“The Reports of My Death Are Greatly Exaggerated”: Humans Can Extract Unconscious Bigram Knowledge in the Artificial Grammar Learning Task

Humans often appear to possess knowledge they cannot fully articulate. For instance, in the landmark artificial grammar learning (AGL) task, participants acquire a complex regularity (grammar) that generates letter sequences, while typically claiming to have little or no subjective access to the regularity, thus suggesting that learning has been largely unconscious. A highly influential argument against this conclusion comes from studies showing that participants can identify simple letter pairs (bigrams) that the grammar is composed of, when asked to do so. Consequently, the unconscious status of learning in AGL has remained highly controversial. Crucially, however, these latter studies have never measured participants’ subjective states when performing the bigram identifications but assumed that bigram identification necessarily reveals conscious knowledge. In this paper, we put to test this critical assumption, by employing novel bigram-based artificial grammars and, simultaneously, by measuring participants’ subjective states when operating with bigram knowledge. For generalization, we employed both typical letter stimuli and face stimuli. Participants (N = 188) memorized strings (of faces or letters), constructed from the grammars; subsequently, they classified novel strings (half grammatical, half nongrammatical), reporting subjective awareness trial-by-trial. We found that letter strings were classified above chance based both on unconscious and on conscious knowledge (all ps < .001, Bayes factors > 106), while data were insensitive for the faces condition (Bayes factors between 0.33 and 3, all ps > .05). Our findings demonstrate unconscious learning of bigrams for letter strings, challenging the influential idea that accurate bigram knowledge necessarily reveals conscious knowledge.