“OpenAI has made significant progress on reducing hallucinations in GPT-4, but the model is not yet at the required level of reliability.”