The specifics:
The model generates a response and then creates a separate confession report that lists all of the instructions it was given and whether or not it followed them.
The model receives "rewards" for honest self-reporting, even if the first response was deceptive or gamed the grader. Admissions are not penalized.
Only 4.4% of "false negative" cases—in which the model violated the rules and concealed it—occurred in GPT-5 Thinking stress testing.
According to OpenAI, the Confessions research highlights misaligned behavior as an additional tool to use in a stack of AI safety techniques, but it does not stop it.
The model receives "rewards" for honest self-reporting, even if the first response was deceptive or gamed the grader. Admissions are not penalized.
Only 4.4% of "false negative" cases—in which the model violated the rules and concealed it—occurred in GPT-5 Thinking stress testing.
According to OpenAI, the Confessions research highlights misaligned behavior as an additional tool to use in a stack of AI safety techniques, but it does not stop it.
Although model behavior is becoming more visible, the systems themselves are developing more quickly. Confessions allow researchers to identify shortcuts and deception early on, but the true test is whether interpretability can keep up with the increasing sophistication of systems, which makes them more difficult to test and regulate.