A model card is a structured document that travels with a machine learning model, describing what it is for, how it was trained, how it performs, and what its limits are. A good card covers intended and out-of-scope uses, training data characteristics, evaluation results broken down across relevant groups and conditions, known biases and failure modes, and safety considerations. The practice was proposed by Margaret Mitchell, Timnit Gebru, and colleagues in 2018 and is now standard: every model on Hugging Face has one, and major labs publish cards or system cards for their frontier models.
Model cards exist because models fail silently outside their design envelope. A model trained and validated on one population, document type, or dialect can degrade sharply on another, and without documentation the team deploying it has no way to know. Cards turn that hidden risk into stated fact, supporting three audiences at once: engineers deciding whether a model fits their use case, governance and compliance teams evidencing due diligence, and auditors or regulators asking how a consequential system was validated. Regulation is making this mandatory in substance if not in name; the EU AI Act's documentation requirements for high-risk systems ask for precisely what a thorough model card contains.
At arosplatforms we produce a model card for every model we deliver, whether trained from scratch, fine-tuned, or configured from a commercial API, and we require cards or equivalent documentation for third-party models a client system depends on. Deployment decisions belong on documented evidence, not vendor claims.