What is a model card in AI?

Where does the idea come from?

Researchers proposed model cards in a paper first posted in October 2018, with Margaret Mitchell as first author. It was presented at the FAT* Conference on Fairness, Accountability, and Transparency in January 2019. The authors describe model cards as short documents that accompany trained machine learning models.

A card reports how a model performs in a variety of conditions. Those include different cultural, demographic or phenotypic groups that matter for its intended use. A card also discloses the context in which the model is meant to be used. The aim, the authors wrote, is to clarify intended uses and to minimize use where a model is not well suited.

How do medical regulators use them?

The FDA suggests model cards for AI-enabled medical devices in draft guidance from January 2025. It says they are not required. It calls them a means to consistently summarize a device's characteristics, performance and limitations. An appendix shows an example card, which lists items such as:

• the device's name, version and model architecture
• its regulatory status
• its intended users and intended use

The FDA's guidance page still listed that guidance as a draft in October 2026.

The FDA's August 2026 discussion paper on generative AI devices takes the idea further. It asks whether foundation model developers could voluntarily submit structured model cards or system cards to the agency. The FDA would hold these files confidentially. The paper is for discussion only and is not guidance.

How do health IT rules use the term?

U.S. rules for certified health IT software also use the label. In a proposed rule published on December 29, 2025, the health IT office of the Department of Health and Human Services proposed changes to its certification program. One change would fully remove what it called the AI "model card" requirements from its decision support criterion. The AI in Medicine Regulation Tracker lists that proposal.