Mitchell et al. argue that learning relies on propositional reasoning in humans, as well as in certain other animals such as rats. They support their view with similar results for both species in certain learning studies. At the same time, they admit that not all nonhuman animals (e.g., Aplysia) draw on propositional processes. Taking their argument seriously, we will show that their stance creates grave problems in accounting for human learning exclusively by propositions.
The first problem arises from the fact that Mitchell et al.'s view effectively partitions all animals into two groups; those, most notably humans, whose learning relies entirely on propositional reasoning and those whose learning is completely without propositions (e.g., Aplysia).
Such partitioning implicitly establishes a “magical” point in human phylogenesis where animals stopped being equipped with non-propositional learning. But why should evolution favor individuals who throw away mechanisms that have proven adaptive for generations of ancestors? It seems much more likely that, instead of disposing of the mechanism, individuals may have complemented it by acquiring other mechanisms such as propositional reasoning. Hence, it seems more reasonable to assume that humans are equipped with two, rather than only one, learning system. Against this background, any account which claims that human learning is entirely propositional is only satisfactory if it also explains why previously existing learning systems have been shut down during phylogenesis. However, such an explanation is not provided by Mitchell et al.
The second problem arises from the methodology used to establish the point of propositional learning in non-human animals. Based on behavioral similarities of nonhuman animals and humans, the learning mechanisms assumed for humans are also thought to be at work in the nonhuman animals. This method for ascribing learning mechanisms, however, cuts both ways. If nonhuman animals and humans behave similarly in learning, one can as well ascribe the nonhuman learning mechanisms to humans. Either Mitchell et al. are not aware of the two-edged nature of their method or they are not aware of results on nonhuman learning which shed doubt on their main claim.
Some of these results have been obtained using honey-bees as subjects. Komischke et al. (Reference Komischke, Giurfa, Lachnit and Malun2002), for example, trained bees on reversal problems, where a previously reinforced conditioned stimulus (CS) was subsequently not reinforced, and vice versa. The bees showed improved reversal learning performance, depending on the number of previously encountered reversals. Thus, one could argue that the bees acquired some kind of a (propositional) reversal rule.
In another series of experiments, Lachnit et al. (Reference Lachnit, Ludwig and Reinhard2007) and Deisig et al. (Reference Deisig, Sandoz, Giurfa and Lachnit2007) investigated the effects of trial spacing on learning performance in humans and bees, respectively. Importantly, the two studies yielded comparable results. For both, humans and bees, increasing the inter-trial interval led to a general increase of the conditioned response (CR) and an improved ability to differentiate between CS+ and CS- in learning positive patterning (PP) and negative patterning (NP) discrimination problems.
Further similarities between the learning of humans and bees have been demonstrated by Kinder and Lachnit (Reference Kinder and Lachnit2003) and Deisig et al. (Reference Deisig, Lachnit, Sandoz, Lober and Giurfa2003). In these studies, subjects were presented with an extended NP discrimination (A/B/C+, AB/AC/BC+, ABC−). Contrary to major associative learning theories available at that time, both bees and humans showed no CR difference for the single- and double-compound CSs.
These five studies show substantial convergence of learning in humans and bees. Direct comparisons indicate that humans and bees behave similarly regarding several learning problems and experimental manipulations. Consequently, following the method of Mitchell et al., humans and bees can be assumed to draw on similar learning mechanisms. Mitchell et al. probably would argue that propositional reasoning is at work, because human learning relies completely on propositional reasoning. Hence, according to Mitchell et al., learning in bees also relies on propositional reasoning.
Yet, the results of Komischke et al. (Reference Komischke, Sandoz, Lachnit and Giurfa2003) clearly contradict this conclusion. Bees had to solve NP and PP discriminations under various conditions. In one condition, the CSs of the NP problem were presented to both antennae of the bees, whereas in another condition, the CSs were presented to only one of the antennae. A third condition simultaneously presented PP to one antenna and NP to the other antenna (with identical CSs for both problems): (a) NP was learned only in condition 1 and 3; (b) the PP discrimination was hampered in condition 2.
These results contradict the idea that bees' learning relies on propositional reasoning. According to the propositional account, learning should degrade in dual-task situations. This was observed for PP; NP, however, improved under the dual-task condition. Hence, bees' learning did not rely on propositional reasoning. If learning in bees is not propositional and learning in bees and humans is comparable, using the method of Mitchell et al., one can only conclude that humans – at least sometimes – also do not learn propositionally.
Interestingly, this conclusion regarding non-propositional learning in humans on PP and NP tasks is further corroborated by Lachnit and Kimmel (Reference Lachnit and Kimmel1993) and Lachnit et al. (Reference Lachnit, Kinder and Reinhard2002). Lachnit and Kimmel (Reference Lachnit and Kimmel1993), employing a design similar to that of Shanks and Darby (Reference Shanks and Darby1998), observed an asymmetric relation of PP and NP in transfer situations: Learning PP hampers NP, but not vice versa. However, utilizing different response systems (spared conditioned response [SCR] and eye-blink) for the different patterning problems, Lachnit et al. (Reference Lachnit, Kinder and Reinhard2002) found no such asymmetry; PP and NP were solved equally well instead. If learning to solve PP and NP discrimination problems relies on the construction of consciously available general rules, why then should the effect of learning these rules be response-specific?
In light of the evidence presented here, the account of Mitchell et al. in the target article seems untenable. Not only do the authors fail to present important information to justify their account, but empirical evidence on human and nonhuman learning also contradicts their claim that human learning is entirely propositional.
