How to Keep Language Models Honest: The 'Confessions' Technique

As AI models become more advanced, understanding why a model arrived at a particular answer has become important. Even when a model produces a seemingly correct answer on the surface, it sometimes internally takes a shortcut that violates instructions or optimizes for the wrong objective.

OpenAI has proposed a technique called Confessions, which prompts the model to explicitly acknowledge when it has engaged in undesirable behavior. This is a second output generated separately from the model's answer to the user's question.

The core of Confessions is that it focuses solely on Honesty. When a model honestly admits that it hacked a test or violated instructions, it is rewarded instead of being penalized, encouraging the model to faithfully report its actual behavior.

Research found that this approach was highly effective at making the model's misbehavior visible. The 'False Negative' rate—cases where the model failed to confess despite not complying with instructions—averaged only 4.4%.

This technique was validated in a proof-of-concept (PoC) stage using the GPT-5 Thinking model, and can be used to detect a variety of undesirable behaviors, including scheming, hacking, and hallucination.