After experiencing an electric shock following a squirt of water to its siphon, Aplysia will show a change in its behavior. The water squirt will come to evoke a dramatic gill withdrawal (a response readily evoked by the shock, but, previously, only weakly by the squirt itself). This is a classic example of associative learning as a phenomenon. What is the mechanism responsible for this phenomenon?

Throughout most of the target article, Mitchell et al. are unambiguous in their answer to this sort of question. The associative learning effect depends, they say, on an effortful, attention-demanding, reasoning process that produces conscious, propositional knowledge about the relationship between events. This is the only mechanism normally allowed. Why do they waver (as they do) when it comes to the case of Aplysia? (Although they focus on human learning, the authors assert that the human system is not likely to be unique and that a process of belief acquisition also underlies animal learning.) The obvious answer – that it seems implausible to attribute such processes to a mollusc equipped with just a few hundred neurons – means that they must find some other explanation for learning in Aplysia. What they offer (a reflexive, stimulus-response [S-R] mechanism) is entirely in accord with what we know from neurophysiological research on this animal (Carew et al. Reference Carew, Hawkins and Kandel1983) and will be widely accepted.

But Mitchell et al.'s acceptance of this analysis has major implications for their central thesis, implications that they scarcely acknowledge. If classical conditioning procedures can produce S-R learning in Aplysia, might they not do so elsewhere? And the evidence currently available (see Hall Reference Hall and Gallistel2002, for a recent review) supports the conclusion that S-R association formation plays a role in generating the conditioned responses shown by higher vertebrates even in the more complex training procedures used for these animals. In allowing the existence of this mechanism, Mitchell et al. have let in, by the back door, a version of the dual-system approach that they profess to reject entirely.

It should be acknowledged, however, that modern studies of classical conditioning in animals, conducted within the associative tradition, have been concerned to show that the effect goes beyond simple S-R learning. As Mitchell et al. themselves point out, it is difficult to explain, by way of the S-R mechanism, the observation that procedures designed to change the value of an outcome (such as sating the animal for a given food) will reduce the vigor of a conditioned response evoked by a stimulus that has previously been paired with that food. This observation has been taken to indicate that the animal has learned something about the relationship between the stimulus and its outcome. Perhaps we need to turn our attention to a more modest version of Mitchell et al.'s thesis, considering its application to just this form of conditioning. Perhaps this form of conditioning, at least, is solely to be explained in terms of propositional reasoning.

Standard associative learning theory offers an alternative interpretation. It suggests that the conditioning procedure establishes an excitatory link between the central representations of the signal and its consequence (i.e., it envisages an S-S [stimulus-stimulus], as opposed to an S-R association). As the signal evokes a response by way of its excitatory effect on the representation of the outcome, the effects of outcome devaluation are readily explained. How are we to choose between this account and one that allows the animal to reason that the signal results in the occurrence of the outcome? Again, we may turn to the behavior shown by (relatively) simple systems to provide an answer. Perhaps the most thoroughly worked-out S-S theory of conditioning is that developed by Wagner (e.g., Reference Wagner, Spear and Miller1981), and he has applied it in detail to the case of eyeblink conditioning in the rabbit. The neurophysiology of this phenomenon has been investigated extensively, and Wagner's theoretical mechanisms map on very well to the systems identified in the cerebellum and brain stem nuclei (Wagner & Donegan Reference Wagner, Donegan, Hawkins and Bower1989). It seems that S-S learning need not involve higher brain structures (decorticate rabbits maintain conditioned responding, Mauk & Thompson Reference Mauk and Thompson1987). The argument is again one of plausibility, but it surely seems more reasonable to endow the cerebellum with the ability to form S-S associations than with the ability to reason about the relationship between events.

If we accept the foregoing arguments (and thus the reality of S-R and S-S excitatory links), then we must reject the central proposal of the target article: that propositional reasoning is the sole source of the associative learning phenomenon. What remains is the far less radical proposal that excitatory link mechanisms play little or no part in generating associative learning effects in human subjects. This is a question that has been worked over repeatedly and has not been resolved. Regrettably, it seems to come down to a matter of personal preference – are we more impressed by the (to me, still surprising) finding that the development of opinions about the allergenic properties of foodstuffs often seems to follow associative principles of the sort embodied in the Rescorla-Wagner model (Rescorla & Wagner Reference Rescorla, Wagner, Black and Prokasy1972) or by the fact that this form of learning shows properties (e.g., the role of awareness; sensitivity to verbal instructions) that lie outside the scope of models of this sort?

Whatever their answers to this question, all are likely to accept the argument of the target article – that it would be a good thing if we could come up with a single theoretical analysis capable of accommodating all the data. Mitchell et al. argue for the propositional analysis. But if the arguments presented above are accepted, we must acknowledge the role of excitatory (S-R and S-S) links in some instances of associative learning. And having done so, parsimony seems to dictate that the next step should be to attempt to extend this sort of account to deal with those features that seem to call for a propositional theory. Success in this enterprise would put paid to the dual-system approach, although not in the way envisaged by Mitchell et al.