Conndence Judgements, Performance, and Practice, in Artiicial Grammar Learning

Artiicial grammar learning is noted for the claim that subjects are unaware of their knowledge. Chan (1992) and Dienes et al. (in press) have demonstrated that subjects are unaware in the sense that they lack meta-knowledge. Dissociations between subjects' performance and their conndence in their decisions suggest that the learning mechanism may be in some sense encapsulated from the \conndence system". Here we tested the alternative hypothesis that the conndence system is initially poorly calibrated, or does not know which aspects of the learning mechanism to attend to, by training and testing subjects over four weekly sessions. On all four weeks we found a strong, near-perfect association between conndence and performance for trained subjects, but a dissociation for untrained control subjects. We discuss possible explanations for these results, and previously observed dissociations. Artiicial grammar learning is notable for the controversial claim that subjects acquire knowledge which allows them to distinguish strings which follow the same rules as a set of previously memorised strings, from those which do not, but that they are not consciously aware of this knowledge (Reber, 1967, 1989). However, measuring subjects' conscious knowledge is plagued with the problems of nding suitably sensitive, explicit tests, and of ensuring that the knowledge these explicit tests measure is the same as the knowledge subjects use to perform (in press) set issues of consciousness aside and focus instead on behaviour: Subjects' ability to make conndence judgements about their performance. This tests meta-knowledge, or \what subjects know about what they know". Chan (1992) showed that subjects' conndence in their judgements was unrelated to the likelihood that those judgements were correct. Similarly, Dienes et al. (in press) showed that even when subjects thought that they were guessing, their performance was above chance (and untrained control) levels. Chan (1992) and Dienes et al. (in press) propose that the dissociation between conndence and accuracy is evidence that subjects lack meta-knowledge. This suggests that the learning mechanism is to a certain extent encapsulated from the \conndence system", so that its inner workings (e.g. the strengths of its rules, or connection weights, or its error signals) are inaccessible. In this paper, we present a study which aimed to test a possible alternative explanation for this dissociation; that meta-knowledge is in principle available, but simply that extraction of this knowledge requires learning or calibration, and is initially inaccurate. Dienes and Perner's (in press) discussion of the extent and …