2.1. Natural Kind and Induction

The notion of natural kind is essentially enmeshed with the problem of induction (Mill Reference Mill[1843] 1905; Quine Reference Quine and Quine1969; Boyd Reference Boyd1990, Reference Boyd1991; Hacking Reference Hacking1991). For, the notion of natural kind singles out those classes about which nonaccidental, scientifically relevant inductive generalizations can be formulated. With this notion, one draws a distinction between two kinds of classes: those about which inductive generalizations can be formulated, e.g., atoms, and those about which no or few generalizations can be formulated, e.g., things that weigh more than 124 kg. Thus, the members of a natural kind are supposed to share nonaccidentally a large number of (logically unrelated), scientifically important properties (or relations) beside the properties (or relations) that are used to identify them.

Since members of natural kinds have many properties in common, natural kinds are the building blocks of scientific generalizations. Many empirical sciences aim at identifying the natural kinds in their domain in order to develop adequate empirical theories. Psychology is no exception. Psychologists look for classes of entities about which scientifically relevant generalizations can be made. They introduce new notions to pick out kinds that are believed to be natural (e.g., the notion of basic emotion) and they reject notions that turn out to pick out kinds that are not natural (Griffiths Reference Griffiths1997).

2.2. Two Red Herrings: Essences and Laws

The basic idea presented above has been cashed out in many ways. Which account is best suited for my present purpose? Clearly, this account has to be applicable to psychological kinds and to allow many classes to be natural kinds. Otherwise, the denial of the natural kind assumption would simply be trivial.

Two accounts are thus inadequate. According to a common account, natural kinds are characterized by the possession of an essence, that is, a set of intrinsic, causally explanatory properties that are necessary and jointly sufficient to belong to the kind (Putnam Reference Putnam and Putnam1975). However, it is unlikely that psychological kinds possess essences—at least, if the functionalist view is true. And this notion is excessively restrictive: for example, species wouldn't be natural kinds according to this account (Hull Reference Hull1978). One might want to broaden the notion of essence, by including relational properties alongside intrinsic ones. However, the account of natural kind under consideration would remain unsatisfactory. For, it would not distinguish the properties that determine membership from the causally explanatory properties. However, in some natural kinds, the properties that are causally explanatory are not necessary conditions of membership, for not all members possess these properties. The class of political parties in democratic societies, which could be a natural kind in political theory, illustrates this point. The conditions of membership in this class are arguably different from the causal properties that explain why most democratic political parties share many properties.

According to another account, natural kind terms feature in laws, that is, in generalizations that are temporally and spatially unrestricted and that support counterfactuals (Collier Reference Collier and Riggs1996). Now, psychological kinds feature in ceteris paribus generalizations—not in laws (Fodor Reference Fodor1974). And, again, this notion is too restrictive: few kind terms feature in laws. One could propose to replace the notion of law by the notion of ceteris paribus generalization. This would be a step in the right direction, as we shall see below. However, an important element would still be missing from this account, namely the idea that there is at least one causal mechanism that accounts for these generalizations. This is an important aspect of the kinds scientists are interested in for inductive purposes.

2.3. The Causal Notion of Natural Kind

A broader notion has been developed by Richard Boyd (Reference Boyd1990, Reference Boyd1991; Griffiths Reference Griffiths1997).

The core idea of this definition is the following. A natural kind is a class about which many generalizations can be formulated: its members tend to have many properties in common. These generalizations are not accidental: there is at least one causal mechanism that explains why its members tend to have those properties. Finally, this class is not a subset of a larger class about which the same generalizations could be formulated.

One might object that ‘large’ is too vague. However, this complaint is unfounded. For the notion of natural kind, like the notion of heap, is vague. Even if there is no way to sharpen the idea of a ‘large set of properties’, it remains that the number of properties that can be projected from one subset of the class to the whole class is one dimension that distinguishes natural kinds from other classes.

Inversely, one might object that a class of entities that would share only few scientifically fundamental properties would be a natural kind. First, one can question whether such classes really exist. These fundamental properties would plausibly cause the members of these classes to share other properties. Moreover, if there were such classes, I would stick to my guns: they would not be natural kinds. For, they would be noticeably different from the paradigmatic natural kinds. Instead of stretching the notion of natural kind in order to include them, it would arguably be better to distinguish two types of scientific kinds: the natural kinds whose members share many properties and, say, the fundamental kinds, whose members share few fundamental properties.

In the present context, two aspects of the definition proposed above are important. First, the properties that characterize a natural kind are not necessarily possessed by all its members. It is only required that members of a natural kind tend to have these properties. Hence, natural kind terms don't have to feature in laws. It is only required that they feature in ceteris paribus generalizations.

Second, essences are only one of the possible causal mechanisms that explain why members of a natural kind share or tend to share many properties. Boyd (Reference Boyd1990) has insisted upon the homeostasis of properties: in some natural kinds, the occurrence of a property causes the occurrence of other properties and is caused by their occurrence. Other mechanisms are possible, including common descent and social causes (Griffiths Reference Griffiths1997).

2.4. The Natural Kind Assumption

The causal notion of natural kind is relevant for my present purpose. It allows psychological kinds to be natural kinds, since natural kinds do not have to possess essences nor to feature in laws. Moreover, this notion has a large extension: substances, e.g., gold, physical entities, e.g., atoms, species, e.g., dogs, and artifacts do qualify as natural kinds. However, the causal notion of natural kind is not vacuous. It implies that nominal kinds, for instance, the class of physical objects that weigh more than 30 kg, are not natural kinds. For their members don't share many (scientifically relevant) properties. Moreover, it implies that many subsets of natural kinds, e.g., white dogs, are not natural kinds either. For the properties that are true of white dogs are true of all dogs (except being white).

In what follows, I use exclusively the causal notion of natural kind. Hence, the natural kind assumption says that the class of bodies of knowledge that are used by default in our higher cognitive processes possesses three characteristics:

• • There is a large set of properties that these bodies of knowledge tend to possess.

• • These bodies of knowledge possess these properties because of some causal mechanism.

• • This set of properties is specific to this class of bodies of knowledge.

Importantly, psychologists who endorse the natural kind assumption are not committed to deny the existence of differences between concepts or kinds of concepts (Medin et al. Reference Medin, Lynch and Solomon2000). For despite these differences, concepts may share a large set of properties.

2.5. The Received View

For more than thirty years, the cognitive psychology of concepts has been built on the view that concepts constitute a homogenous class about which specific, scientifically relevant generalizations can be formulated. Although some psychologists have recently have voiced some doubts (e.g., Ashby et al. Reference Ashby1998; Hahn and Chater Reference Hahn and Chater1998; Smith et al. Reference Smith, Patalano and Jonides1998), many still endorse this view, particularly prototype theorists (e.g., Hampton Reference Hampton, Hahn and Ramscar2001; Smith Reference Smith2002), exemplar theorists (e.g., Juslin and Persson Reference Juslin and Persson2002; Zaki and Nosofky Reference Zaki and Nosofsky2004), and theory theorists (e.g., Carey Reference Carey2000). Barsalou and colleagues assume that most concepts, if not all, are perceptual symbols (Barsalou Reference Barsalou1999). The recent search for hybrid models of concepts (see Section 6) suggests similarly that psychologists have not renounced the belief that concepts constitute a homogeneous class of representations. Several philosophers of psychology share this commitment. For instance, Prinz (Reference Prinz2002) proposes that most concepts, if not all, are proxytypes, that is, context-sensitive perceptual representations.

Moreover, the received view has driven the methodology of the psychology of concepts and of the higher cognitive functions. In substance, psychologists have tried to find data that were easily accounted for by their pet theory of concepts, but that were problematic for rival theories. Rival theories could account for these data only with ad hoc assumptions. Now, this methodology assumes that the class of concepts is homogenous and, therefore, that all the data about concepts have to be accounted for by a unique theory. This is tantamount to assuming the natural kind assumption.

3.1. The Heterogeneity Hypothesis

Given the causal notion of natural kind, there are three options to rebut the natural kind assumption:

• • It could not be the case that the bodies of knowledge that are used by default in our higher cognitive processes have many properties in common.

• • These mental representations could have many properties in common by accident.

• • They could have those properties because they are a subset of a larger set of representations: no empirical generalization could be made specifically about them.

I propose that the available empirical evidence about our concepts supports strongly the first option: it is not the case that concepts have in common a large set of properties.

Indeed, I will argue that the class of concepts is divided into several kinds of representations that have little in common. I call this claim the heterogeneity hypothesis. There are several readings of this hypothesis. One could claim that different types of categories, say the folk biological categories and the artifact categories, are represented by different kinds of concepts and that these kinds have little in common. For example, there could be few properties common to DOG and TABLE, because in general, concepts of biological kinds and concepts of artifacts could have little in common. One could also claim that different cognitive processes are defined over different kinds of concepts and that these kinds have little in common. For example, a concept of dog used in the categorization process could have little in common with a concept of dog used in induction, because in general, concepts that are used in the categorization process and concepts that are used in the induction process could have little in common. I reject both readings.

Instead, I claim that most categories are represented by several concepts that belong to kinds that have little in common. For example, we have several concepts of dog, each of which belongs to a different kind of concepts. There are few properties common to these concepts because in general, the corresponding kinds of concepts have little in common. I call these kinds the basic kinds of concepts. Each of these concepts of dog can be used in our higher cognitive processes. We can use DOG₁, DOG₂, DOG₃, …, to categorize, to reason inductively, to reason deductively, etc. They qualify as concepts because they are poised to be used in our higher cognitive processes. They are bodies of knowledge that are used by default by our higher cognitive processes. Save for specific circumstances, we rely on these bodies of knowledge, when we decide whether something is a dog, when we reason about dogs, etc. Beliefs about dogs that are not part of these bodies of knowledge can be used in our cognitive processes as well. However, they are not used by default by our higher cognitive processes, when we think about dogs. They are retrieved from long-term memory merely occasionally. This version of the heterogeneity hypothesis is summarized by Figure 1.

Figure 1.

The heterogeneity hypothesis.

This view of our higher cognition stands in sharp contrast with the received view. Our higher cognitive processes are not defined over a homogenous class of mental representations. Thus, the class of concepts is not useful to draw scientifically relevant generalizations about our mental representations. Instead of trying unproductively to characterize the properties that are generally possessed by concepts, psychologists and other cognitive scientists’ explicit goal should be to characterize each basic kind of concepts. I submit that the properties that are characteristic of each basic kind of concepts are not characteristic of the other basic kinds. These basic kinds of concepts are the right level for theorizing about our mental representations.

Some might object that we knew all along that we use different kinds of knowledge in our higher cognitive processes. Isn't it obvious that we can use any kind of knowledge to decide whether something is a dog? Granted. But this objection is beside the point. For the claim is (i) that we have specific bodies of knowledge in our mind that are used by default in our cognitive processes and (ii) that these bodies of knowledge divide into different kinds that have little in common. That we may use any kind of information in our higher cognitive processes does not say anything about the class of bodies of knowledge that are used by default in our higher cognitive processes. Psychologists and some philosophers of psychology aim at formulating generalizations about this class, assuming thereby that it is homogeneous. I deny this claim.

3.2. The Argument from Explanatory Necessity

It is now possible to state the following argument against the natural kind assumption:

The core idea of this argument is straightforward. Concepts are theoretical entities that are posited in order to explain the properties of our higher cognitive processes. In order to provide a satisfactory explanation, we have to posit several kinds of concepts that have little in common. Which ones? At least those that have been assumed by the main paradigms of concepts in the recent psychological literature.

Clearly, this argument is empirical. The truth of Premise 3 is an empirical question. Thus, the whole argument claims that the available empirical evidence about concepts and about the higher cognitive processes supports the denial of the natural kind assumption. Since this argument is empirical, new empirical data could result in a different conclusion.

I will take Premise 1 for granted. Almost all psychologists and many philosophers of psychology (e.g., Prinz Reference Prinz2002) assume it. Concepts are those bodies of knowledge that we use by default in our higher cognitive processes. Their properties explain how we reason, how we categorize, etc. In what follows, I argue for Premises 2 and 3.

4.1. Three Theoretical Entities: Prototypes, Exemplars, Theories

Since the rejection of the classical view of concepts in the 1970s, cognitive psychology has produced three main views, the prototype view of concepts, the exemplar view of concepts, and the theory view of concepts.Footnote ²

The prototype view of concepts is built around the idea that concepts are prototypesFootnote ³ (e.g., Rosch and Mervis Reference Rosch and Mervis1975; Smith et al. Reference Smith, Osherson, Rips and Keane1988; Smith and Minda Reference Smith and Minda2002):

Prototypes single out the properties of the members of the denoted categories in statistical terms. In most prototype models (but not in all), prototypes are assumed to encode some information about the typical properties of the denoted category, that is, the properties that are judged to be possessed by most members of the category. Importantly, these properties are not judged to be necessary. For example, to have a concept of dog is to think by default of dogs as having generally, but not necessarily, such form, as having generally a tail, etc.

According to the exemplar view of concepts, our higher cognitive processes are defined over representations of specific category members, which stand for the whole category (e.g., Medin and Shaffer Reference Medin and Shaffer1978; Nosofsky Reference Nosofsky1988; Juslin and Persson Reference Juslin and Persson2002).

For example, to have a concept of dog is to think by default of dogs as being the class of entities that are similar to Fido, Pluto, and Spot and to use these individual representations to reason about dogs in general.

According to the theory view of concepts, our mental representations are similar to scientific theories and our higher cognitive processes are similar to scientific forms of reasoning (Murphy and Medin Reference Murphy and Medin1985; Carey Reference Carey1985; Keil Reference Keil1989; Gopnik and Meltzoff Reference Gopnik and Meltzoff1997; Keil and Wilson Reference Keil and Wilson2000). The theory view is often put very loosely in the empirical literature. One finds usually two ideas: (i) concepts are enmeshed within theories and (ii) concepts are theories. I am interested in the second claim.

To possess a concept of dog is to think by default of dogs as being the class of entities that satisfy some kind of theory, i.e., some nomological, modal, causal and functional generalizations (e.g., a theoretical concept of dog may say of dogs that they necessarily engender individuals that belong to the same species).

4.2. Prototypes, Exemplars and Theories Have Little in Common

The claim is that these three theoretical constructs have little in common. Now, how does one evaluate how different various kinds of mental representations are? Clearly, exemplars, theories and prototypes have plenty of irrelevant properties in common, for example not being cups of tea. What matters, however, are those properties that psychologists take to be important when they characterize mental representations. Psychologists focus on the representational properties of mental representations and on their functional properties, i.e., what mental representations represent and how they are used in cognitive processes. Hence, if exemplars, prototypes and theories have different representational and functional properties, this would suggest that they have little in common. Of course, other properties, for example neural properties, may turn out to be important to characterize representations. Then, we would have to reevaluate the denial of the natural kind assumption.

I argue first that prototypes, exemplars, and theories are assumed to encode three different types of knowledge. According to exemplar theorists, applying a concept to an individual is tantamount to thinking that this individual is similar to some specific individuals, which are themselves described as having such and such properties. Thus, the mind stores some information about the properties of specific individuals, for example, some descriptions of Fido and Spot.

According to prototype theorists, applying a concept to an individual is tantamount to thinking that this individual possesses to a sufficient degree a sufficient number of typical properties. Thus, the mind stores some statistical information about the properties that characterize classes of individuals (some prototype theories are more complex).

According to theory theorists, applying a concept to an individual is tantamount to thinking that this individual possesses some properties either with some modal force, or because of some law, or for some causal or functional reason. Thus, the mind is storing some modal, nomological, causal, and functional information about classes of individuals.

Notice that for most categories, it is possible to encode some information about the typical properties of their members, some modal, nomological and causal information about their members and some information about the properties possessed by some of their members. Hence, most categories could be represented by a prototype, a theory and a set of exemplars.

I argue now that prototypes, exemplars and theories are assumed to be used in three different kinds of cognitive process. Exemplar theorists and prototype theorists suppose that the higher cognitive processes involve the computation of the similarity between some representations, and respectively, sets of exemplars and prototypes. To make this idea a bit more concrete, consider how a prototype theorist might explain the categorization process. Peter sees a dog. He forms a representation that attributes some properties to this dog. The properties that are attributed are matched with the properties that are represented by several prototypes, including the prototype of dog. The similarity between the dog and each prototype is computed. Since the similarity between the dog and the prototype of dog is higher than the similarity between the dog and the other prototypes, Peter concludes that he is seeing a dog. Various properties of our other higher cognitive processes are similarly explained (see Osherson et al. Reference Osherson1990 on induction; Hampton Reference Hampton1982, Reference Hampton1988 and Sloman Reference Sloman1998 on deduction; Smith and Osherson Reference Smith, Osherson, Vosniadou and Ortony1989 on probabilistic reasoning).

Advocates of the exemplar view would provide a similar explanation of the categorization process, except that the dog would be compared to sets of exemplars, instead of prototypes (Nosofsky Reference Nosofsky1988). Many properties of the other higher cognitive processes could be similarly explained (e.g., Juslin and Persson Reference Juslin and Persson2002).

Advocates of the prototype view and of the exemplar view agree on the role of similarity in our cognitive processes. However, they assume different ways of computing the similarity between a represented object, and respectively, prototypes and sets of exemplars. In brief, prototype theorists assume that the similarity measure is a linear function of the number of properties common to the concept and the object to be categorized, while exemplar theorists take it to be a nonlinear function.

The theory view of concepts assumes that our higher cognitive processes do not rely on similarity. On the contrary, they are supposed to be similar to the reasoning processes that are used in science. Since theory theorists are less concerned with our cognitive processes than prototype and exemplar theorists, their views about our cognitive processes are less developed. Theory theorists often compare our reasoning processes to explanations (Keil and Wilson Reference Keil and Wilson2000) or to inferences to the best explanation (Murphy and Medin Reference Murphy and Medin1985). They generally assume that our cognitive processes follow the standard rules of logic and of probabilistic reasoning (Smith et al. Reference Smith, Langston and Nisbett1992).

Murphy and Medin (Reference Murphy and Medin1985) provide the following toy example. If at a party, a guest jumps in the swimming pool entirely dressed, we may conclude that she is drunk. This categorization does not rest on any matching between our concept of drunk people and our singular representation of this guest. On the contrary, we infer that the most plausible explanation of the behavior of this guest is that she is drunk. This inference is supposed to follow the rules of probabilistic inference.

Finally, it is important to notice that the heterogeneity hypothesis suggests that several cognitive processes underlie categorization, several processes underlie inductive reasoning, and so on. If prototypes, exemplars and theories all exist, and if prototypes, exemplars, and theories are usually used in different types of cognitive process, then we have plausibly several ways to categorize, to draw an induction, and so on. To put it a bit differently, we would have several categorization processes (or several induction processes, etc.), which would use different kinds of representations. This suggestion raises an exciting, but yet understudied question: How do these cognitive processes interact or compete? To illustrate, when we categorize, how do our prototype-based categorization process, our exemplar-based categorization process, and our theory-based categorization process interact, if they do? This is a topic for further research.

I have shown first that prototypes, exemplars and theories are assumed to possess different representational properties. I have also shown that they are assumed to possess different functional properties. Hence, given the properties that are relevant to characterize mental representations, I conclude that the main psychological theories of concepts assume three kinds of mental representations that have little in common. This conclusion supports Premise 2. In other words, I have established the following conditional claim: if it can be shown that prototypes, exemplars, and theories exist and are used by default in our higher cognitive processes, then concepts are not a natural kind.

5.1. Concept Combination

Concept combination is the capacity to form new bodies of knowledge about classes for which we have no concept permanently stored in long-term memory. Those representations, e.g., GRANDMOTHER SPY, are formed on the basis of other representations stored in long-term memory, e.g., GRANDMOTHER and SPY. A concept is simple if and only if it is a permanent body of knowledge in long-term memory. There is a finite stock of simple concepts. A complex concept is not stored permanently in long-term memory, but has to be produced on the fly for various cognitive purposes (e.g., for language comprehension, reasoning, etc.).Footnote ⁶ The claim that concepts combine means that there is a mental mechanism that operates on simple concepts to produce complex concepts (or on less complex concepts to produce more complex concepts).Footnote ⁷

To understand concept combination is to characterize the processes that underlie the creation of a new body of knowledge out of the bodies of knowledge that are stored in long term memory. As a result, we would explain why some information turns out to be represented by complex concepts. The claim defended here is that in order to explain concept combination, one has to assume the storage of prototypes, exemplars, and theories in long-term memory.

5.2. Prototypes and Property Inheritance

When subjects are given two expressions, say “Oxford graduate students” and “carpenter,” and they are asked to list the properties that are typical of Oxford graduate students who are carpenters, they often find the task meaningful and easy (Hampton Reference Hampton1987; Kunda et al. Reference Kunda, Miller and Claire1990; Johnson and Keil Reference Johnson, Keil, Keil and Wilson2000). In some experiments, they are also asked to determine which properties are typical of the two original categories, in our example, Oxford graduate students and carpenters (Hampton Reference Hampton1987; Johnson and Keil Reference Johnson, Keil, Keil and Wilson2000). The psychologist James Hampton has shown that there is a correlation between the properties that are judged to be typical of the members of the original categories, Oxford graduate students and carpenters, and the properties that are judged to be typical of the members of the resulting category, Oxford graduate students who are carpenters (Hampton Reference Hampton1987).

This correlation suggests that the creation of a complex concept involves determining which properties are typical of the denoted category, Oxford graduate students who are carpenters, on the basis of the typical properties of the original categories, Oxford graduate students and carpenters. Now, prototypes are supposed to encode the knowledge about typical properties. Thus, it is plausible that when we produce on the fly a complex concept, we retrieve the prototypes of the original categories and we transmit some of the typical properties they represent to the complex concept. This is known as property inheritance (Hampton Reference Hampton, Lamberts and Shanks1997). The inheritance of properties by the complex concept is (partly, as we shall see) driven by the assumption that the typical properties of the categories denoted by the compounded concepts are also typical of the category denoted by the complex concept. Notice that I am not claiming that complex concepts are prototypes. Although complex concepts do encode some information about the typical properties of the denoted categories, as do prototypes, they may also encode some modal and some causal information—see below. Property inheritance is summarized by Figure 2.

Figure 2.

Property inheritance.

One might object that the results described above tell something about concept combination only if it can be shown that our beliefs about the properties that are judged to be typical of Oxford graduate students who are carpenters constitute a concept, that is, if they are used in our higher cognitive processes. Otherwise, these results would not bear on the creation of complex concepts. Now, several phenomena suggest that these beliefs are indeed used by our cognitive processes. Particularly, subjects are able to evaluate the typicality of individuals qua Oxford graduate students who are carpenters (e.g., Medin and Shoben Reference Medin and Shoben1988). That is, when subjects are given the description of an individual, they are able to determine whether this individual is a typical Oxford graduate student who is a carpenter. This suggests that people use their beliefs about which properties are typical of Oxford graduate students who are carpenters to evaluate the typicality of individuals.

One might reply that these typicality judgments are based on the typicality of these individuals qua Oxford graduate students and on their typicality qua carpenters (Huttenlocher and Hedges Reference Huttenlocher and Hedges1994). That is, we may judge that an individual is a typical Oxford graduate student who is a carpenter, if we judge that she is a typical Oxford graduate student and a typical carpenter. Hence, these judgments may not rely on a complex concept that represents the typical properties of Oxford graduate students who are carpenters. In other words, to judge whether an individual is a typical Oxford graduate student who is a carpenter, we may not need to produce a complex concept, OXFORD GRADUATE STUDENT WHO IS A CARPENTER. This objection fails, however. For, the typicality of items with respect to a complex concept (OXFORD GRADUATE STUDENT WHO IS A CARPENTER) is often not a function of their typicality with respect to the combined concepts (OXFORD GRADUATE STUDENT and CARPENTER). The former is a function of the latter only when the membership in one category (Oxford graduate students) is believed to be independent from the membership in the other (carpenters) (Hampton Reference Hampton1987, 57; Huttenlocher and Hedges Reference Huttenlocher and Hedges1994). Hence, the best explanation of the typicality judgments under consideration is that subjects do produce a complex concept that represents typical properties and use this complex concept to make typicality judgments. Together with other phenomena (Hampton Reference Hampton1982, Reference Hampton1987, Reference Hampton1988, Reference Hampton1996), this discussion suggests that we produce complex concepts by retrieving prototypes from our long-term memory.

5.3. Theories and Property Inheritance

The story is however more complex, for psychologists have shown that the typicality of a property is not the only factor that determines whether it is represented by the complex concept. To illustrate, whether the property being ambitious is typical of Oxford graduate students and of carpenters is not the only factor that determines whether it is represented by the complex concept OXFORD GRADUATE STUDENT WHO IS A CARPENTER. Property inheritance is also influenced by our theoretical knowledge about the categories that are combined (Medin and Shoben Reference Medin and Shoben1988; Murphy Reference Murphy1988, Reference Murphy1990; Rips Reference Rips1995; Johnson and Keil Reference Johnson, Keil, Keil and Wilson2000).

Concept combination uses some modal information about the properties of the original categories, for example, about the properties of grandmothers and of spies. Which properties are represented by the complex concept, GRANDMOTHER SPY, is a function of their modal force. Studies show that if the input concepts represent impossible properties (properties that are typically possessed by the members of one category, say male for spies, but that are believed to be impossible for the members of the other category, say grandmothers), these are never represented by the complex concept; if the input concepts represent necessary properties (properties that are believed to be necessarily possessed by the members of one category, say being the mother of a parent for grandmothers), these properties are always represented by the complex concept (Hampton Reference Hampton1987).

Moreover, concept combination requires some causal information about the properties of the original categories (Medin and Shoben Reference Medin and Shoben1988; Murphy Reference Murphy1988, Reference Murphy1990; Johnson and Keil Reference Johnson, Keil, Keil and Wilson2000). So-called central properties are preferentially represented by complex concepts. A property is central when it explains causally the possession of other properties. For example, being curved is a central property of boomerangs because subjects believe that boomerangs come back because they are curved. Being made out of wood is less central because it explains few properties of boomerangs.Footnote ⁸

These results show that during concept combination, we access some modal and causal knowledge. Now, prototypes do not represent any modal nor any causal knowledge, but only some statistical knowledge. Consequently, during concept combination, we access some knowledge that is not represented by prototypes. Theories are precisely supposed to represent this form of knowledge. Thus, to account for concept combination, we need to posit a second type of concepts, i.e., theories, beside prototypes.

Not so easily, however. For, one could object that the evidence does not show that we have some theoretical concepts, that is, bodies of theoretical knowledge that are used by default in a large number of cognitive tasks. We may simply rely on some background modal and causal knowledge about the categories that are combined. This is indeed how Hampton himself thinks of the use of modal knowledge in concept combination (Hampton Reference Hampton, Lamberts and Shanks1997). Being a prototype theorist, he believes that this modal knowledge is specifically extracted for the purpose of building complex concepts. Hence, this modal knowledge is not used by default in our cognitive processes. I disagree with this interpretation of the evidence. For, we routinely use this very same causal and modal knowledge in other higher cognitive processes (see Keil Reference Keil1989 on categorization; Coley et al. Reference Coley, Medin and Atran1999 on inductive reasoning). The theoretical knowledge that is used in concept combination is poised for being used in many cognitive processes. Thus, the bodies of knowledge that encode this knowledge are concepts.

5.4. Emergent Properties, Exemplars, and Theories

Not all properties that are represented by a complex concept are represented by the original prototypes. The properties not represented by the original prototypes are said to be emergent and this phenomenon is known as property emergence (Hampton Reference Hampton1987, Reference Hampton, Lamberts and Shanks1997; Kunda et al. Reference Kunda, Miller and Claire1990; Johnson and Keil Reference Johnson, Keil, Keil and Wilson2000; Costello and Keane Reference Costello and Keane2000). For example, Oxford graduate students who are carpenters may be judged to have an artistic character, while neither Oxford graduate students nor carpenters are judged to have an artistic character (Kunda et al. Reference Kunda, Miller and Claire1990). Several explanations of property emergence have been suggested. They are not mutually exclusive. For the sake of argument, I focus on the explanation that relies on exemplars.

Some properties derive from our knowledge of specific members of the complex category (Hampton Reference Hampton1987; Medin and Shoben Reference Medin and Shoben1988, 183). For example, if a subject has to produce on the fly a complex concept like PRESIDENT FROM TEXAS, she may look in her long-term memory for someone who is president and who comes from Texas. Plausibly, she would retrieve the singular representation of G. W. Bush and she would transmit to the complex concept the properties that are represented by this singular memory. She would thus represent presidents from Texas as being right-wing. Hence, concept combination accesses some singular representations. This suggests that these may be poised to be used in our higher cognitive processes, that is, that they are concepts.

5.5. Consequences

Empirical evidence suggests that concept combination rests on three kinds of knowledge: some statistical knowledge about the typical properties of the relevant categories, some causal and modal knowledge about these categories, and some knowledge about some members of these categories. Prototypes, theories and exemplars are respectively supposed to encode these kinds of knowledge. Since other cognitive processes rely also on these bodies of knowledge, they qualify as concepts. This conclusion is evidence for Premise 3 of the argument from explanatory necessity: in order to explain concept combination, we need to assume that prototypes, exemplars and theories exist.