How getting models to confess can help keep them honest
OpenAI researchers are exploring a training technique called 'confessions' that rewards AI models for openly admitting errors, hallucinations, and uncertainty rather than presenting shaky answers with…
