Large language models often lie and cheat. We can’t stop that—but we can make them own up. OpenAI is testing another new way to expose the complicated processes at work inside large language models.