OpenAI’s fix for hallucinations? Incentivising AI models to say ‘I don’t know’