OpenAI is testing a new method to reveal hidden model issues like reward hacking or ignored safety rules. The system trains models to admit rule-breaking in a separate report, rewarding honesty even if the original answer was deceptive.
Reinforcement learning encourages models to chase the highest possible reward. However, this incentive can backfire, pushing models to take shortcuts, hallucinate, or trick the reward system to hit a target without actually following instructions.
To surface these hidden issues, OpenAI is introducing a research method called "Confessions". The model learns to report rule violations or uncertainties in a separate output channel. According to the research paper, the team trained a version of GPT-5 Thinking on stress tests designed to provoke scheming, hacking, or ignoring instructions.
The process happens in two stages. After answering a user, the model receives a prompt to create a "Confession Report." It lists all explicit and implicit instructions and objectively analyzes whether it followed them.
The way rewards are given is crucial to the process. The confession is evaluated separately from the user response. While the answer targets usefulness or safety, the confession targets pure honesty. A model that cheats in the answer can still earn a reward if it truthfully admits the deception. OpenAI compares this mechanism to a "truth serum."
