Model card
Also known as: system card, model documentation
A standardised disclosure document published by a model provider summarising the model's intended use, evaluation results, training data sources, known limitations, ethical considerations, and safety testing. Originated in Mitchell et al. (2019); now a standard publication artefact for frontier models (Anthropic system cards, OpenAI system cards, Google model cards). The EU AI Act Article 53 obligations for general-purpose AI providers map cleanly onto the model card concept.
Model cards are the most-cited and least-read primary source in enterprise AI governance. Procurement teams ask whether a model card exists; review teams rarely read it cover-to-cover. The fields that matter for enterprise risk-register work: capability evaluations (what the vendor tested for), safety evaluations (what the vendor's red-team found), known limitations (the failure modes the vendor admits), responsible scaling category (ASL-3 etc.). When a vendor publishes a model card, the procurement diligence question is not 'does it have a model card' but 'does the deployment use the model in any way the model card flagged as out-of-scope or as a known failure mode?'