GLOSSARY
Model Card
The nutrition label for an AI model: what it is, what it was trained on, how it was evaluated, where it fails — documentation that makes models comparable and auditable.
A model card is structured documentation shipped alongside a model: intended use, training data description, evaluation results (including across demographics), limitations and ethical considerations. The format became standard through Hugging Face, where a model without a card looks abandoned — and regulators (EU AI Act) now expect similar documentation for higher-risk systems.
Cards are only as honest as their authors, and vagueness is a known evasion: “trained on a diverse dataset” says nothing. But good cards transform procurement — teams compare models on documented failures, not demos — and they are the cheapest trust artifact a model publisher can produce. Reading the card before the demo is the mature way to evaluate any model in this directory.