OpenAI says its latest chatbot should make fewer “hallucination” errors based on a measurement system the company devised. So how does it work, and what does it show?