There is much in this target article that we can fully endorse, but unfortunately, not the general conclusion that Pavlovian conditioning depends on high-level cognitive processes type “conscious propositional reasoning,” the core outcome of which is propositional knowledge about stimulus relations. We believe there is something fundamentally wrong with this proposal in general, but we start with the more modest claim that, both empirically and conceptually, Mitchell et al.'s approach in the target article does not capture the essence of evaluative conditioning (EC) (Baeyens et al. Reference Baeyens, Vansteenwegen, Hermans, Eelen and Columbus2001b).

There are many demonstrations of EC effects going counter to participants' conscious propositional beliefs, including situations (1) in which unconditioned stimulus (US) occurrence is attributed to an irrelevant, non-correlated characteristic of the conditioned stimuli (CSs), but acquired (dis)liking follows the objective CS-US contingency (Baeyens et al. Reference Baeyens, Eelen, Van den Bergh and Crombez1990b; Reference Baeyens, Vansteenwegen, De Houwer and Crombez1996b; Reference Baeyens, Eelen, Crombez and De Houwer2001a); (2) in which US occurrence is correctly believed to be conditional (modulated), but acquired (dis)liking is unconditional (Baeyens et al. Reference Baeyens, Crombez, De Houwer and Eelen1996a; Reference Baeyens, Hendrickx, Crombez and Hermans1998); or (3) in which US occurrence is no longer expected, but conditioned (dis)liking still persists (Vansteenwegen et al. Reference Vansteenwegen, Francken, Vervliet, De Clercq and Eelen2006). These dissociations suggest that EC falls beyond the scope of Mitchell et al.'s model.

Moreover, Mitchell et al.'s account of EC faces more serious, conceptual problems: (i) How does one infer CS (dis)liking from the propositional-declarative knowledge type “CS refers to the (dis)liked US”? From a logical/rational/inferential point of view, there is no causal connection at all between entertaining the propositional belief, “CS refers to (dis)liked US,” and (dis)liking a CS. “Disliking” a flavor-CS does not follow any more from entertaining the belief that “This flavor refers to the bad tasting US Tween20,” than that it would follow from it that the CS should be “liked,” or should evoke feelings of pride, envy, or plain misery. Related to this: (ii) How does one explain that a (correct) propositional belief about the CS-US relationship is not a sufficient condition for EC effects (Baeyens et al. Reference Baeyens, Eelen, Van den Bergh and Crombez1990b)? Finally, (iii) if EC would be based upon a proposition with truth-value, how does one deal with the observation that EC is resistant to extinction; that is, that acquired (dis)liking is not affected by the subsequent experience (and resulting propositional belief) that the CS is no longer accompanied by the US? The alternative is to accept that EC does not depend on propositional attitudes (thinking/believing that “CS refers to US”), but reflects a less-than-propositional state (thinking of US) resulting from the operation of an associative mechanism: CS presentation activates a representation of a (dis)liked US, and the activation of this representation causes or instantiates an approach/avoidance response-tendency that phenomenologically equals a feeling of “(dis)liking,” such that “(dis)liking” the CS causally follows from mere presentation of the CS. Such an association is not true or false, but is simply established or not established; and once it has been formed, it remains there forever.

One could argue that EC is just a special case, and that Mitchell et al.'s approach is valid for all instances of Pavlovian learning except for EC. We are not sympathetic to this possibility either. The problem of linking conditioned responding (“performance”) to specific “belief” states (resulting from entertaining a particular propositional mental content) is not restricted to EC, but spreads to the whole domain of Pavlovian conditioning. Mitchell et al. indeed admit that a detailed causal/mechanistic account of the translation of propositional beliefs (and concomitant/resultant expectancies) to specific physiological/behavioral responses is not the forte of their model. It is a long way from the propositional attitude, “I believe that the tone predicts a shock,” to increases in heart rate, muscle tension, breath regulation, hormone release, or the activation of escape/avoidance responses. It is not just a long way, but also a way that is not specified at all by the (content of the) propositional attitude. In principle, Pavlovian conditioned responses are not rationally connected to propositional knowledge about the CS-US relationship and cannot (logically) be inferred from it; nor do they (necessarily) result from the interaction with other propositional beliefs (i.e., from reasoning) (Shanks & Dickinson Reference Shanks and Dickinson1990). A model that can explain all but conditioned behavior, is lacking something quintessential. Again, the alternative is to accept that conditioned responding causally results from the associative activation of the US-representation – a theory that may indeed require (much) refinement, but at least offers an account that works in principle.

Why is it that Mitchell et al. have arrived at this problematic account of learning? First, in some arguments purportedly favoring the “conscious reasoning/propositional knowledge” idea, two issues are erroneously mixed up. The data on the role of awareness, and the influence of cognitive load/secondary tasks, favor an account of Pavlovian learning that acknowledges the importance of “controlled” processing indeed; but these arguments do not bear at all upon the issue whether the acquired knowledge should take the form of structured mental representations, or of non-propositional associative links between representations. Second, many of the observations that do favor an account in terms of representations with combinatorial syntax and semantics (and structure sensitivity of processing), are derived from experiments that invoke processes/faculties that simply go beyond the scope of what associative learning theory reasonably could be expected to explain. The findings on the influence of verbal instructions, abstract rules, and deductive reasoning show that people indeed can use language and reason, and can transform complex, structured mental representations in situations that require more than simple registration of stimulus co-occurrence. But how does this favor a “conscious reasoning/propositional knowledge” account of associative learning any more than does a demonstration that humans can play chess or understand poems?

Mitchell et al.'s account of learning tries to get rid of a dual-system approach, and while doing so, throws away the very notion of associations. According to our analysis, there is not much ground to justify this radical stance. Moreover, Mitchell et al. still adhere to a dual-system approach of mental processes in general. Even though the very act of learning is supposed to take place in a conscious reasoning system, this system is said to get inputs from a non-conscious, automatic perceptual/memory system; and the only thing Mitchell et al. ultimately propose is to resect plasticity from one part of the dual system. Parsimony, where art thou?