What motivates human behaviour in social dilemmas? The results of public goods games are commonly interpreted as showing that humans are altruistically motivated to benefit others. However, there is a competing 'confused learners' hypothesis: that individuals start the game either uncertain or mistaken (confused) and then learn from experience how to improve their payoff (payoff-based learning). Here we (1) show that these competing hypotheses can be differentiated by how they predict contributions should decline over time; and (2) use metadata from 237 published public goods games to test between these competing hypotheses. We found, as predicted by the confused learners hypothesis, that contributions declined faster when individuals had more influence over their own payoffs. This predicted relationship arises because more influence leads to a greater correlation between contributions and payoffs, facilitating learning. Our results suggest that humans, in general, are not altruistically motivated to benefit others but instead learn to help themselves.