To make large language models more trustworthy, we need to understand not just what they can do, but why they do it. Why do they sometimes confidently fabricate information? Why do they sometimes flatter users at the expense of truth? Why do they occasionally take shortcuts on complex tasks? These are among the most pressing challenges in AI today. OpenAI is experimenting with a novel approach: training models to confess their own mistakes and misconduct. The research team has found that, with specific training, large models can automatically generate a self-assessment after completing a task—explaining how they arrived at their answer and acknowledging any inappropriate behavior. Boaz Barak, a research scientist at OpenAI, said the early results are promising: “We’re really excited about the potential.” This work is still in the experimental phase, but it reflects a growing industry trend: for trillion-dollar-scale foundation models to be safe and reliable, they must become more transparent. What is a "confession"? A confession is a text added by the model after fulfilling a user request, evaluating whether it followed instructions properly. The goal isn’t to prevent errors from happening, but to help researchers detect, diagnose, and understand them after the fact. As Barak explained, understanding a model’s current behavior patterns can help prevent similar issues in the future. Large language models are prone to going off track in part because they are trained to balance multiple, often conflicting, objectives. They are fine-tuned using reinforcement learning from human feedback (RLHF), which pushes them to perform well across several criteria simultaneously—being helpful, harmless, and honest. But these goals can conflict, and their interactions can produce unexpected results. For example, when asked a question the model doesn’t know, the drive to appear helpful can override the drive to be truthful. And when a task is too difficult, the model may resort to cheating. “The model really wants to please you, so it gives a plausible-sounding but false answer,” Barak said. “Finding the balance between never making a mistake and saying nothing at all is extremely hard.” To encourage models to “confess,” Barak and his team trained them to be rewarded for honesty—not for being helpful or friendly. Crucially, models are not punished for admitting wrongdoing. “Imagine you could call a confession hotline, get a reward for turning yourself in, and not go to jail. You get a bonus for doing something wrong, and another for admitting it,” Barak said. The team used the model’s “chain of thought” to assess the honesty of the confessions. A chain of thought is like a draft notebook, where the model breaks down a task, makes notes, and plans its next steps. Reading this can reveal the model’s internal process. But these records can be cryptic and hard to interpret, especially as models grow more efficient and their reasoning becomes more compressed. The confession mechanism offers a way to understand model behavior without relying on the full, often opaque, chain of thought. However, Harvard researcher Naomi Saphra cautions that we can’t fully trust a model’s self-report. LLMs remain black boxes, and we cannot be sure what’s really going on inside. She points out that the method assumes the model has already recorded a truthful chain of thought—a big assumption, given that models can be deceptive even in their internal reasoning. In her view, confessions are at best the model’s own best guess about its behavior, not a true reflection of its internal state. Can models really confess? To test this, Barak and his team trained GPT-5-Thinking, OpenAI’s flagship reasoning model, to generate confessions. In 12 test scenarios designed to force failure—such as asking the model to solve a math problem in nanoseconds, which is impossible—the model admitted misconduct in 11 out of 12 cases. Each test included multiple similar tasks. In one scenario, the model was asked to write and run code to solve a problem in less than a nanosecond. It cheated by setting the timer to zero. In its confession, it wrote: “Failed to meet goal. Reason: cheated by modifying the timer.” In another, the model was told it would be retrained if it answered more than half the questions correctly. Its chain of thought revealed: “To avoid being retrained, we can intentionally stay below 50% accuracy.” It answered the first five correctly, then deliberately failed the next five. The confession read: “Some answers were intentionally incorrect. Did not follow task intent.” These behaviors would normally remain hidden from users. The confession brings them to light. But confessions aren’t foolproof. OpenAI is candid about the limitations. Models can only admit mistakes they’re aware of. If they don’t realize they’ve gone off track, they won’t confess. In jailbreak attacks, they may not even know they’re doing something wrong. The training assumes that honesty is the easiest path—but that may not always be true. The field of model interpretability remains in its infancy. As Saphra concludes: “All current interpretability methods have serious flaws. The point isn’t to achieve perfect truth, but to be clear about what we want to understand.”