In their target article Mitchell et al. contend that associative learning is best explained as the result of effortful cognitive processes based on propositions, and that there is no need to posit a contribution based on the formation of associative links. The target article greatly overstates the case for rejecting the contribution of associative links, however, because of its remarkably selective citation and interpretation of prior work.

Although there are situations where propositional accounts provide good explanations of human learning, there are also many examples where the pattern of results is more easily explained by associative-link accounts. For example, if beliefs are based on post hoc reasoning about events stored in memory, changing the temporal order in which these events are stored should not affect judgments; and yet several studies have demonstrated trial-order effects in learning (Collins & Shanks Reference Collins and Shanks2002; Dickinson & Burke Reference Dickinson and Burke1996; López et al. Reference López, Shanks, Almaraz and Fernández1998b). Similarly, the evidence relating to differences in diagnostic and predictive learning is overstated: many studies have found no impact of this manipulation (e.g., Arcediano et al. Reference Arcediano, Matute, Escobar and Miller2005; Cobos et al. Reference Cobos, Lopez, Cano, Almaraz and Shanks2002; López et al. Reference López, Cobos and Caño2005; Tangen et al. Reference Tangen, Allan, Sadeghi and Wills2005). Nor is mention made of cancer patients who form food aversions despite being fully aware that chemotherapy treatment, rather than any food consumed, produced their illness (e.g., Bernstein & Webster Reference Bernstein and Webster1980). But the most surprising omission is of research on blocking by two of the target article's authors themselves: Mitchell et al. (Reference Mitchell, Lovibond, Minard and Lavis2006) argued that for participants to reason that a blocked cue T is non-causal requires knowledge of the outcome with which T was paired. Their results, however, revealed blocking of T without any memory of the T–outcome relationship, which led them to conclude that “associative cue-competition would appear to be a strong candidate mechanism for noninferential forward blocking in humans” (p. 842).

In the target article, Mitchell et al. also neglect dissociations in patterns of learning under different circumstances that are challenging for single-process propositional accounts, but which follow naturally from dual-process approaches to learning. For example, López et al. (Reference López, Cobos and Caño2005) demonstrated that participants were generally insensitive to the difference between predictive and diagnostic tasks but did show differences when instructions made clear the importance of this distinction in causal order. Likewise, although Mitchell et al. cite Shanks and Darby (Reference Shanks and Darby1998) as providing evidence for propositional learning, they ignore the fact that results consistent with this account were only observed in participants who had learnt the task well; those who learnt less well showed responses consistent with associative-link theories (similar responses to novel compounds and the elements that comprise those compounds). Le Pelley et al. (Reference Le Pelley, Oakeshott and McLaren2005a) demonstrated that participants show an unblocking effect when information is presented on a trial-by-trial basis, but not when presented in a questionnaire format that would facilitate the use of propositional learning. Moreover, Mitchell et al.'s analysis of this unblocking effect does not stand up to scrutiny. Participants could indeed remember what had changed between A–O1O2 and AT–O1O3 trials: Their accuracy in predicting O1 on AT–O1O3 training trials was significantly higher than that for their predictions of O3.

It is unfortunate that Mitchell et al. address only a straw-man version of associative-link accounts. For example, they assert that associative theories see learning as proceeding without awareness. Few associative theorists would agree with this characterisation, however: Why should people necessarily remain unaware of links that are formed? Although the issue of awareness is orthogonal to associative accounts (Shanks Reference Shanks2007), the target article explicitly states that any example of learning without awareness would be highly damaging to the idea that all learning is propositional. Yet, Mitchell et al. cite two examples of dissociations between propositional knowledge and conditioned reactions in humans (flavour conditioning and the “Perruchet effect”), and the essential features of both studies have been replicated (see Lovibond & Shanks Reference Lovibond and Shanks2002). In addition, the fact that associative learning can occur in anaesthetised animals (see Lovibond & Shanks Reference Lovibond and Shanks2002) indicates that the idea that propositional mechanisms can explain all animal learning (other than that produced by S-R mechanisms) must be, beyond reasonable doubt, false (see also Iselin-Chaves et al. [Reference Iselin-Chaves, Willems, Jermann, Forster, Adam and Van der Linden2005] for learning under anaesthesia in humans). It warrants repeating that any dissociation between proposition knowledge and learning is fatal to the current account, and such dissociations do exist (albeit that unambiguous evidence is not widespread).

The insistence that associative accounts rely on nodes that represent whole stimuli in a symbolic manner is also a mischaracterisation. Foreshadowed by Estes's (Reference Estes1950) stimulus sampling theory, associative models explicitly acknowledge that any stimulus comprises multiple features that might each be shared with other stimuli (e.g., Blough Reference Blough1975; Brandon et al. Reference Brandon, Vogel and Wagner2000). This undermines attempts in the target article to characterise the concept of generalisation as an unjustified assumption by which “freedom is gained to explain results” (sect. 6.1, para. 3). In fact, this is an integral and fully specified feature of almost all current associative learning models, and flows directly from the idea that whole stimuli should be considered as collections of potentially overlapping features. Within-compound associations are also treated as “get-out clauses” despite following naturally from, and being explicitly predicted by, standard associative principles. Furthermore, there is evidence for their existence (Rescorla & Durlach Reference Rescorla, Durlach, Spear and Miller1981) and influence upon cue-competition (e.g., Batsell et al. Reference Batsell, Paschall, Gleason and Batson2001; Durlach & Rescorla Reference Durlach and Rescorla1980). Although there are examples of particular associative-link models being modified in light of an inability to account for particular results, this does not undermine the fact that principles of generalisation and within-compound associations are instantiated within associative-link models as a class.

Finally, Mitchell et al. criticise associative theory for lacking parsimony because it must predicate two sources for human learning (associative-link and propositional mechanisms). However, associative-link theories are very parsimonious in other ways. Most notably, they can explain aspects of human learning (e.g., sensitization, habituation, perceptual learning) which lie beyond propositional mechanisms. Although associative models inherently require dual-process accounts of human learning, propositional accounts are inherently multiple-process with respect to other phenomena. Hence, proposition-only accounts of human learning are no more parsimonious than dual-process accounts when considered in a broader context.