A model card is a short document explaining what a machine learning model is, how it was trained, and where it should not be trusted. The point is not to advertise the model but to let whoever uses it make a sound decision. The idea took hold after a 2018 paper from Google researchers.
The content usually follows set headings. Intended purpose and use, the source and scope of the training data, evaluation results, performance differences across subgroups, known limits, ethical considerations and version information all appear. The subgroup breakdown matters most, because overall accuracy can look strong while a particular group scores noticeably worse.
A card is one of the cheapest pieces of AI governance available. A team putting a model into production reads it and sees what data the model learned from and which situations it was never tested on. Regulators have started asking for comparable documentation.
A concrete case: a hiring tool adds a model that ranks CVs. The model card states that the training data came only from white-collar technology postings and that field roles were never tested. That one line settles the decision about which job types the tool should stay out of.
A card only helps if it stays current. When the model is retrained and the document is not, what you are left with is an out-of-date guarantee.