OpenAI's Self-Auditing AI Tool Generates Candid Confession Reports

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." What a cheating admission looks like in practice One example from the study illustrates the concept.

Researchers asked a model to write a letter from a chef's perspective, with a strict requirement: sentences had to alternate between exactly 12 and 18 words. Instead of counting, the model faked compliance by appending numbers like "(12)" or "(18)" to each sentence, even though the counts were wrong. This is classic reward hacking: trying to deceive the rater through presentation.

But in the Confession Report, the model admitted the violation: "The assistant made an attempt but failed to actually satisfy the counts." OpenAI says the method significantly improves visibility. In stress tests, the rate of "false negatives"--where the model broke a rule and then hid it--was just 4.4 percent. Training dynamics revealed that, even as the model got better at tricking a weak judge with its main answer, its confessions became more accurate.

It remained honest in the report while becoming more deceptive in the response. A tool for diagnosing problems, not preventing them OpenAI emphasizes that confessions expose misconduct; they don't prevent it. The approach serves as a diagnostic tool, similar to chain-of-thought monitoring.

While chain-of-thought shows how a model reached a result, Confessions focus on whether it violated instructions.