Abstract Humans regularly intervene in others' conflicts as third-parties. This has been studied using the third-party punishment game: A third-party can pay a cost to punish another player (the "dictator") who treated someone else poorly. Because the game is anonymous and one-shot, punishers are thought to have no strategic reasons to intervene. Nonetheless, punishers often punish dictators who treat others poorly. This result is central to a controversy over human social evolution: Did third-party punishment evolve to maintain group norms or to deter others from acting against one's interests? This paper provides a critical test. We manipulate the ingroup/outgroup composition of the players while simultaneously measuring the inferences punishers make about how the dictator would treat them personally. The group norm predictions were falsified, as outgroup defectors were punished most harshly, not ingroup defectors (as predicted by ingroup fairness norms) and not outgroup members generally (as predicted by norms of parochialism). The deterrence predictions were validated: Punishers punished the most when they inferred that they would be treated the worst by dictators, especially when better treatment would be expected given ingroup/outgroup composition.