# Model Card A standardized documentation format for AI models — proposed by Mitchell, Gebru, and colleagues in 2019 — that discloses a model's intended use cases, training data, performance metrics across demographic groups, known limitations, and ethical considerations. The concept is explicitly modeled on nutrition labels: just as food packaging discloses ingredients and allergens, model cards disclose what went into the AI and what risks it carries. ## Why the term matters Model cards represent the **documentation-as-governance** approach to AI accountability. The premise: if you can't regulate what's inside the model (the weights are opaque, the training process is proprietary), you can at least require disclosure about the model's characteristics, limitations, and intended uses. The practice has been widely adopted — [[Hugging Face]] hosts model cards for thousands of models, Google and Meta publish them for major releases, and the EU AI Act's transparency requirements draw on the model-card framework. But adoption varies wildly in quality: some model cards are substantive documents with detailed performance breakdowns across demographic groups; others are pro-forma checkbox exercises that disclose little of value. The irony: model cards were created by researchers (Mitchell, Gebru) who were subsequently fired from Google for their AI ethics work. The governance tool they created was adopted by the industry; the governance *ethos* they represented was not. ## See Also - [[AI Safety]] — the governance framework model cards serve - [[Interpretability]] — the deeper transparency that model cards can't provide - [[Responsible AI]] — the corporate framework model cards live within