Abstract: Animals can acquire information from the environment privately, by sampling it directly, or socially, through learning from others. Generally, private information is more accurate, but expensive to acquire, while social information is cheaper but less reliable. Accordingly, the 'costly information hypothesis' predicts that individuals will use private information when the costs associated with doing so are low, but that they should increasingly use social information as the costs of using private information rise. While consistent with considerable data, this theory has yet to be directly tested in a satisfactory manner. We tested this hypothesis by giving minnows (Phoxinus phoxinus) a choice between socially demonstrated and non-demonstrated prey patches under conditions of low, indirect and high simulated predation risk. Subjects had no experience (experiment 1) or prior private information that conflicted with the social information provided by the demonstrators (experiment 2). In both experiments, subjects spent more time in the demonstrated patch than in the non-demonstrated patch, and in experiment 1 made fewer switches between patches, when risk was high compared with when it was low. These findings are consistent with the predictions of the costly information hypothesis, and imply that minnows adopt a 'copy-when-asocial-learning-is-costly' learning strategy.