Context card
Model behavior policy and model card: two documents, two time directions
What is a model behavior policy, and how does it differ from a model card?
The short answer
A model behavior policy — published under names such as model spec, constitution or code of conduct — is a document in which a model provider states how its models should behave: which instructions they follow, what they refuse, and how they may describe themselves. It is prospective: it is written to steer training, technical controls and evaluation of future versions. A model card is retrospective and version-bound: it reports what one released model was built from and how it scored.
For: Technical writers and AI governance roles
Key points
- A behavior policy sets intended behavior for models that are not finished yet.
- A model card documents one existing version: intended use, data, results, limitations.
- A draft policy is a statement of intent, not evidence about current products.
- Both are governance documents and neither by itself demonstrates that an obligation is met.
- Version labels and dates decide which document applies to which model.
The context
What a behavior policy says
A model behavior policy states the behavior a provider wants from its models: which instruction sources outrank which, what the model refuses, how it handles sensitive requests, and how it may present itself and its own status. It is written to steer later work, feeding objectives into training, into technical controls, into operational monitoring and into internal review. Because it looks forward, a draft or freshly published policy is a statement of intent. It is not evidence that current releases already behave as described, and careful providers say so in the preface — including the caveat that written objectives cannot guarantee the behavior of a trained model.
Where the model card sits
A model card runs in the opposite time direction. It is tied to one model version and reports what was built and measured: intended and excluded uses, data characteristics, evaluation results, known limitations and residual risks. The policy tells you what an organization aims for across versions; the card tells you what one version did on a fixed set of tests. Neither replaces the other, and neither on its own shows that a legal, contractual or standards-based obligation has been satisfied.
Why both belong in one documentation set
In practice both are objects in a set held together by AI governance, next to usage policies, system descriptions and, in regulated sectors, formal technical documentation. For writers the useful move is traceability: a sentence in the policy should be findable as a named behavior, that behavior as a test in an evaluation report, and the result as a limitation in a model card. Where the chain breaks, readers cannot tell whether a promise was dropped, re-worded or simply never tested.
Dates and version labels carry as much weight as wording. A policy still in public consultation, a policy in force, and the policy that actually shaped a shipped model may be three different documents, and quoting one as if it were another is the most common error in reporting on model behavior.
Questions readers ask next
- Is a behavior policy the same as a system prompt?
- No. A system prompt is runtime text sent with a request; a behavior policy is a governing document that may influence training data, reward signals, system prompts and evaluations alike.
- Who writes these documents in a company?
- Usually a mix of policy, safety and research roles, with legal review. Technical communicators are often involved in making the behavior list, its versioning and its publication consistent.
- Does a published policy apply to all of a vendor's products?
- Only to the scope it names, which is often a specific model family rather than every product built on it. The scope statement, not the brand, decides.
Sources
- Model Spec — OpenAI, December 18, 2025
- openai/model_spec (repository) — OpenAI on GitHub, October 6, 2026
- ISO/IEC 42001 explained — ISO, October 6, 2026
Review log and changes
Every context card is checked against its sources before it is published, and again whenever it changes; the date under the byline is the last review. Corrections (something was wrong) and additions (something was missing) are logged below with date and time (Berlin time). Typos, formatting and link fixes are not listed.
Reviewed
No corrections or additions since publication.