OpenAI’s GPT-5 hallucinates less than previous models do, but cutting hallucination completely might prove impossible go.nature.com/4gkzr4A