RLHF is intended to align the model with human preferences, but human preferences tend to reward 'confident and detailed' responses, creating a paradox that reinforces hallucination