Model card

Model cards were proposed by Mitchell and colleagues as a transparency practice: every released model should travel with a document explaining what it is for, how it was evaluated and where it should not be used. A typical card covers model details, intended and out-of-scope uses, factors such as demographic groups or environments, metrics, evaluation and training data, disaggregated results, ethical considerations and caveats.

The practice supports several obligations without being a legal form in itself. The EU AI Act requires technical documentation for high-risk systems (Annex IV) and documentation for general-purpose AI models, and auditors use model cards as evidence of what the provider knew and disclosed. Model cards complement datasheets for datasets, which document the data rather than the model.

Example: a team publishing a speech recognition model reports error rates separately by accent and notes that the model was not evaluated for medical dictation, steering deployers away from that use.