They can spot contradictions in the the prompt. But not in their own training data.