In plain words
Medicines come with a leaflet: what it treats, the dose, who should not take it, the side effects. AI models now often come with something similar, called a model card.
A model card is a short document written by the people who built the model. It says what the model is meant to be used for, and what it should not be used for. It describes the data it learned from, how well it did in tests, and whether it works equally well for different groups of people.
Model cards help developers choose the right model, and help everyone else spot risks before a model is used for something it was never meant to do. Big AI labs publish longer versions, often called system cards, for each new model.
Everyday example
The label on a packet of food
A packet of namkeen tells us the ingredients, the nutrition per serving, allergy warnings like "contains peanuts", and a best-before date. We do not need to visit the factory to decide whether it suits us. A model card does the same for an AI model: what went in, how it performed, and who should be careful.
Where the example breaks: Food labels follow legal rules and are checked by regulators. Most model cards are written by the model's own makers, follow no single required format, and are rarely checked by anyone independent.
How it actually works
Model cards were proposed in 2018 by researchers at Google, with Margaret Mitchell as first author.
The original idea
The paper "Model Cards for Model Reporting" called them "short documents accompanying trained machine learning models". Its two examples were a model that spots smiling faces in photos and one that flags toxic comments.
Results broken down by group
The key idea was to report performance not just overall but separately for different groups of people, and for combinations of groups. A model that does well on average can still do badly for one group; a card is meant to show that. See AI bias.
Intended use and out-of-scope use
Cards say what the model was built and tested for, and warn against uses it was not designed for, for example a model tested only on English text being used on other languages.
Model cards on Hugging Face
On Hugging Face, every model's page is its model card: a text file with the description, intended uses, limitations, training data and evaluation results, plus tags for licence, languages and the base model it was built from.
System cards from AI labs
Labs such as Anthropic and Google DeepMind publish a card with each major model. Anthropic says its system cards "document the capabilities, safety evaluations, and responsible deployment decisions" for its models, often including red teaming results.
- Proposed
- October 2018, "Model Cards for Model Reporting", Margaret Mitchell and colleagues at Google; presented at the FAT* conference in January 2019
- Typical sections
- Model details, intended use, out-of-scope uses, training data, evaluation results, results by group, limitations
- In law
- The EU AI Act requires makers of general-purpose AI models to keep technical documentation and publish a training data summary, from 2 August 2025
- Related term
- System card: a longer report on a deployed model and its safety testing
Questions people ask
What is the difference between a model card and a system card?
A model card describes one model: purpose, data, tests and limits. A system card, the name several big labs use, usually covers the model as deployed, with its safety tests and safeguards. Anthropic, for example, publishes a system card with each Claude model.
Where can I find a model's model card?
For open models, on the model's page on Hugging Face, where the card is the main text. For models from big labs, look on the company's website for its model or system cards page. Google DeepMind and Anthropic both keep one.
Are model cards required by law?
Not under that name. But the EU AI Act, since 2 August 2025, requires providers of general-purpose AI models to keep technical documentation, share capability and limitation details with businesses building on them, and publish a summary of training content. See AI regulation.
Can I trust what a model card says?
Treat it as the maker's own report. It is very useful, but it is usually not audited, and it shows the tests the maker chose to run. Independent benchmarks and outside testing give a fuller picture.