ISO Reference Pack for AI Compliance
← All guidesTemplates & Documentation Practices

TEMPLATE_model-cards

Source
—

Model Cards — an AI transparency documentation template

This file is OTG's original research on "Model Cards," written after confirming via public search that the concept traces to a real, citable academic paper and that genuinely free, public templates implementing it still exist and are actively maintained. It is not a rewrite of any one source; it draws on the sources listed at the end.

Origin

Model Cards were introduced in the paper "Model Cards for Model Reporting" by Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru (Google), presented at FAT* '19 (the ACM Conference on Fairness, Accountability, and Transparency) in January 2019. The paper is freely available on arXiv (ID 1810.03993) and via Google Research's own publication page — it is not paywalled.

The paper's proposal: a model card is a short document, shipped alongside a trained ML model, that reports how the model performs across different conditions — including across demographic, cultural, or phenotypic groups where relevant — rather than reporting only a single headline accuracy figure. The original paper specifies nine sections: model details, intended use, factors, metrics, evaluation data, training data, quantitative analyses, ethical considerations, and caveats and recommendations. The stated goal was to make trained models more transparent to the range of people who might build on, deploy, or be affected by them, as a step toward more responsible and accountable machine learning practice.

Current public status — genuinely free, actively maintained templates exist

Two lineages of freely usable Model Card templates could be confirmed via public search:

  • Google's own tooling. Google released the open-source Model Card Toolkit (announced July 2020 on the Google AI / Research blog), which provides a JSON/proto schema for model card fields plus pre-built Jinja templates (HTML and Markdown) that can be freely used or modified. It's documented as part of TensorFlow's public "Responsible AI Toolkit" documentation site.
  • Hugging Face's model card format. Hugging Face maintains a public model card template — a Markdown file with a YAML metadata header — used across the Hugging Face Hub for documenting published models. Hugging Face publishes an annotated version of the template explaining what to put in each section, a "Model Card Guidebook," and the raw template file itself in the open-source huggingface_hub library on GitHub. Any of these can be copied and reused without cost or license restriction beyond the library's own open-source license.

Both lineages descend from the same nine-section structure in the original Mitchell et al. paper, adapted to their respective platforms' needs (Hugging Face's version is oriented around a single published model page; Google's toolkit is oriented around a pipeline that generates the card programmatically).

Relationship to ISO/IEC standards — described generically, not overclaimed

Public search did not turn up a specific ISO/IEC standard that names "model cards" by that term in its normative text. What could be verified:

  • ISO/IEC 12792:2025, Transparency taxonomy of AI systems, specifies a taxonomy of information elements meant to help AI stakeholders identify and address AI transparency needs — including documenting an AI system's internal functioning and its datasets. This covers the same underlying documentation practice a model card implements (structured, standardized disclosure of what a model does, what data it was trained on, and its known limits), but it should be described as a related, generic transparency-documentation practice, not as a standard that specifically mandates the "model card" artifact by name.
  • ISO/IEC 42001 (AI management system) requires "documented information" under its Clause 7.5 as part of running an AI management system. Several third-party ISO/IEC 42001 implementation guides and auditor commentary (not the standard's own text, as far as could be verified) describe model cards as one practical way organizations satisfy that documented information requirement for a given AI system, and note that certification auditors commonly ask to see model cards or equivalent system documentation. This is a secondary, practitioner-community claim about how the standard is applied — not a verified requirement written into ISO/IEC 42001 itself — and is presented here with that caveat rather than as settled fact.

In short: model cards are a widely adopted, freely available AI documentation practice that maps naturally onto the "system documentation" and "transparency" concerns that several AI governance standards raise in more general terms — but readers should not assume any specific ISO/IEC standard requires the exact Mitchell et al. template or uses the term "model card."

Sources