A model card is a short document that accompanies a trained machine learning model and explains what it is for, how well it works, and where it should not be used. The idea was introduced in “Model Cards for Model Reporting,” submitted to arXiv on October 5, 2018 and revised in January 2019, by Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru, and presented at the FAT* conference.
The paper’s central recommendation is disaggregated evaluation. Rather than reporting a single headline accuracy number, a model card breaks performance down across different cultural, demographic, or phenotypic groups - and their intersections - so that a model that works well on average but poorly for a particular population is not quietly shipped as “accurate.” A model card also records intended use cases, evaluation conditions, training data, and ethical considerations.
Model cards became a widely adopted practice; major model hubs and AI providers now publish them as a matter of routine, and they pair naturally with datasheets for datasets on the data side. For a business reader, a model card is the closest thing to a nutrition label for an AI system: a concise, structured way to judge whether a model is appropriate and safe for a given purpose before deploying it.