Glossary
Model card
Written with AI from the sources below and checked against our facts sheet. Last checked . How this site is made · Spot an error?
Part of: EU AI Act
A model card is a short, standardized document that describes a trained AI model. It states what the model is for and not for, how it was built and evaluated, and how well it performs across groups. It also lists known limits, helping deployers, auditors and users decide whether the model fits their use.
In plain English
A model card is a nutrition label for an AI model. A food label tells you the ingredients, the serving size and the allergens. A model card tells you what the model is for and what it was trained and tested on. It also shows how well it performs for different groups, and when not to use it.
It is short enough to read in a few minutes. It is not the source code, the raw training data or a full compliance file. The idea comes from a 2019 paper by Mitchell and colleagues.
Example
Harrow Analytics, a fictional vendor, releases a model that flags home-insurance claims for fraud review. Its model card reads, in summary:
| Section | What Harrow wrote |
|---|---|
| Model details | Version 2.1, gradient-boosted classifier, released October 2026 |
| Intended use | Prioritizing claims for review by trained fraud investigators |
| Out-of-scope uses | Automatically denying claims; pricing policies; life or health insurance |
| Factors | Claim type, region, policy age |
| Metrics | Precision and recall at the default threshold |
| Quantitative analyses | Results by region and claim type; recall is lower for flood claims |
| Training and evaluation data | Summary of 2023 to 2025 claims, linked to datasheets |
| Caveats and recommendations | Retest after major weather events; never act on a flag without investigation |
A customer that wants to auto-deny claims can see at once that Harrow ruled that use out.
Why it matters
Origin. Mitchell et al. proposed model cards so performance is reported across groups, not hidden in one average. Their companion idea is the datasheet for datasets by Gebru and colleagues.
EU AI Act. The Act never says "model card", but it requires overlapping content:
- High-risk providers must give deployers instructions for use (Art. 13), including intended purpose, accuracy levels and metrics, and known limitations and risks.
- High-risk providers must also keep technical documentation (Art. 11, Annex IV). This is the much more detailed file for authorities.
- GPAI model providers must give downstream providers information on capabilities and limitations (Art. 53(1)(b), Annex XII). The GPAI Code of Practice, published 10 July 2025, includes a Model Documentation Form for this.
Governance. By naming out-of-scope uses, a model card helps forecast and limit secondary uses. Deployers rely on it in vendor due diligence and in their own impact assessments.
How the AIGP exam tests it
The BoK names the model card as part of release readiness (lesson III.C.2), fed by training and testing records (lesson III.B.6).
Common traps:
- Model card vs AI inventory. A model card summarizes one model for deployers and users. An inventory is the internal register of every AI system, with owners and incident plans.
- Model card vs technical documentation. Art. 11 documentation is the detailed proof of compliance for regulators.
- Model card vs datasheet. A datasheet describes a dataset, not a model.
- What does not belong. Source code, raw training records and every incident plan the organization holds.
The tell: if a question asks which document should have listed excluded uses, the answer is the model card.
Frequently asked questions
Is a model card required by the EU AI Act?
Not by that name. But the Act requires similar content. High-risk providers must give deployers instructions for use covering intended purpose, accuracy and limits (Art. 13). GPAI model providers must give downstream providers documentation on capabilities and limitations (Art. 53(1)(b)). Many providers use a model card to deliver part of this.
What should a model card include?
The 2019 format has nine sections: model details, intended use, factors, metrics, evaluation data, training data, quantitative analyses, ethical considerations, and caveats and recommendations. The most useful parts in practice are the out-of-scope uses, performance broken down by group, and known limitations.
What is the difference between a model card and a datasheet?
A model card documents a trained model: its purpose, performance and limits. A datasheet documents a dataset: why it was created, what it contains, how it was collected and labeled, and its limits on use. A good model card links to the datasheets for its training and evaluation data.
Who writes a model card?
The provider or developer that built the model, because only the builder knows how it was trained and tested. Deployers read and challenge it during due diligence and impact assessments, but they do not write it. Much of its content is gathered during training and testing and finalized before release.
Related
Sources
- Mitchell et al., Model Cards for Model Reporting (2019)
- Gebru et al., Datasheets for Datasets
- Regulation (EU) 2024/1689 (EU AI Act), Arts. 11, 13 and 53, EUR-Lex
- EU AI Act Service Desk, Article 53: Obligations for providers of general-purpose AI models
- European Commission, The General-Purpose AI Code of Practice