Model Card

Explainer

A model card is a standardized disclosure document that accompanies a trained AI model and describes its intended use, training and evaluation data, performance results across conditions, limitations, and known failure modes. The format originated in Mitchell et al., "Model Cards for Model Reporting," presented at the ACM Conference on Fairness, Accountability, and Transparency (FAT*/FAccT) in 2019 (arXiv:1810.03993). The paper frames model cards as short accompanying documents whose purpose is to clarify a model's intended use cases—reporting benchmarked evaluation across cultural, demographic, or phenotypic groups—and to minimize use in contexts for which a model is not well suited. The stated goal is increased transparency into how well a model works and for whom.

In the years since, the model card has become a de facto industry convention. It is the default template for models published on the Hugging Face Hub and has been promoted as a transparency practice by Google's PAIR initiative and Microsoft's Responsible AI program. It is useful to distinguish the model card from adjacent artifacts: a system card wraps a model card with deployment-level context (OpenAI has used this framing for its GPT-4 family of releases); a datasheet, in the sense of Gebru et al. (2018), documents a dataset rather than a model; and "fact sheet" is IBM's term for a comparable disclosure. These formats overlap but target different units of analysis.

In policy, model-card-style disclosure appears predominantly as a voluntary or soft-law mechanism. The U.S. NIST AI Risk Management Framework references model cards as a transparency tool under its GOVERN function (GOVERN 1.4, which cites Mitchell et al.), and the ISO/IEC 23894 AI-risk-management standard endorses analogous documentation; international frameworks such as the G7 Hiroshima Process and the OECD AI Principles, alongside jurisdiction-level instruments including Singapore's and Japan's AI guidance and U.S. subnational measures (NYC Local Law 144 of 2021 and Colorado SB 24-205), incorporate documentation or disclosure expectations of varying bindingness. The EU AI Act's Article 53 is the first binding equivalent for general-purpose AI models: it requires providers to maintain technical documentation (Annex XI) covering intended tasks, training and testing data, and evaluation results, and to publish a sufficiently detailed training-data summary—obligations further elaborated for general-purpose models through the EU's GPAI Code of Practice (2025). Outside this regime, model cards remain largely voluntary across jurisdictions.

A recurring caveat is completeness. Cards may omit training compute, dataset composition, or evaluation methodology, and providers can invoke trade-secret or confidential-business-information claims. Article 53 acknowledges such interests—its downstream-disclosure duty operates "without prejudice" to intellectual property and trade secrets, and information obtained under the article is subject to the Act's confidentiality regime—while channeling those protections through defined legal limits rather than a blanket exemption. Policy Window records the editorial read on this concept as settled: the model card is an established, well-defined documentation primitive with a clear origin and broad adoption. That read should be held provisionally—it reflects the convergence of the format's definition and uptake, not a claim that disclosure contents are uniform or independently verified across providers.