Hallucinations are also trained by the incentive structure: reward for next-token prediction, no penalty for guessing.