Language is composed of both familiar and novel combinations. As a result, language production requires balancing automaticity and creativity. The most likely continuation of a given utterance can be produced automatically using preceding context as a cue. In contrast, production of novel sequences requires generalizing forms to new contexts. In this talk, I will argue that generalizing a form to novel contexts is predicted by the strength of top-down cues to that form, and that experiences that increase reliance on top-down cues contribute to generalization. I will then share data from two studies showing that contextual variability decreases reliance on preceding-context cues and increases reliance on top-down cues. This simple mechanism can explain a wide range of patterns from difficulties faced by children with Developmental Language Disorder (DLD) to patterns of disfluencies in adults’ everyday language use.