Our ability to learn about the reputations of others--that is, who is likely to cooperate versus cheat--contributes greatly to cooperativeness in society. There has been recent debate whether humans employ memory bias favoring cheaters (i.e., there is an evolved module for the detection of cheaters) or whether no such bias exists (i.e., reputation learning is flexibly modulated by contextual factors). We report 3 experiments that address this issue by comparing persistence against extinction--which is a reliable measure of prepared fear learning (Öhman & Mineka, 2001)--between memories regarding cheaters and cooperators. In all experiments, participants learned to classify unfamiliar persons as either cooperators or cheaters, and, then, they were instructed to disregard those learned associations and told that they had been determined arbitrarily, which simulated a verbal extinction procedure in the fear conditioning paradigm (Hugdahl & Öhman, 1977). The results indicated that while postlearning changes in perceived trustworthiness were modulated by a contextual factor (appearance of the facial stimulus), the persistence of learning exhibited a cheater advantage: Cheaters remained perceived as untrustworthy to a greater extent than cooperators as trustworthy at the extinction period. Thus, there exists a cheater bias in human reputation learning, the proximate and ultimate mechanisms of which warrant further study.