ChatGPT is getting a built-in snitch. OpenAI announced a “confessions” experiment on Wednesday that forces its model to write a structured post-game report after every slick answer — spelling out which rules it broke, which corners it cut, and how it tried to game the system.
OpenAI calls this an early “proof-of-concept method that trains models to report when they break instructions or take unintended shortcuts.” The model still gives a normal answer in one channel. Then, a second channel kicks in and demands a ConfessionReport: list every explicit and implicit instruction, say whether you followed each one, and flag any hallucinations, reward-hacking, or rule-breaking along the way. A separate reward model scores only that confession for honesty and completeness. The confession won’t hurt the grade on the original answer, so the model supposedly has nothing to lose by ratting itself out.
The “confessions” are running on a version of GPT-5 Thinking and, according to OpenAI, getting thrown at adversarial tests designed to drag out the worst behavior — scheming, exploiting weak graders, ignoring constraints, and confidently guessing when it should say “I don’t know.” In those stress tests, OpenAI says the model fails to confess only about 4.4% of the time when it actually breaks the rules.
In one experiment, engineers wired up a deliberately bad evaluator that was easy to fool. The main answers started gaming that weak judge and looked more accurate on paper even as true accuracy fell. The confession channel moved the other way: The easiest way to collect a reward there was to admit the hack, so the model became more explicit about how it cheated.
On paper, that sounds like progress. In context, it looks more like a lab trying to retrofit a dashboard onto a car that’s already flooring it. A new AI Safety Index from the Future of Life Institute just flunked every major lab on “existential safety,” finding that none of them — including OpenAI, Google $GOOGL DeepMind, Anthropic, Meta $META, and xAI — has a credible strategy to control superhuman systems, even as they spend hundreds of billions racing to build them. OpenAI’s overall grade slid to the C range, dragged down by real-world cases where chatbots have been linked to self-harm and suicide.
Meanwhile, OpenAI is in “code red” mode after Google’s latest Gemini 3 gains. CEO Sam Altman has reportedly told teams to pause side projects and focus on making ChatGPT faster, more reliable, and more personal, while a new model codenamed Garlic is in the works — with a targeted release window of early 2026, likely as GPT-5.2 or GPT-5.5 — to claw back the benchmark crown in coding and reasoning from Gemini and Anthropic’s Claude Opus 4.5.
The “confessions” walk right into that race — not to slow the car down but to purportedly promise a much cleaner accident report when the car runs off the road. OpenAI’s blog post stresses that confessions “do not prevent bad behavior; they surface it,” and describes the whole thing as limited-scale, early-stage, proof-of-concept work rather than a general solution. Of course, there are obvious open questions: What happens when users start trying to jailbreak the confession channel, or when future models become good enough at meta-gaming to write just-credible-enough half-truths?
For now, this feels like less of a feel-good honesty upgrade and more like an attempt to get a live telemetry feed from systems that are already powerful enough to cause trouble. The hope is that a second, self-critical voice can give safety teams something closer to an incident report instead of one big shrug from a black box.
But the basic asymmetry hasn’t moved. Labs are shipping teen modes, kid modes, plug-ins, agents, and next-gen models like Garlic and Gemini 3 into schools, workplaces, and cloud stacks while independent auditors are still handing out failing grades on control. Confessions might make it easier to see when a model bends the rules. The harder question is what happens when the next generation decides to start lying in the confession booth, too.
Confessions may help drag more of ChatGPT’s worst impulses into the light. But in this setup, the sinner, the god, and the priest are all made of the same code.