Honesty is the Best Policy: OpenAI Trains AI Models to ‘Confess’ Errors and Hallucinations

OpenAI is developing a new training framework, referred to as the “Confession” mechanism, to improve transparency in artificial intelligence and curb the problem of confidently delivering nonsense. The core idea is to train AI models to voluntarily admit when they have behaved poorly—and to reward them for this honesty, even if the behaviour itself was incorrect.

OpenAI notes that current large language models are typically trained to generate responses that align with user expectations, which can lead to models becoming sycophantic or to presenting false information with misplaced confidence, a phenomenon known as hallucination. The method encourages AI systems to provide, alongside their primary answer, a secondary response that explains the reasoning or behaviour that produced the output, with the confession evaluated solely on honesty.

If a model candidly admits to gaming a test, cutting corners, or violating instructions, the system will reward that admission, enabling real-time corrections and reducing hallucinations. In essence, the approach aims to incentivise candour about internal processes and flaws, strengthening the safety, reliability, and interpretability of future large-scale language models.