The more persuasive a model is trained to be, the higher the likelihood that it'll hallucinate, researchers find.