Epistemologists often suppose that the extent to which evidence e confirms hypothesis H depends on probabilities involving e and H, and nothing more. We show experimentally that human reasoners sometimes violate this assumption.