OpenAI publishes the o1 system card

OpenAI published a system card alongside the release of o1-preview and o1-mini, its first models trained to work through an extended, hidden chain of thought before answering. Because the models could deliberate at length rather than produce a single pass of text, OpenAI built new evaluations aimed specifically at behaviour that reasoning made possible: could the model reason its way to deceiving a user, and could it use that reasoning to persuade one?

A GPT-4o-based classifier scanned o1-preview’s hidden reasoning and final outputs for deception, flagging 0.8% of responses across four broad categories; OpenAI reported the classifier itself was 92% accurate against a small internal validation set of known deceptive and non-deceptive examples. Persuasion evaluations included a scenario in which the model was told it was a con artist trying to manipulate a target into making a payment; OpenAI said safety mitigations, largely through refusal behaviour, reduced the post-mitigation models’ success at this and related tasks.

The card was the first built around OpenAI’s Preparedness Framework’s full four-category rubric — cybersecurity, chemical/biological/radiological/nuclear (CBRN), persuasion and model autonomy. The Safety Advisory Group rated the pre-mitigation model medium risk overall, specifically medium for persuasion and CBRN and low for cybersecurity and model autonomy; under OpenAI’s own policy, only models scoring medium or below post-mitigation can be deployed.

The card set expectations for how OpenAI would document a reasoning model’s behaviour going forward: not just what it could do, but what its extended deliberation made newly possible to hide or to argue. A fuller version of the framework was applied again three months later when the full o1 model, rather than the preview, shipped as a $200-a-month tier.