We offer a theory of polarization as an optimal response to ambiguity. Suppose individual A's beliefs first-order stochastically dominate individual B's. They observe a common signal. They exhibit polarization if A's posterior dominates her prior and B's prior dominates her posterior. Given agreement on conditional signal likelihoods, we show that polarization is impossible under Bayesian updating or after observing extreme signals. However, we also show that polarization can arise after intermediate signals as ambiguity averse individuals implement their optimal prediction strategies. We explore when this polarization will occur and the logic underlying it. (JEL D81, D82, D83)