Model card
A short structured document published with an AI model — declares intended uses, training data overview, performance across subgroups, known limitations, and risk factors.
Model cards (Mitchell et al., 2018) are the model equivalent of a nutrition label. Sections cover model details, intended use, factors affecting performance, evaluation data, training data, quantitative analyses, ethical considerations, and caveats.
In 2026, model cards are expected by procurement teams in regulated industries (health, finance, legal, government) and required by some jurisdictions for high-risk AI. OpenAI, Anthropic, Google, Meta, and most open-source releases ship them by default.
For agent builders, a model card matters when you are the procurement target. Sales cycles in enterprise grind to a halt without one. For internal models, a model card is also a hygiene practice — it forces you to enumerate your assumptions.
Frequently asked
Is a model card legally required?+
Under the EU AI Act, yes, for high-risk systems. Under NIST AI RMF in the US, recommended. In practice, enterprise procurement asks for one regardless of jurisdiction.