In their target article, Mitchell et al. present a detailed and very thoughtful evaluation of the potential evidence in favour of automatic association formation as a source of associative learning effects. They convincingly argue that there is presently very limited, if any, evidence for the existence of a separate, non-propositional association formation module. They rightfully point out that, given the obvious need to postulate the existence of a propositional module in order to explain many aspects of human associative learning, an association formation module that does not add explanatory power to the propositional module is entirely redundant.
Still, the claim that any instance of human associative learning (i.e., any change in performance that is a result of the presence of regularity in the relation between events in the world) must by necessity be due to the operation of controlled, propositional reasoning processes, is a strong one. Prima facie, Mitchell et al.'s claim seems ill-fitted to the existence of a phenomenon such as evaluative conditioning: If, for example, I have developed a liking for white wine because of spending many pleasant holidays in France, it would seem that this liking for white wine does not need to reflect any knowledge about the relation between white wine and anything else, beyond the fact that white wine makes me think of France. Essentially, the fact that white wine makes me think of France is non-propositional (I cannot be right or wrong for being reminded of France upon smelling white wine). As such, evaluative conditioning effects very strongly appear to result from the mind being carried from one idea or representation to another, without any intermediate processing, much like what is the presumed mode of operation of an association (Fodor Reference Fodor2003).
It is then perhaps not surprising that some of the best evidence for automatic association formation comes from an evaluative conditioning study (Baeyens et al. Reference Baeyens, Eelen and Van den Bergh1990a). Still, the description of the propositional approach as offered by Mitchell et al. leaves open the possibility that even evaluative conditioning effects, although perhaps resulting from automatic, non-propositional memory retrieval processes (an object “automatically” making you think of something pleasant), do necessitate the conscious, falsifiable establishment, in propositional form, of a link between events (if only of the form “event A co-occurred with event B”) at some earlier point in time. That is, the fact that evaluative conditioning effects, at performance, are almost by nature non-propositional (the fact that A makes you think of B is not something that you can subsequently evaluate as correct or wrong), does not preclude that they perhaps only occur if people at some point have consciously noticed some sort of real-world relationship between A and B (such as “A has repeatedly co-occurred with B,” a statement which you can obviously evaluate to be true or false).
Does this render the propositional account unfalsifiable? Surely, the fact that performance may reflect automatic memory retrieval of propositional knowledge stored earlier and may moreover reflect propositional knowledge indirectly (such as when stored propositional knowledge about the co-occurrence of two events influences your subsequent evaluation of one of both), does make falsification of the general framework difficult, but not impossible. It would suffice to convincingly demonstrate associative learning about entirely subliminally presented CSs to rule out a role for propositional reasoning altogether. The debate about whether such evidence already exists still seems to be open (see Wiens & Öhman Reference Wiens and Öhman2002 vs. Shanks & Lovibond Reference Shanks and Lovibond2002).
However, the most important contribution of the propositional approach to associative learning is not to be situated in proving the association formation approach wrong. As Mitchell et al. point out, what is perhaps most important, is that it has provided a new perspective on conditioning, not only in humans (where at least a contribution of reasoning processes to learning has long been acknowledged), but also in animals. This perspective has not only enabled us to unveil the importance of rule learning in animal Pavlovian fear conditioning (Beckers et al. Reference Beckers, Miller, De Houwer and Urushihara2006), but also to highlight the parallels between extinction learning and rule learning in terms of context sensitivity and generalisation (Wheeler et al. Reference Wheeler, Beckers and Miller2008). As such, the propositional approach has opened up a whole new framework for the understanding and the prediction of human and animal conditioning phenomena, the impact of which is bound to further increase over the coming years.
And perhaps this is where a caveat about the propositional approach to associative learning, in turn, is warranted. Notwithstanding the impressive amount of evidence that the propositional approach is more veridical than the association formation approach, it seems beyond argument that models developed within the association formation tradition have continuing heuristic value as well. As an example, just recently Leung and Westbrook (Reference Leung and Westbrook2008), in a series of extremely elegant experiments, demonstrated that the degree of additional extinction accrued by a cue exhibiting spontaneous recovery is governed by both individual prediction error of the cue and common prediction error of all cues present during an extinction trial. Does such a finding invalidate the propositional nature of associative learning? Not necessarily (probably not, one might even argue). Still, it is obvious that experiments like these would never have been designed, and these findings never revealed, on the basis of our current understanding of propositional reasoning. As such, it may simply be too early for one truth to govern our inquiries into human and animal associative learning. Keeping our antennas open to discover empirical phenomena in the realm of associative learning and conditioning will probably necessitate a willingness to entertain a variety of models and approaches for some time to come.
