Model card
A fact sheet for an AI model that tells you what it was built for, how well it works for different groups of people, and what it should not be used for.
The Simple Version
A fact sheet for an AI model that tells you what it was built for, how well it works for different groups of people, and what it should not be used for.
Detailed Explanation
The model card concept was introduced by Mitchell et al. (2019) at Google and has since become an industry standard for model transparency. A model card typically covers: model description (architecture, intended use, out-of-scope uses), training data summary, evaluation metrics disaggregated by demographic subgroup, known limitations and failure modes, and ethical considerations. Model cards do not satisfy all of EU AI Act Annex IV requirements but cover a significant portion of the required content and are widely used in the AI industry as a baseline transparency mechanism. Some organisations augment model cards with 'data cards' for training datasets.
Key Characteristics
- Provides concise, structured transparency about a specific model
- Includes performance metrics disaggregated by subgroup to surface bias
- Widely adopted in industry as a transparency baseline
- Covers a substantial portion of EU AI Act Annex IV documentation requirements
Why It Matters
Model cards are an efficient transparency mechanism for AI product teams, they communicate essential governance information to deployers, procurement teams, and auditors without exposing proprietary model internals.
Real-World Analogy
Like a medication package insert, a standardised document that informs prescribers and patients about intended use, dosage, contraindications, and known side effects, without revealing the drug's proprietary synthesis process.
Common Misconceptions
- A model card fully satisfies EU AI Act technical documentation requirements. Annex IV requires more detailed development methodology, data governance, and risk management documentation than a standard model card.
- Model cards are only relevant for open-source models, proprietary model cards are equally valuable for communicating capabilities and limitations to enterprise deployers.