AI Model Card.

Model cards were proposed in the 2018 paper "Model Cards for Model Reporting" (arXiv:1810.03993) by Margaret Mitchell et al. at Google, later presented at the ACM FAccT (then FAT*) 2019 conference. They standardise transparent reporting for ML models with sections such as intended uses and out-of-scope uses, training and evaluation data, quantitative performance disaggregated by demographic or operational slice, known ethical considerations, and recommendations. Major providers — Google, Hugging Face, OpenAI, Meta, and Microsoft — publish model cards, and on Hugging Face the card is simply the README.md with YAML metadata (license, datasets, metrics) that powers the Hub's search and filtering.

The format underpins regulatory transparency: the EU AI Act requires technical documentation for high-risk and general-purpose models, and the NIST AI Risk Management Framework (AI RMF 1.0) treats model documentation as a core "Map/Measure" practice. From a security standpoint, a model card is a supply-chain artifact — closely related to an AI Bill of Materials (AI-BOM) — that lets defenders reason about data lineage, licensing risk, and known failure modes before integrating a third-party model.

But cards are self-reported and unsigned: they can be stale, incomplete, or outright falsified, so a card is a starting point for due diligence, not proof of safety. Pairing model cards with signed provenance (e.g. C2PA content credentials, Sigstore-signed weights) and independent red-team evaluation closes that trust gap.

flowchart TD
  T[Model trained / fine-tuned] --> A[Author writes model card]
  A --> S1[Intended & out-of-scope uses]
  A --> S2[Training / eval data and licence]
  A --> S3[Sliced performance metrics]
  A --> S4[Ethical and safety considerations]
  S1 & S2 & S3 & S4 --> C[Published model card / README.md]
  C --> R{Consumer due diligence}
  R -->|maps to| G[AI governance, EU AI Act, NIST AI RMF]
  R -->|verify with| V[Signed provenance + red-team review]