In an experimental task with novel words, we find that some lexical statistical regularities of Turkish phonotactics are productively extended in nonce words, while others are not. In particular, while laryngeal alternation rates in the lexicon can be predicted by the place of articulation of the stemfinal stop, by word-length, and by the preceding vowel quality, this laryngeal alternation is only productively conditioned by place of articulation and word-length. Speakers’ responses in a nonce word task demonstrate that although they are attuned to the place of articulation and size effects, they ignore preceding vowels, even though the lexicon contains this information in abundance. We interpret this finding as evidence that speakers distinguish between phonologically-motivated generalizations and accidental generalizations. We propose that Universal Grammar, a set of analytic biases, acts as a filter on the generalizations that humans can make: UG contains information about possible and impossible interactions between phonological elements. Omnivorous statistical models that do not have information about possible interactions incorrectly reproduce accidental generalizations, thus failing to model speakers’ behavior.