Explore our new Cybersecurity Terms Glossary, explained by our experts.
Cybersecurity Terms Glossary
A model card is a short structured document that states what an AI model is for, how it was built and evaluated, where it should not be used, and what is known to go wrong.

What Is a Model Card?
A model card is the datasheet an AI model should arrive with. Margaret Mitchell and co-authors proposed the format in 2019, borrowing the idea from the datasheets that accompany electronic components: a short, structured document covering intended use, training data, evaluation results, limitations, and ethical considerations. The proposal addressed a real gap—models were being shared, downloaded, and embedded in products with no statement of what they were built to do or where they fail.
Hugging Face made the format mainstream by attaching cards to repository pages, and providers of frontier models publish longer system cards covering capability and safety evaluations for the whole system rather than the weights alone.
Cards answer questions that otherwise require guesswork. Does this model carry a licence that permits commercial use. Was it fine-tuned, and from what base, which determines whether a disclosure affecting the base model affects you. What evaluation was run, and does it include any adversarial testing. Which subgroup did it perform worst on, which matters if the model informs decisions about people. A model that arrives with no card is not necessarily unsafe, but it is undocumented, and documenting it becomes your job before it goes anywhere near production. This is the same discipline described in our note on AI supply chain security, where provenance is the control that makes everything else possible.
The format is only as good as the candour behind it. Common weaknesses are easy to spot once you look for them: an intended-use section so broad it excludes nothing, evaluation on public benchmarks that resemble nothing in your domain, no subgroup breakdown, no mention of adversarial testing, and a training data section that says “publicly available sources”. Cards also go stale, because models get fine-tuned, quantised, and re-released while the documentation stays where it was. Treat the card as a claim to verify on your own data rather than as assurance, and write your own card for any model you fine-tune, recording what you changed and what you re-evaluated.
ISO/IEC 42001 expects documented AI system information as part of a management system, and the NIST AI Risk Management Framework places transparency documentation across its Map and Measure functions. The EU AI Act turns this into law: providers of high-risk systems must maintain technical documentation to a prescribed content list, and providers of general-purpose models must publish documentation for downstream deployers plus a summary of training content. In the UK, the Department for Science, Innovation and Technology’s AI assurance work points in the same direction through its portfolio of techniques, and public sector buyers increasingly ask for model documentation in procurement. For a deployer, the practical value is simpler than any of this: a card gives you the evidence to answer a regulator, a client questionnaire, or an incident review without reconstructing history from commit logs. Pair it with the records described in AI auditing and traceability. At Privia Security, our Regulation and Compliance Services help teams produce AI documentation that stands up to review.
They overlap but differ in scope. A model card documents a model: its training, evaluation, and limits. A system card documents a deployed system, including the model, the guardrails, the retrieval sources, the tools it can call, and the human oversight around it. For most organisations the system card is the more useful artefact, because that is what actually faces users.
You need documentation for your deployment, which can incorporate the provider’s card by reference. Record the model and version in use, the configuration, what you added through fine-tuning or retrieval, the evaluations you ran on your own data, and the limits you place on its use. The provider documents the model; you document the system.
The model and version, training data provenance at least in summary, evaluation results including subgroup breakdowns, adversarial and safety testing performed, known failure modes, licence terms, update and deprecation policy, and a security contact. If a vendor cannot supply these, treat that as a finding in your assessment rather than an administrative gap.
On every material change: a new version, further fine-tuning, quantisation, a change of training data, or a newly discovered failure mode. Tie the card to the model registry entry so it versions with the artefact. A card that describes a model two releases old is worse than none, because it gives false confidence to anyone relying on it.
You May Be Interested In These
We use a few cookies that are needed to make this website work, and they are always on. We would also like to set analytics cookies, to see which pages are useful; functional cookies, so that our live chat works; and advertising cookies, to measure our campaigns. We will not set any of those until you say yes, and you can change your mind at any time from the Cookie Settings link in the footer.
Choose the cookies you are happy for us to use. Strictly necessary cookies are always on; everything else is off until you turn it on.
These cookies keep the website working: they hold your security checks, your form submissions and the choice you make here. The site cannot run without them, so they are always on.
These let our live chat open and remember your conversation, and let the newsletter sign-up work. Switch them off and the chat window will not appear.
These tell us which pages are read and how people find us, so that we can improve the site. The reports are about groups of visitors, never about you personally.
These let us measure our advertising and show you our adverts on other websites, such as LinkedIn, Google and Meta. Switch them off and you will still see adverts, but they will be less relevant.