Despite many years of investigation into implicit learning in nonlinguistic domains, the potential for implicit learning to deliver the kinds of generalizations that underlie natural language competence remains unclear. In a series of experiments, we investigated implicit learning of the semantic preferences of novel verbs, specifically, whether they collocate with abstract or concrete nouns. After reading sentences containing the verbs, participants were required to judge the familiarity of pairs of novel verbs and nouns and to indicate their confidence or the basis of their judgment (i.e., guess, intuition, memory). Although all of the words had occurred in the texts, none of the critical items had actually occurred together. However, endorsement rates were significantly higher for pairs that respected the semantic preference rules than those that did not. Through analysis of subjective measures and verbal report, we argue that, for the majority of participants, this effect was based on unconscious knowledge. We argue that implicit learning of the kind of generalizations underlying semantic preferences is possible even after limited exposure.