Abstract
The present study examined the accuracy of conscious representations that emerge from implicit statistical learning (ISL), a fundamental cognitive process through which we extract regularities in the environment. While ISL has several characteristics of unconscious processing (e.g., it operates unintentionally, produces subjectively unconscious knowledge), participants in ISL experiments always report some fragmentary conscious knowledge. Thus, the notion that ISL is truly an unconscious process has been the subject of perpetual debates. In the present study, we challenge the assumption that these conscious reports reflect direct access to the acquired knowledge. Combining previously collected and novel data, we tested the hypothesis that participants' conscious reports in ISL reflect a post hoc conscious model of their nonconscious knowledge. Across two experiments, participants were exposed to sequences of stimuli (letters, faces, or body movements in VR) generated by different regularities (artificial grammars and grammar-like bigram regularities). In a subsequent test, they decided whether novel strings were grammatical or not and reported their subjective awareness trial-by-trial. In both experiments, we found extreme Bayesian evidence that repetitions embedded in the testing strings made participants more aware of the knowledge driving their grammaticality decisions, above and beyond their influence on responses or accuracy. This suggests that, lacking access to the true basis of their decisions, participants attributed their responses to the most salient feature available: the repetitions. Thus, we find evidence that our conscious experience can misrepresent not only the external world but also our own unconsciously learned representations.