TLDR:
A model card is a standardized documentation artifact for an AI model—covering its intended use, training data, performance characteristics, limitations, ethical considerations, and known risks. The model card concept was introduced by Google researchers in 2018 and has become standard practice across responsible AI deployment.
Standard Model Card Contents
A complete model card typically includes: model details (name, version, type, developer, citation), intended use (primary use cases, users, out-of-scope uses), factors (relevant factors that affect performance—demographics, environmental conditions), metrics (chosen evaluation metrics and disaggregated performance across subgroups), evaluation data, training data, quantitative analyses including fairness analysis, and ethical considerations including known risks and mitigations.
Why Model Cards Matter
Model cards serve multiple stakeholders: developers downstream understand whether the model fits their use case; users understand limitations; auditors and regulators have a structured artifact to assess; and the broader community can compare models on consistent dimensions. Major foundation model providers (Anthropic, OpenAI, Google, Meta) publish model cards or system cards for new releases. Hugging Face hosts thousands of community model cards.
Regulatory Adoption
Model cards have moved from voluntary best practice toward regulatory requirement: the EU AI Act mandates technical documentation that closely mirrors model card content for high-risk AI systems; the EU’s GPAI Code of Practice requires model documentation for general-purpose AI; sectoral regulators (FDA for medical AI, financial regulators for credit models) increasingly require structured model documentation. Enterprises building AI systems should adopt model cards as standard practice early—retrofitting documentation across many models is significantly more expensive than building documentation alongside development.