Mitchell et al. state that a demonstration of conditioning without awareness would be a major problem for their propositional approach. They claim the literature is consistent with this approach because convincing evidence for learning without awareness is scarce, and because there is a “clear concordance between conditioning and awareness” (sect. 4.1, para. 6). Lovibond and Shanks (Reference Lovibond and Shanks2002) established clear principles for the appropriate assessment of awareness, and found most demonstrations of unconscious learning used awareness tests that were insufficiently sensitive or rigorous. Over time, their evaluation of the literature has fuelled the conclusion that learning cannot occur without awareness. It is worth examining that conclusion again. Aside from the observation that the dubious nature of the evidence for unconscious learning is not evidence of absence, one can make a strong case that the concordance between learning and awareness does not inform the current debate at all. This is because a realistic dual-process account should still predict that learning will be harder to observe and less likely to occur in unaware participants.
First, it is wrong to assume that link formation should be observable in all situations and in all participants. Dual-process theories usually assume that conscious reasoning can have a large impact on behaviour, large enough in some circumstances to obscure other more subtle behavioural influences. This is even true of participants who reason incorrectly or erratically during an experiment. We should not be surprised if participants who were aware of an associative relationship showed behavioural evidence of learning. But those who were unaware may fail to show learning because they are deliberating on a spurious hypothesis or concentrating on something unrelated to the task. This may disrupt performance, regardless of whether learning has taken place.
It is also wrong to assume that, just because a learning process is automatic, it will always occur (and always to the same degree). If link formation is affected by selective attention or memory load, then the evidence showing a close correspondence between awareness and learning is highly predictable by any account. To suppose that link formation is affected by cognitive processes is not contrary to its conception as an automatic process. Changing the input to a link-based network inevitably leads to changes in what is (or what is not) learned by that system. A model of automatic link formation would be impervious to manipulations of selective attention and working memory if, and only if, its input was restricted to the earliest levels of sensory processing.
Mitchell et al. argue that if link formation is dependent on cognitive resources, then “one of the reasons for postulating the existence of an automatic link-formation mechanism has been removed” (sect. 4.2, para. 3). But this argument only holds if capacity limitations affect link formation and propositional learning in the same way. Instead, they might affect what the cognitive system does in the sense of the inferences that are drawn about the relevant events, whereas those same limitations might affect what the automatic system receives in the way of input. With this in mind, it would not be at all surprising if consistently attentive participants were more likely to consciously identify the relevant contingencies and also more likely to learn, whereas participants who attended erratically were less likely to be aware and less likely to learn.
Although Mitchell et al. provide examples where learning is sensitive to rules and instructions, these examples only confirm what both approaches already assume – that conscious inferences and beliefs can influence decisions and behaviour. More importantly, some learned behaviours do not show this sensitivity. While Mitchell et al. discuss Perruchet's (Reference Perruchet1985) dissociation, other prominent examples include conditioned taste aversions and anticipatory nausea and vomiting elicited by cues associated with chemotherapy (Bernstein & Webster Reference Bernstein and Webster1980; Carey & Burish Reference Carey and Burish1988). Both appear to be very clear cases of uncontrollable, automatically learned responses that have nothing to do with the beliefs of the sufferer.
Mitchell et al. wish to extend their approach to describe learning in other animals. But the utility of the approach is questionable given the obvious obstacles in establishing whether other animals even have awareness or beliefs, let alone in measuring either. And if we assume they do, what can we deduce from an animal's behaviour about the content of its beliefs? As one example, when a pigeon pecks or drinks a key-light during an autoshaping procedure (e.g., Jenkins & Moore Reference Jenkins and Moore1973), does it believe that the key-light is food or water? What reasoning process could give rise to such a proposition? Or should we conclude that the pigeon's behaviour does not reflect the content of its belief? If the latter, then we must concede that the learning process cannot be adequately investigated by examining animal behaviour.
At a superficial level, positing a single mechanism for learning seems more parsimonious than assuming two mechanisms, and Mitchell et al. encourage us to abandon link-based learning on this basis. However, conscious propositional reasoning is computationally more expensive and much less well-specified than the operations underlying the strengthening of associative links. The propositional learner is one who retrieves information from episodes and consciously rationalises which events signal other events within the same episode. This comes with its own set of assumptions that imply considerable computation. For a start, in order to calculate relationships symbolically, the propositional system needs to decide what each event is. It also needs to store events in an episodic fashion that includes temporal information. This contrasts with the simple mechanisms by which associative links are strengthened or weakened according to statistical regularities in the environment.
The comparison begs the question: Is it really parsimonious to conclude that all instances of learning are the consequence of an elaborate and cumbersome set of cognitive operations just because we know these operations affect human behaviour in certain circumstances? We conclude that it is more parsimonious to assume learning is the product of a very simple link mechanism, but to evaluate exceptions to this rule where the evidence necessitates. Of course, the issue of parsimony is at the very heart of the single-versus-multiple process debate.
