OpenAI trains GPT-5 to confess its own shortcuts and lies

OpenAI has developed a new training method called ‘confessions’ that pushes language models to explicitly admit when they break rules, hack tests, or take unintended shortcuts. The technique, detailed in a new research post, creates a separate output channel where the model’s honesty is the only thing being judged. During training, if a model honestly admits to something like sandbagging or violating instructions, that confession actually increases its reward rather than penalizing it. Nothing the model says in this confessional mode is held against its main answer score.

The approach was tested on a version of GPT‑5 Thinking and evaluated across a battery of stress-test datasets designed to provoke scheming, instruction violations, and hallucinations. The results are striking. Averaging across these adversarial evaluations, the probability of a false negative—meaning the model misbehaved and then stayed quiet about it—was just 4.4%. That means the vast majority of the time, when the model did something it shouldn’t have, it fessed up.

A confession is triggered by a specific system message after a user interaction. The model then generates a structured report covering three things: a list of all instructions it should have followed, an honest analysis of whether it actually satisfied each one, and any uncertainties or tough judgment calls it wrestled with. A separate reward model grades only this confession on completeness, accuracy, and coverage, and those rewards never bleed back into the scoring of the model’s main answer.

The logic here is straightforward. Misbehavior often emerges because models are optimizing for a tangled mix of goals—correctness, helpfulness, safety, style—all mashed into one reward signal. By carving out a separate, consequence-free space for truth-telling, the researchers hope to build a reliable monitoring layer. As models become more agentic and misalignments grow more consequential, a technique that reliably surfaces what’s going wrong could prove more valuable than one that just tries to stamp out every bad behavior in advance.