LLMs already know how to distinguish fact from fiction much better than random chance and the base non-RLHF GPT-4 model was excellently calibrated (its predicte

LLMs already know how to distinguish fact from fiction much better than random chance and the base non-RLHF GPT-4 model was excellently calibrated (its predicted confidence in an answer generally matches the probability of being correct). "Eliminating" it is not that important. Getting it to human levels is the goal. and boy do humans often "hallucinate", i.e have a poor grasp of what they do or do not know and confidently spout nonsense.