OpenAI is teaching AI models to 'confess' when they hallucinate — here's what that actually means

OpenAI has introduced a new research method called “confessions,” which trains AI models to self-report when they take shortcuts or break ...