Decisions in social dilemmas lead to outcomes for oneself and others. These outcomes can be gains or losses, yet we lack a full understanding of how people’s decisions depend on which outcomes are above or below zero. We systematically varied whether the outcomes of social dilemmas (Prisoner’s Dilemma, Stag-Hunt, Chicken) were losses, gains, or combinations thereof. Across 7 experiments (4 preregistered; N Offline = 197, N Online = 1,653), participants consistently tried to avoid losses altogether (loss avoidance), but they did not try to minimise losses (loss aversion). If cooperation avoided losses, people cooperated more, if defection avoided losses, people defected more, even if this imposed a loss on someone else. Loss avoidance was larger for one-shot than for iterated games and was present in all games studied. Our results suggest that loss avoidance, rather than loss aversion, systematically influences how people cooperate.