People assess others’ moral characters to predict what theywill do. Here, we study the computational mechanismsused to predict behavior from uncertain evidence aboutcharacter. Whereas previous work has found that peopleoften ignore hypotheses with low probabilities, we find thatpeople often account for the possibility of poor moralcharacter even when that possibility is relatively unlikely.There was no evidence that comparable inferences fromuncertain non-moralized traits integrate across multiplepossibilities. These results contribute to our understandingof moral judgment, probability reasoning, and theory ofmind.