Journal article: We study social learning by boundedly rational agents. Agents take a decision in sequence, after observing their predecessors and a private signal. They are unable to make perfect inferences from their predecessors' decisions: they only understand the relation between the aggregate distribution of actions and the state of nature, and make their inferences accordingly. We show that, in a discrete action space, even if agents receive signals of unbounded precision, there are asymptotic inefficiencies. In a continuous action space, compared to the rational case, agents overweight early signals. Despite this behavioral bias, eventually agents learn the realized state of the world and choose the correct action.