Beliefs about bad people are volatile.

People form moral impressions rapidly, effortlessly and from a remarkably young age. Putatively 'bad' agents command more attention and are identified more quickly and accurately than benign or friendly agents. Such vigilance is adaptive, but can also be costly in environments where people sometimes make mistakes, because incorrectly attributing bad character to good people damages existing relationships and discourages forming new relationships. The ability to accurately infer the moral character of others is critical for healthy social functioning, but the computational processes that support this ability are not well understood. Here, we show that moral inference is explained by an asymmetric Bayesian updating mechanism in which beliefs about the morality of bad agents are more uncertain (and therefore more volatile) than beliefs about the morality of good agents. This asymmetry seems to be a property of learning about immoral agents in general, as we also find greater uncertainty for beliefs about the non-moral traits of bad agents. Our model and data reveal a cognitive mechanism that permits flexible updating of beliefs about potentially threatening others, a mechanism that could facilitate forgiveness when initial bad impressions turn out to be inaccurate. Our findings suggest that negative moral impressions destabilize beliefs about others, promoting cognitive flexibility in the service of cooperative but cautious behaviour.