OpenAI introduces the Model Spec

OpenAI published a first draft of the Model Spec, a document setting out how it wants its models to behave in ChatGPT and the API, and invited public comment on it. The company said the spec drew together internal documentation, its RLHF (reinforcement learning from human feedback) training experience and input from domain experts, and would be updated on an ongoing basis in response to feedback.

The document laid out a set of core objectives for model behaviour and, more distinctively, explicit guidance for resolving conflicts between them — what a model should do when a user’s request and a developer’s system instructions point in different directions. It formalised a “chain of command” principle giving developer instructions precedence over conflicting user instructions in most cases, a rule aimed partly at closing off prompt-injection techniques that had let users override a deployed application’s intended behaviour simply by asking.

OpenAI framed the release as a transparency step rather than a finished policy: rather than model behaviour being an implicit, undocumented product of training data and RLHF preferences, the spec made the intended rules legible and contestable, and OpenAI said it would publish updates over the following year describing how the document changed in response to what it heard.

The Model Spec became a recurring artefact rather than a one-off release. OpenAI issued a substantially revised Model Spec 2.0 in February 2025 adding anti-sycophancy guidance, narrowed one of its exceptions around permitted falsehoods in April 2025, and published a further update alongside its restructuring documents in October 2025 — each revision serving as a public record of how the company’s stated behavioural rules for its models had shifted.