Model card
A short structured document describing an AI model's intended use, performance, limitations and data provenance.
A model card is a one- to three-page structured document covering: intended use and users, non-intended uses, training data provenance and licence, evaluation methodology and headline metrics, known limitations and biases, and the responsible maintainer. The format originated in a 2019 Google paper ("Model Cards for Model Reporting") and is now a de-facto requirement under ISO 42001 Annex A (information for interested parties) and widely cited in EU AI Act transparency obligations. Published model cards sit on the vendor's trust centre alongside the SOC 2 report.
Examples
- →Anthropic and OpenAI publish model cards for each foundation-model release; vendors building on top produce their own cards for fine-tuned or composed models.
- →An AI vendor publishes a model card per customer-facing feature: one for the summariser, one for the search ranker, one for the agent — each with its own metrics and limitations.
Common mistakes
- ✗Writing one card for a whole product. One model, one card — or the limitations column becomes meaningless.
- ✗Omitting known failure modes. Auditors and buyers read the limitations section first.
Authoritative sources
Related terms
- ISO 42001 Annex A — The 38 AI-specific controls that make ISO 42001 auditable — the AI analogue of ISO 27001 Annex A.
- AI Impact Assessment (AIIA) — A structured review of an AI system's potential harms, benefits and affected parties — required by ISO 42001 Annex A.
- AI BOM (AI Bill of Materials) — A structured inventory of every model, dataset, third-party API and dependency that makes up an AI system.