Rogers & McClelland's (R&M's) deeply impressive book, Semantic Cognition (Reference Rogers and McClelland2004), outlines a plausible alternative to theory-theory, and it shows how parallel distributed processing (PDP) models can capture many phenomena in the literatures on categorisation, naming, and concepts. We are interested in implications of R&M's approach for what we know about inductive reasoning and its relation to knowledge. Our understanding of their position is that in many cases inductive reasoning and semantic cognition are just different names for the same thing. However, in our view, although knowledge is very important to an understanding of thinking, there are limits to what can be explained by recourse to knowledge and the processes by which it is attained. First, we will describe some effects, our own and other people's, which appear to challenge accounts that equate thinking with semantic cognition. Then we will speculate as to what kinds of account might best capture those effects.

The literature on deductive reasoning contains the clearest evidence that there is more to thinking than semantic cognition. For example, Handley et al. (Reference Handley, Capon, Beveridge, Dennis and Evans2004) used a belief bias task where 10-year-old participants were asked to reason about arguments the validity and believability of whose conclusions had been orthogonally manipulated. Participants also completed measures of inhibitory control and working memory. Successful performance on this task calls for the inhibition of outputs from semantic cognition, and Handley et al. observed that inhibitory control and working memory were independent predictors of the ability to respond in accord with logical validity.

Of course, R&M make no claims about deduction. However, some of our own work asks whether inductive reasoning can be wholly captured by fast and parallel knowledge-based processes or whether slow, resource-demanding processes also play a role. For example, Feeney (Reference Feeney2007) studied inductive projection using arguments with multiple premises. Such arguments can be used to study whether people are sensitive to diversity and amount of evidence when evaluating inductive arguments, and sensitivity to these phenomena has been modelled in wholly similarity-based ways (Osherson et al. Reference Osherson, Smith, Wilkie, López and Shafir1990; Sloman Reference Sloman1993). Feeney showed that a measure of IQ is associated with people's sensitivity to these principles. The results are complex, but particularly in the case of diversity, those participants who scored highest on the IQ test tended to be most sensitive to the diversity of the premises. One interpretation of correlations between IQ and performance on particular thinking tasks is that they indicate the involvement of slow, symbol-manipulating processes in thinking (see Stanovich Reference Stanovich1999). That is, inductive reasoning is more than semantic cognition, and is based on more than processes that allow for the calculation of similarity between representations.

A related finding concerns when sensitivity to properties of the premises of an inductive argument develops. Wilburn and Feeney (Reference Wilburn and Feeney2007) have shown that sensitivity to diversity begins to emerge at age 7, whereas sensitivity to amount of evidence does not begin to emerge until age 13. We interpret this finding as suggesting that in a category-based inductive argument, sensitivity to amount of evidence requires the reasoner to know that larger samples make for sounder inferences, whereas sensitivity to diversity can be demonstrated on the basis of similarity calculations alone. This finding also suggests that there is more to thinking than mere similarity.

Like R&M (Semantic Cognition, Ch. 8), we have also been concerned with the effects of knowledge about causal relations on inductive generalisation. A particularly interesting case comes from Medin et al. (Reference Medin, Coley, Storms and Hayes2003), who demonstrated the category-based conjunction fallacy. They compared strength ratings for the following argument:

to the mean strength ratings for the causally near generalisation from lead to pipes and to the causally distant generalisation from lead to plumbers. (We term lead and pipes causally distant because the reasoner has to infer the involvement of pipes to explain the transmission of Property X from lead to plumbers.) Medin et al. (Reference Medin, Coley, Storms and Hayes2003) demonstrated that, on average, people commit the conjunction fallacy. That is, the strength rating for the argument with the conjunctive conclusion is higher than the average strength rating for other two arguments.

Feeney et al., (Reference Feeney, Shafto and Dunning2007) followed up on this finding and showed that the near generalisation from lead to pipes is rated strongest, whereas the distant generalisation from pipes to plumbers is rated weakest. In addition, we found that participants highest in IQ were more likely to rate the near generalisation stronger than the conjunctive argument. In further follow-up experiments (Crisp et al., under review) we asked participants to concurrently perform a working memory task whilst rating generalisation strength. The secondary task increased rates of the conjunction fallacy observed when ratings for the conjunctive argument were compared to the distant case, but not when compared to the near case. Our interpretation of these findings is that in the distant case, people resisted the conjunction fallacy because they explicitly reasoned about causal relations and reconstructed the causal chain linking, for example, lead to plumbers. Having reconstructed the causal chain, they assigned equally high-strength ratings to distant and conjunctive arguments. A concurrent task impeded their ability to engage in this causal reasoning in the distant case, whereas it had no effect in the near case because the stronger causal relation was immediately available. The same basic pattern was obtained when participants were encouraged to answer quickly. Thus, the individual differences, secondary task, and speeded task data suggest that some effects of knowledge on thinking are moderated by processes that are associated with IQ and working memory, and which take time.

Our preferred explanation for these findings is that there are at least two types of thinking (see Evans Reference Evans2006; Sloman Reference Sloman1996; Stanovich Reference Stanovich1999), a fast and associative form of thinking, and a slower and sequential type of thinking. The first type of thinking performs, among other operations, similarity calculations, whereas the second type applies rules and makes some (but not all) inferences about causal relations. It has been studied by researchers interested in models (Johnson-Laird Reference Johnson-Laird2006), rules (Rips Reference Rips1994), or simulations (Evans & Over Reference Evans and Over2004) for reasoning. R&M's models of semantic cognition appear more relevant to the first type of thinking than they do to the second.

Impressed as we are by R&M's book, we cannot see how their current models can capture our data. Of course, R&M have anticipated our concerns and questions (see Semantic Cognition, pp. 371–73), but it may take another book to convince us of their answer.