This paper presents the results of an artificial grammar learning experiment that supports a theory of analytically biased learning in which learners form hypotheses about novel phonological processes based on their knowledge of grammatical principles. Specifically, learners naturally inferred directionality as the source of spreading over the typologically implausible 'majority-rules' strategy. Adult native speakers of English were trained on a back/round vowel harmony language that was ambiguous between an unattested 'majority rules' spreading pattern and a natural right-to-left spreading pattern (or, in a separate condition, a left-to-right spreading pattern). Participants preferred the 'majority rules' alternation only when it was consistent with the direction of spreading that they were trained on, supporting the existence of analytic learning biases that lead learners to infer unmarked, natural phonological processes.