Nonhuman and human brains are made of the same stuff. Yet, as Penn et al. point out, there appear to be deep discontinuities between them. Owing to nonlinear interactions between genotypes, environment, and the resulting phenotypes, functional discontinuities are a common product of continuous evolutionary tinkering. At issue is whether apparent discontinuities in human mental function result directly from biological adaptations such as the “supermodule” hypothesized by the relational reinterpretation (RR) hypothesis or whether human mental abilities – differing quantitatively from those of nonhuman animals – are made qualitatively different by external symbol systems made possible by language, culture, and education.

At the same time as Penn et al. present the evidence against symbolic thinking in nonhuman animals, they tacitly assume that human cognition is innately symbolic and propositional. For instance, they claim that the “propensity to evaluate [similarity] … based on causal-logical and structural characteristics … rather than on […] shared perceptual features appears quite early and spontaneously in all normal humans” (sect. 2.1, para. 1). Yet, one of the authors has himself argued that generating spontaneous analogies poses substantial challenges even for adults who can be easily misled by perceptual similarities (Gick & Holyoak Reference Gick and Holyoak1980). Although subsequent work went on to provide numerous demonstrations of children's and adults' sensitivity to structural relations (e.g., Holyoak et al. Reference Holyoak, Junn and Billman1984), it is not obvious that this type of reasoning arises spontaneously. For instance, Gentner and colleagues (Kotovsky & Gentner 1996; Rattermann & Gentner Reference Rattermann, Gentner, Holyoak, Gentner and Kokinov1998b) have argued for the role of relational labels aligning object representations: Hearing three differently sized objects referred to as “daddy, mommy, and baby” seemed to highlight the size relationship among the objects, enabling the 4- to 5-year-old children to transfer the relation to new objects. Without the relational labels, the relationship among the objects remained opaque. Namy and Gentner (Reference Namy and Gentner2002) further argued that hearing common labels for objects facilitates taxonomic choices, leading children to group objects in more abstract ways (although Penn et al. mention some of this evidence, their dismissal of it is perhaps premature). Importantly, there is evidence that language is not only used as a “training tool,” but may continue to play an online role in relational thinking, as suggested by studies using patients with linguistic impairments. For instance, the patient LEW (Druks & Shallice Reference Druks and Shallice1996), whose primary impairment is severe anomia showed a similar pattern of performance to that of 4- and 5-year-old children on tasks requiring relational reasoning. The addition of meaningful labels for stimuli induced a similar increment in performance to that found in children (Davidoff & Roberson Reference Davidoff and Roberson2004). It seems that when external aids (here, words) are unavailable, performance becomes more concrete and, to a greater degree, driven by perceptual similarities.

Humans can certainly reason analogically and perform relational judgments. The critical question is where these abilities come from. Are they the natural outcomes of the human genome? Or are they made possible by external aids such as the use of relational language? Penn et al. correctly point out that “normal human cognition clearly depends on normal linguistic capabilities” (sect. 9.3.1, para. 1). It remains possible that human performance that is qualitatively different from nonhuman performance may depend on an immersion in human culture and language. The evidence discussed in the target article fails to rule out this possibility.

The authors' assumption of innate symbolic reasoning by humans is also apparent in the section on language in which Penn et al. claim that it is “widely recognized […] that the ability to freely generalize relational operations over role-based variables is a necessary condition for using human languages” (sect. 3, para. 1). Although many researchers do hold this view, there is substantial evidence to the contrary. Some of this evidence is reviewed in the very article used by the authors to support their contention: Gomez and Gerken (Reference Gomez and Gerken2000). For instance, both infant and adult learners become more sensitive to the invariant structure (long-distance dependencies) in an artificial-grammar learning task when the variability of the intervening elements is increased (Gomez Reference Gomez2002) – the kind of token-based performance Penn et al. argue is characteristic of the performance of nonhuman animals (e.g., sect. 2.2., para. 2). Penn et al.'s discussion of language acquisition also omits the work on construction grammars (Goldberg Reference Goldberg2006) and item-based learning (Tomasello Reference Tomasello2003), which have offered ample demonstrations that children's language learning is intensely shaped by specific examples both at the lexical and syntactic levels. Computational modeling provides sufficiency proofs that item-based learning can produce the appearance of abstract role-based categories (Elman Reference Elman2004). Evidence also indicates that adult language comprehension, rather than demonstrating knowledge of abstract role fillers (e.g., noun and verb phrases) instead demonstrates fluid interactions between word-specific knowledge and syntactic frames (Hare et al. Reference Hare, McRae and Elman2003; Reference Hare, McRae and Elman2004). Thus, although it remains to be explained how humans formulate explicit theories about language, actual human language use may rely less on abstract rules than the authors admit.

The human ability to reason about unobservable causes, to draw inferences based on hierarchical and logical relations, and to formulate highly abstract rules is not in dispute. Much of this thinking is compatible on an intuitive level with Penn et al.'s RR hypothesis. But although it is indeed “highly unlikely that the human ability to reason about higher-order relations evolved de novo and independently with each distinctively human cognitive capability” (sect. 11, para. 7), it is not unlikely that such uniquely human abilities depend on the use of external symbol systems. The ability to invent such systems and benefit from them in turn may depend on quantitative improvements in a range of domains: memory, imitation, shared attention, sequence learning, and so on (Elman Reference Elman2005). The hypothetical child magically kept alive by itself on a desert island will inherit these quantitatively superior abilities in a range of cognitive domains. But would this child have all the abilities Penn et al. list as being uniquely human in the absence of the scaffolding afforded by external symbol systems offered by language and culture more broadly? The evidence that Penn et al. offer is insufficient to dismiss the conclusion that such a child would “not differ very much” from other great apes (Tomasello & Rakoczy Reference Tomasello and Rakoczy2003). Although the authors provide a compelling demonstration for an insensitivity to structural relations and the use of symbols by nonhuman animals, in taking for granted the biological basis for these abilities in human animals, the very premise of a biologically based fundamental discontinuity between human and nonhuman minds remains unfulfilled.