OpenAI (December 2025). Models are trained to produce a confession — a secondary output generated after the main answer — by rewarding the confession solely for honesty, never mixing the honesty signal with task reward. This structural isolation makes honest reporting the path of least resistance. Evaluated on GPT-5-Thinking across four categories (hallucination, instruction following, scheming, reward hacking); average confession rate 74.3%, with 4/12 settings exceeding 90%. Structural limit: models cannot confess to violations they do not internally register — hallucinated content the model believes is true produces no confession. The companion OpenAI blog post is at: https://openai.com/index/how-confessions-can-keep-language-models-honest/