This article investigates the learnability filter (LF) hypothesis, according to which the set of logically possible grammars predicted by linguistic theory is reduced to a proper subset of learnable grammars by external principles of language learning. Antifaithfulness constraints (Alderete, 2001a) provide a linguistic theory that predicts the existence of circular chain shifts of two segment types, including purely phonological exchanges that are unattested cross-linguistically (Anderson and Browne, 1973). Overt data representing such systems are fed into a standard OT learning model in which learners have IOantifaithfulness constraints at their disposal. Despite this evidence and these constraints, learners always select grammars in which segmental exchanges are restricted to morphologically defined environments, consistent with typological findings. These results are shown to have implications for the nature of constraints in Optimality Theory, the correct analysis of morpho-phonological alternations, and a host of representational assumptions in phonology and morphology. #