Penn et al. conflate two issues: one is the degree to which human and animal minds approximate a LoT (language of thought)-based system, and the other concerns the underpinnings of distinctively human learning and reasoning capacities.

A LoT requires a combinatorial syntax and semantics, and syntax-sensitive processes. The LoT hypothesis is noncommittal both about the details of the syntax and semantics of an organism's thoughts and about whether any particular process or rule is implemented as “hardware,” “software,” or something in-between (Fodor Reference Fodor1987). Certainly, it does not address what specific information the organism can learn or reason about. Thus, even the thought capacities of a very simple mind could approximate one LoT-based system to the same extent as a human's could approximate another.

The evidence suggests that the minds of honeybees approximate a LoT-based system to a high degree (Tetzlaff Reference Tetzlaff2006). First, honeybee navigational capacities exhibit systematicity. The effects of panoramic context on the performance of route flight segments and the ability of bees to learn sequences of route segments provide one of many examples (M. Collett et al. Reference Collett, Harland and Collett2002; T. S. Collett et al. Reference Collett, Fry and Wehner1993): If a bee has the capacity to learn the route sequence [distance n to landmark L, then distance m to landmark K], then it thereby also has the capacity to learn any of the route sequences (1) [distance m to L, then distance n to K], (2) [distance n to K, then distance m to L], and (3) [distance m to K, then distance n to L].

The vector-navigational capacities of bees are extremely difficult to explain without positing algebraic rules (Collett & Collett Reference Collett and Collett2000; Gallistel Reference Gallistel, nd and Osherson1998; Menzel et al. Reference Menzel, Greggers, Smith, Berger, Brandt, Brunke, Bundrock, Hülse, Plümpe, Schaupp, Schüttler, Stach, Stindt, Stollhoff and Watzl2005). Moreover, the capacity to freely generalize universally quantified one-to-one mappings requires the use of algebraic rules (Marcus Reference Marcus2001). Honeybees are able to freely generalize the solar ephemeris for their locale: Based on minimal exposure to the sun, their solar-ephemeris learning mechanism produces a record that allows them to estimate the sun's azimuthal position at times when they have not seen it or can never see it (Dyer & Dickinson Reference Dyer and Dickinson1996). Also, recent Y-maze experiments show that bees can solve a delayed-matching-to-sample task, where their solution allows them to generalize to novel stimuli across sensory modalities (Giurfa et al. Reference Giurfa, Zhang, Jenett, Menzel and Srinivasan2001). This suggests that bees can acquire novel rules such as [choose the x-marked arm if x was at the entrance].

Finally, honeybees seem able acquire non-perceptually based information about particulars. Path integration requires a global accumulator that tracks distance and direction from the hive. But there is also a need to posit one or more local accumulators that tie distance and direction information to specific locations, such as the place of release after displacement by an experimenter (Collett et al. Reference Collett, Harland and Collett2002). Significantly, a bee's local-flight-segment learning mechanism can be active without being tied to any specific perceptual features. This could occur, for example, when a displaced bee, after playing out its feeder-to-hive vector (say), arrives at a perceptually uniform location that would have been the location of the hive in the absence of displacement (Menzel et al. Reference Menzel, Greggers, Smith, Berger, Brandt, Brunke, Bundrock, Hülse, Plümpe, Schaupp, Schüttler, Stach, Stindt, Stollhoff and Watzl2005; Riley et al. Reference Riley, Greggers, Smith, Stach, Reynolds, Stollhoff, Brandt, Schaupp and Menzel2003). It appears, then, that a bee can acquire information akin to [distance n and direction d from o], where o represents a particular location (the origination point) independently of any of its perceptual features.

Because humans are by no means unique in possessing a LoT, the real issue is how many distinctively human learning/reasoning mechanisms there are. Penn et al. claim that there is just one, for reasoning about higher-order relations. An objector whom they never properly address would claim that there are many distinctively human learning mechanisms, for “theory of mind” (ToM), for intuitive physics, for natural language, and more (Barrett & Kurzban Reference Barrett and Kurzban2006; Carruthers Reference Carruthers2006). Penn et al.'s brief argument against this “multiple-module” hypothesis is that it can't explain the commonalities they claim to have identified across the various domains of distinctively human thinking. (Other than this, all of their arguments presuppose a single-module opponent, e.g., the claim that language is what explains all the cognitive differences between humans and other animals or that ToM does so.) But this presumes, of course, that there are such commonalities. We deny that there are.

Penn et al. claim that the distinctive thing common to all the phenomena they examine is that humans are reasoning about relations among relations in ways that are not available to nonhuman animals. But when one looks closer at the details of what humans are doing, it is impossible to make this description fit all (or even the majority) of the data. The ToM domain, for example, has nothing to do with relations among relations. It is a matter of identifying and drawing inferences involving the causal variables underlying observed behavior (beliefs, desires, and the rest). Likewise in the case of intuitive physics: this is about identifying and drawing inferences about the unperceived causal forces that govern the physical world (notably gravity and momentum). In contrast, the capacity to engage in transitive inference is not about underlying causal structure at all; nor is it about relations among relations. It is a matter of identifying those relations in respect of which transitive inference is valid, and reasoning accordingly. And the ability to learn category-based rules such as “noun-noun-verb” or “noun-verb-noun” is surely something else again. The only cases that clearly fit Penn et al.'s “relations-between-relations” interpretation are analogical reasoning and higher-order spatial reasoning (although the latter probably depends more crucially on ToM abilities).

On closer consideration, there is no reason whatever to think there could be a single learning mechanism that can engage in all of the distinctively human forms of learning that Penn et al. discuss. The only possibility that they present for consideration is LISA (Learning and Inference with Schemas and Analogies). But this takes semantically interpreted sentences as input and does analogical discovery and reasoning. No arguments are given for thinking that LISA (or anything like it) could also discover the mental variables underlying observed behavior, or the causal forces at work in the perceived physical world. Indeed, it seems most implausible that LISA could be extended into these domains without becoming an entirely different sort of learning mechanism.

At the end of the day, then, the multiple learning mechanisms hypothesis remains, not only unscathed, but as the best extant explanation of the comparative data.