed above suggests and as Lombrozo accepts, theories and prototypes are distinct concepts.
6. Deux types de théories?
Poirier et Beaulac soutiennent que la notion de théorie recouvre deux types très différents d’ensembles de connaissances : d’un côté, les théories causales étudiées en psychologie cognitive développementale par Susan Carey, Elizabeth Spelke, Renée Baillargeon, et leurs collègues, de l’autre, les théories apprises par l’intermédiaire du langage.
Poirier et Beaulac ont raison de noter que ces deux types de théories ont des propriétés différentes. En particulier, elles semblent être acquises de manière différente et nous semblons avoir un accès conscient au contenu du second type de théories causales, mais pas au contenu du premier type.
En revanche, je ne suis pas convaincu que cela justifie l’idée que ces deux types de théories forment deux types fondamentaux de concepts, c’est-à-dire deux types ayant peu à voir l’un avec l’autre. Il me semble plus probable que ces deux types de théories constituent deux sous-types du même type fondamental de concepts — les théories — parce qu’elles ont sans doute de nombreuses propriétés en commun, y compris le fait de contenir le même type de connaissances.
7. Reiterating the Argument against Piccinini
Piccinini and Poirier and Beaulac defend a very similar (though not identical) picture of the organization of our knowledge in long-term memory. Both appeal to an influential kind of theories of cognitive architecture: dual-process theories. According to these theories (which I criticize in section 5.2.2 of DwC), the mind consists of two types of processes or systems, sometimes called System 1 and System 2 processes (for review, see, e.g., Evans, 2008; Evans and Frankish, 2009). Piccinini and Poirier and Beaulac concur that System 2 processes involve linguistic and conscious thoughts and processes. They also agree that a distinction between two fundamental kinds of concepts corresponds to these two types of processes: A first kind of concepts is used in System 1 processes, while a second kind of concepts is used in System 2 processes. While Piccinini rejects (erroneously, as we have seen) the view that exemplars are concepts and speculates (erroneously too) about the existence of prototype-theory hybrids, Poirier and Beaulac accept that exemplars, prototypes, and theories are concepts; but they insist that these are used only in System 1 processes and that a distinct fundamental kind of concepts is used in System 2 processes.
In an earlier paper (2006), I have criticized a previous, less developed formulation of these ideas (Piccinini and Scott, 2006), appealing to the fact that the concepts of meaningless, non-lexicalized categories and those of lexicalized categories have very similar properties (e.g., both give rise to typicality effects). Piccinini replies to this criticism in his comment. Before examining Poirier and Beaulac’s as well as Piccinini’s appeal to dual-process theories, I would like to explain why Piccinini’s reply fails, and, thus, why the objection formulated a couple of years ago still undermines Piccinini’s proposal as well as Poirier and Beaulac’s.
The gist of Piccinini’s reply is the following. As Piccinini rightly mentions, his hypothesis proposes that a lexicalized category (e.g., dogs) is represented by two concepts—one used in linguistic processes and one used in non-linguistic processes. Poirier and Beaulac make a similar claim.Footnote 2 So, the experiment showing that the properties of a concept of dogs are similar to those of a concept of an unlexicalized category (e.g., a category of points or of drawings) does not show that the concepts used in linguistic and in non-linguistic processes have similar properties because these experiments have elicited the concept of dogs that is used in non-linguistic processes.
This response fails for a simple reason. The experiments showing that concepts of lexicalized categories, such as dogs, and of non-lexicalized categories have similar properties are all linguistic. For instance, some tasks involve the application of words to objects (linguistic categorization), and other tasks involve evaluating the strength of inductions that are linguistically formulated. If there really is a kind distinction between concepts used in linguistic and non-linguistic processes, it is plausible to expect linguistic tasks, which rely on linguistic processes, to involve the former kind of concepts. But if this is correct, then—pace Piccinini and pace Poirier and Beaulac—the concepts used in linguistic and non-linguistic processes do have similar properties and cannot form two distinct fundamental kinds of concepts.
In fact, Piccinini’s own views about concepts entail that his reply fails. Piccinini speculates that abstract concepts belong to the kind of concepts used in linguistic processes. The problem is that abstract concepts too have some important properties in common with concepts of unlexicalized categories. In particular, many abstract concepts produce typicality effects (Hampton, 1981).
Poirier and Beaulac deal with my objection somewhat more successfully than does Piccinini. They propose that the System 2 processes have access to the concepts used in non-linguistic processes (these being, in their opinion, prototypes, exemplars, and theories). This proposal might explain why, in linguistic tasks, concepts of lexicalized categories seem to have the same properties as concepts of non-lexicalized categories (although it is not entirely clear to me how it explains this finding). However, it is worth noting that this proposal is a striking departure from the gist of dual-process theories, since typically the bodies of knowledge used by one kind of process (System 1 or System 2 processes) are thought to be inaccessible to the other kind of processes.
8. Les théories postulant deux processus (dual-process theories)
Comme Piccinini, Poirier et Beaulac distinguent deux types de concepts en faisant appel aux théories postulant deux processus. Contrairement à eux, je reste sceptique quant à la valeur de ce type de théories et, en conséquence, je ne pense pas que faire appel à elles soit utile pour développer une théorie correcte des concepts. J’explique pourquoi dans cette sectionFootnote 3.
Les théories postulant deux processus souffrent de deux problèmes principaux. Le premier problème, sur lequel j’insiste dans DwC, est que ces théories sont en général caractérisées de manière très vague. Il n’est pas clair du tout quelles propriétés caractérisent chaque type de processus et les relations entre ces processus ne sont souvent pas clarifiées. Cela est en partie dû à l’habitude regrettable qu’ont les théoriciens postulant deux processus («dual-process theorists») d’assimiler les différentes théories postulant deux processus comme si les différences entre ces théories avaient peu d’importance. Ce premier problème a pour conséquence qu’en règle générale les théories postulant deux processus ont un pouvoir prédictif faible et qu’elles se contentent de fournir des explications post hoc.
On peut avoir différentes attitudes à l’égard de ce premier problème. Piccinini le rejette du revers de la main sans expliquer pourquoi. Dans un article examinant les théories postulant deux processus, Richard Samuels (2009) reconnaît l’existence de ce problème, mais y voit un problème mineur : il fait partie du développement normal d’une science de préciser et de clarifier progressivement les concepts qu’elle utilise. Mais l’imprécision caractéristique des théories postulant deux processus est un défaut bien plus sévère. Ces théories existent depuis maintenant plus de vingt ans et ont inspiré, et continuent à inspirer, de nombreux travaux expérimentaux. Parce qu’elles sont si vaguement formulées, elles sont compatibles avec un grand nombre de données et sont très rarement testées par des tests sévères. En vertu de leur imprécision, ces théories apparaissent donc accumuler un support empirique qui est en partie illusoire. Finalement, Poirier et Beaulac soutiennent que les développements récents des théories postulant deux processus ont résolu ce premier problème. Il est vrai que les théoriciens comme Jonathan Evans (2008) ont fait un effort analytique pour clarifier leurs théories et pour distinguer des variantes différentes plutôt que d’assimiler en vrac ces théories, mais ces efforts en sont à leurs débuts.
Plutôt que de poursuivre cette première critique des théories postulant deux processus, je souhaite insister sur un deuxième problème. Pour ce faire, il est important de clarifier un peu ces théories. Il faut en particulier garder à l’esprit que ces théories n’affirment pas simplement que certains processus sont lents, sous contrôle intentionnel et conscients, tandis que d’autres sont rapides, automatiques et inconscients. Il va de soi que certains processus ont les premières propriétés et que d’autres processus ont les secondes propriétés, et je ne nie nullement une thèse aussi triviale. Ce qui rend intéressantes les théories postulant deux processus, c’est qu’elles combinent les deux thèses suivantes :
1) Les processus «système 1» possèdent un ensemble de propriétés typiques (en sus d’être rapides, automatiques et inconscients, ils sont évolutionnairement ancients, ce sont des processus heuristiques, ils sont contextualisés, ils ne sont pas linguistiques, ils dépendent de la ressemblance et non de règles, ils sont universels et ne varient pas d’un individu à l’autre, ils ne font pas appel à la mémoire de travail, etc.) tandis que les processus «système 2» possèdent un ensemble distinct de propriétés typiques (en sus d’être lents, sous contrôle intentionnel et conscients, ils sont évolutionnairement récents ou même le produit de l’apprentissage plutôt que de l’évolution, ils ne sont pas heuristiques, ils sont abstraits et décontextualisés, ils consistent en l’application de règles, ils varient d’un individu à l’autre, ils impliquent la mémoire de travail, etc.);
2) un processus cognitif quelconque est ou bien un processus «système 1» ou bien un processus «système 2».
Samuels (2009, p. 131) exprime très bien la première thèse :
Dual-process theories do not claim merely that cognitive processes can be divided in terms of single dichotomies on the Standard Menu [la liste des propriétés habituellement attribuées aux processus des systèmes 1 et 2]. No doubt some cognitive processes are more automatic than others, some faster than others, and so on. [A] theory that claimed merely that such distinctions can be drawn — that there are, for example, relatively fast and relatively slow processes — would be a rather uninteresting one. But fortunately, the dual process theorist’s position is far bolder than this. What they claim is that such distinctions line-up: that processes which exhibit one property from a column typically, though not invariably, possess the others.
Je doute que ces deux thèses soient correctes, car je suspecte que les dichotomies lent/rapide, conscient/inconscient, linguistique/non-linguistique, récent/ancient, non-heuristique/heuristique, impliquant des règles/impliquant la ressemblance, etc., ne sont nullement alignées. C’est-à-dire, je soutiens que ces dichotomies sont orthogonales les unes aux autres : certains processus sont lents et non-linguistiques, certains processus sont heuristiques, linguistiques et conscients, certains processus sont ancients et non-heuristiques, et ainsi de suite. Cette objection contre les théories postulant deux processus n’est en rien nouvelle (voir Gigerenzer et Regier, 1996), mais elle n’a nullement été réfutée par les partisans des théories postulant deux processus. Je ne connais guère de travaux cherchant à montrer que ces dichotomies sont alignées, et la recension d’Evans (2008), mentionnée par Poirier et Beaulac, ne fournit pas de données supportant les thèses 1 et 2. En outre, il est très problable que ces dichotomies ne soient pas alignées. Certains processus non automatiques sont anciens et non-linguistiques (Evans, 2008), certains processus automatiques ne sont pas heuristiques (voir par exemple Woodward et Allman, 2009), certains processus automatiques (comme les habilités des experts) sont acquis par apprentissage, et ainsi de suite.
Où cela nous mène-t-il ? Dans la mesure où les théories postulant deux processus supposent que les thèses 1 et 2 sont correctes, je ne vois guère de raisons de les accepter et de baser une théorie des concepts sur elles, comme le font Poirier et Beaulac et Piccinini. Bien sûr, on pourrait être tenté d’embrasser une version minimaliste des théories postulant deux processus, selon laquelle ces théories affirmeraient simplement qu’il y a des processus qui sont lents, sous contrôle intentionnel et conscients et d’autres qui sont rapides, automatiques et inconscients. Mais, tout d’abord, il semble qu’aussi bien Poirier et Beaulac que Piccinini affirment davantage que cela. Leur commentaire soutient par exemple que les processus «système 2» sont linguistiques, tandis que Piccinini ajoute qu’ils sont spécifiquement humains. Ensuite, il est douteux que l’on puisse baser une distinction entre deux types fondamentaux de concepts sur une théorie postulant deux processus comprise d’une manière si minimale. Il ne va pas de soi que les processus lents, sous contrôle intentionnel et conscients et les processus rapides, automatiques et inconscients impliquent deux types de concepts fondamentalement différents.
Mettons de côté pour un instant ces doutes généraux sur la valeur des théories postulant deux processus. Les versions particulières de ces théories mises en avant par Piccinini et par Poirier et Beaulac rencontrent aussi des difficultés. Piccinini spécule que les processus «système 2» sont linguistiques, spécifiquement humains et d’apparition récente dans l’évolution humaine, tandis que les processus «système 1» sont non-linguistiques, présents chez d’autres animaux et anciens. Malheureusement, ces spéculations sont peu probables, comme le reconnaît Evans, l’un des partisans des théories postulant deux processus. Il note ainsi (2008, pp. 259-260) :
[T]here is considerable evidence of a distinction between stimulus-bound and higher-order control process in many higher animals (…), including rodents, which could be seen as the biological foundations for the System 1 and 2 cognition in humans. (…) The proposal that System 1 cognition is evolutionarily old and shared with other animals is also problematic because it is almost certainly not one system with a single evolutionary history.
Poirier et Beaulac caractérisent aussi les processus «système 2» comme étant linguistiques, mais il est loin d’être certain que la distinction linguistique/non-linguistique s’aligne avec la distinction rapide et automatique versus lents et sous contrôle intentionnel. En effet, les recherches sur la cognition numérique exacte montre que des processus non-conscients sont linguistiques (Dehaene et al., 1999).
9. Language and Concepts
Let’s bracket Poirier and Beaulac’s and Piccinini’s appeal to dual-process theories and lets examine the other aspects of their proposals independently of their commitment to these theories. Particularly, one can read their commentaries as proposing, first, that the acquisition of language in the human lineage created a new fundamental kind of concept and, second, that these concepts are definitions. I will first comment briefly on the relation between language and concepts before focusing on the issue of definitions in the next section.
Piccinini speculates that the evolution of the linguistic faculty in the human lineage changed the nature of concepts since these have to possess syntactic properties to be linguistically expressed. While this speculation might seem plausible, it is entirely unclear whether it is correct.
First, it overlooks the fact that representations could have linguistic properties even in non-linguistic creatures since, according to an influential approach to cognition (e.g., Fodor, 1975), mental representations form a language and thinking consists in manipulating these representations according to their formal properties. Perhaps Piccinini takes this approach to be dubious, but he does not explain why.
Second, Piccinini assumes that one can read off the properties of the conceptual expressions expressed linguistically from the properties of the vehicles that express them (i.e., sentences), but this assumption is again far from being obviously correct. Particularly, many theories of language understanding do not think that the conceptual representations expressed by linguistic expressions, such as sentences, possess the syntactic properties of these expressions. This is the case, e.g., of Johnson-Laird’s theory of mental models, of most theories in cognitive linguistics (e.g., Talmy’s), and of the neo-empiricist approach to lexical semantics (e.g., Glenberg, 1997).
More generally, the speculation that the evolution of language in the human lineage resulted in a fundamental modification of our conceptual repertoire is seductive and commonly endorsed, but it is supported by very little evidence. There is no doubt that having a language has a tremendous impact on which concepts people have because linguistic communication enables people to acquire numerous concepts and because it also enables cultures to develop and to transmit to their members concepts that would have not been developed without linguistic communication. But acknowledging this does not amount to recognizing that the evolution of language resulted in the evolution of a new, distinct fundamental kind of concepts.
10. Definitions
Let’s now turn to the issue of definitions. First, I need to acknowledge that Poirier and Beaulac’s criticism of one of my arguments against definitions—that postulating definitions does not explain the psychological phenomena discovered since the 1970s, see pp. 80-1 of DwC—is correct. This argument does assume that concepts form a natural kind, an assumption I explicitly reject.
Second, Poirier and Beaulac are quite right that some terms are associated with definitions. Some sciences (not all, however: consider psychology and ecology) do provide definitions of the properties they are interested in. For instance, chemistry provides a definition of acids, astronomy has recently redefined what a planet is, and mathematics defines what a limit is. Definitions are also to be found in other domains of human life, such as in the legal domain (think of, e.g., the definition of battery) and in sports. I also find it plausible that language plays an important role in human groups’ capacities to develop and transmit definitions, although I am not convinced that other species are unable to form definitions. (I do not see why primates could not learn to classify an object if it is red and square and start learning things and making inductions about the class of red squares.)
So, are definitions a fourth fundamental type of concepts, one whose development is facilitated by humans’ possession of a language? I doubt it. Concepts are meant to be central to our higher cognitive life: Most categorization judgments, most inductions, and most analogies involve them. Their properties explain the properties of our higher cognitive capacities—the ways we categorize, the ways we draw inductions, etc. By contrast, definitions’ place in our cognitive life is marginal. For those few categories (events, substances, etc.) that are defined, their definitions are used in cognitive processing only in particular conditions. Furthermore, their use seems to be only intentional: We need intentionally to decide to use them, and we need to control ourselves to ensure we are using them. Consider for instance the categories of grandmothers and of bachelors. Both are among the few well-defined categories, and people know their definitions. But I suspect people rarely use these definitions in reasoning, to categorize, and so on. Instead, people seem to use prototypes, exemplars, or causal theories of grandmothers or bachelors. Evidence is consistent with these suspicions. Work in terminology is moving away from the assumption that dictionaries should present definitions (Temmerman, 2000); psychological studies of the meaning of predicates often find that these express prototypes (Hampton, 1979; on abstract terms, Hampton, 1981; on emotion terms, Shaver et al., 1984; on disease terms, Bishop and Converse, 1986) or causal theories (e.g., Ahn and Smith, 2008 on disease terms); sociolinguistic studies confirm these results (Labov, 1973); etc. These results suggest that even when words are associated with definitions—which is the exception rather than the general case—definitions are not the bodies of information people tend to rely on, although they can rely on them intentionally in particular contexts.
If definitions are not concepts, what are they then? The discussion here has to become speculative. Consider the fact that the existence of definitions seems to depend on the existence of expert communities such as scientific communities, legal communities, or sport communities. These communities develop definitions (e.g., the definition of a battery, the definition of an offside in football) which provide norms for how the relevant words should be used in some specific contexts (e.g., “battery” in a trial or “offside” on a football pitch). In these contexts, people intentionally use these words according to these definitions. In some domains, particularly in science, these definitions also influence the kind of information people acquire about properties, categories, or substances, a phenomenon well described by Poirier and Beaulac. These phenomena suggest that definitions are better viewed as social norms that people intentionally strive to comply with in specific contexts. But they are distinct from the bodies of knowledge people usually use to reason, categorize, etc. In fact, to comply with definitions in reasoning, etc., people have to prevent their concepts from underlying their judgments; so to speak, they need to take their concepts off-line.
11. Concept Eliminativism
In contrast to Poirier and Beaulac, both Piccinini and Lombrozo are sceptical of the eliminativist conclusion drawn in DwC (see also Machery, 2005). Piccinini puts forward two reasons to resist this conclusion: It’s too early to draw this conclusion, and getting rid of the word “concept” would be impractical (on this second point, see also Mercier, 2010; Strohminger and Moore, 2010). I do not know how much evidence Piccinini thinks is needed and how many decades of research on concepts he deems appropriate. Second, when needed, the description “bodies of knowledge used by default in higher cognition” could be used, which is not particularly ungainly.
Lombrozo correctly notes that the eliminativist conclusion depends on the characterization of concepts I have proposed (viz. C), a characterization she is inclined to question (see section 2, above). Furthermore, she notes that the benefits I argue would follow from the elimination of “concept” follow in fact from the recognition of some form of conceptual pluralism. While this is correct, I suspect that Lombrozo underestimates the costs of keeping “concept” in the theoretical vocabulary of psychology. I suspect that the presence of this term in this vocabulary invites psychologists to look for unified theories of the bodies of knowledge used by default in higher cognition.
Finally, she wonders whether I would recommend eliminating a term like “mental representation,” which is plausibly not a natural kind. However, the case of concepts is very different from the case of, say, representations. Psychologists do not attempt to discover generalizations about representations in general and to encompass these generalizations within a theory of representations, while they do precisely this for concepts. There is thus no theoretical habit to curb in the case of representations. The costs of keeping “representation” and the costs of keeping “concept” are thus altogether different.
