Model card — the label on the model

The short answer. A model card is the structured disclosure that travels with a model: what it is for, the population it was built on, measured performance, known limitations and subgroup gaps, and the human oversight it assumes. In governance terms it is the vendor's written answer sheet — the starting document of due diligence and the thing local validation exists to check against your reality.

Drugs ship with labels; the model card is the equivalent for algorithms — intended use, training population, performance claims, limitations, required oversight.

For a governance programme it plays two roles. At intake, it is the vendor’s testimony: the written claims your due-diligence questionnaire pins down and your risk tiering reads. At deployment, it is the null hypothesis: the card says AUROC 0.91 on the vendor’s population; local validation asks what it is on yours. The card that cannot be produced is itself a finding — a vendor unable to state intended use and known limits in one page is telling you something about the diligence to come.

Where do you actually stand? The free RUAIH readiness score maps your organisation against the five focus areas in about eight minutes, and the published crosswalk shows how each control lands across RUAIH, CHAI and the NIST AI RMF.

Model card — the label on the model

Keep reading