I model a spatial election where citizens can flexibly acquire costly information about their policy preferences before voting. In equilibrium, learning about policy preferences creates voter polarization. In a one-dimensional policy setup, the distribution of political preferences induced by learning is bimodal even if the true underlying distribution is unimodal. With a multi-dimensional policy space, voter preferences induced by learning are aligned across different issues even if issues are independent according to the true preferences. A lower cost of information leads to more polarization of voters and of candidate's platforms, reducing voter welfare.