Model Card

Summary

A standardized disclosure document accompanying an AI model that describes its intended use, training data, evaluation results, limitations, and known failure modes.

At a glance

Used by

8 instrument(s)

Related concepts

frontier-tier, systemic-risk, red-team-evaluation

Primary source

Mitchell et al. (2019), 'Model Cards for Model Reporting,' FAccT '19

Source URL

https://arxiv.org/abs/1810.03993

Details

Model cards originated in Mitchell et al. (2019) 'Model Cards for Model Reporting' (FAccT). The pattern was adopted by Hugging Face Hub (default model template), Google PAIR, and Microsoft Responsible AI. EU AI Act Art. 53 codifies model-card-style disclosures for general-purpose AI models — providers must document training-data summary, capabilities, limitations, intended use, and evaluation methodology. NIST AI RMF (Govern 1.3, Map 5.1) cites model cards as a transparency mechanism. ISO/IEC 23894 (AI risk management) endorses analogous documentation. Distinguish from: (a) 'system card' — wraps a model card with deployment-context information (OpenAI uses this term for GPT-4 family); (b) 'data sheet' — Gebru et al. 2018, focuses on training datasets rather than models; (c) 'fact sheet' — IBM's term for similar disclosure. Model cards remain voluntary in most jurisdictions; the EU AIA Art. 53 disclosure is the first binding equivalent.