I present evidence from Navajo and English that weaker, gradient versions of morpheme-internal phonotactic constraints, such as the ban on geminate consonants in English, hold even across prosodic word boundaries. I argue that these lexical biases are the result of a MAXIMUM ENTROPY phonotactic learning algorithm that maximizes the probability of the learning data, but that also contains a smoothing term that penalizes complex grammars. When this learner attempts to construct a grammar in which some constraints are blind to morphological structure, it underpredicts the frequency of compounds that violate a morpheme-internal phonotactic. I further show how, over time, this learning bias could plausibly lead to the lexical biases seen in Navajo and English.