What is a model card, and what should an AI launch tell you?
A model card is a short public document released with an AI model, like a nutrition label. It says what the model is meant for and not meant for, how it was trained, its test results, known limitations and the safety checks it passed. A responsible launch adds staged rollout, monitoring and a way to report problems.
▶ Try it
Why it matters
The last station on the line is the launch. A good model card lets anyone, from developers to journalists to you, see what a model can do, where it fails and how it was checked.
In 6 minutes you'll be able to
- Look for intended uses, known limits, named test results and safety findings.
- Vague claims like “excellent results” or “no limitations” are red flags.
- Responsible launches check data, scores and attacks before release, and keep watching after.
What you'll practise
What goes on the card?
Pick everything a good model card should include.
- What the model is meant for, and not meant for
- Known limitations and weak spots
- Test results, with the names of the tests
- Safety testing and red-team findings
- What kinds of data it was trained on
- A promise that it never makes mistakes
- Every user's private chats with it
Spot the red flags
Quickstep AI, an invented start-up, rushed out its model card. Tap the parts that should worry you.
Nimbus-1 · model card
Model: Nimbus-1, a chat model.
Size: 3 billion parameters.
Intended use: Any task, including medical and legal advice.
Training data: Data from the internet.
Languages: English and Spanish.
Evaluation: Tested on lots of benchmarks. Results: excellent.
Known limitations: None.
Licence: Research use only.Run Vireo's release line
Tap a release check, then the stage where it can catch its problem. Then run the release and watch five changes head for the public.
collect → test → attack → launch → public
Release checks
- Data check → collect
The data check belongs at collect: private or unwanted data has to come out before training, not after. - Benchmark suite → test
Benchmarks run at test, once there's a trained model to score and compare with the last version. - Red-team review → attack
Red teamers do their work at attack: tricky prompts against the finished model, before anyone else can try. - Model card review → launch
The card is checked at launch: every promise on the launch page must match what the tests actually showed.
Release log
- Ines: Add a new batch of web text
Caught at collect by Data check: a leaked customer list hidden in the batch. Fix: Remove it and rebuild the training set. - Ravi: Speed-up tweak to the model
Caught at test by Benchmark suite: maths scores dropped 9 points. Fix: Roll back the tweak and re-test. - Jun: Friendlier personality
Caught at attack by Red-team review: role-play now unlocks harmful answers. Fix: Add safety training for role-play and re-test. - Elif: Fix a typo in the welcome text
Released safely - Noah: Launch page: “safe for medical diagnosis”
Caught at launch by Model card review: a promise no test supports. Fix: List medical advice under “not intended for”.
- Data check → collect
Remember
A model card says what a model is for, where it fails and how it was tested. No card, or a vague one, is a warning sign.
Key facts
- Model cards were proposed in 2019 as short documents that report a model's intended use, performance across groups and limitations. · Source: Mitchell et al., Model Cards for Model Reporting, FAT* 2019
- A sister idea, datasheets for datasets, asks dataset creators to document where the data came from, how it was collected and what it should be used for. · Source: Gebru et al., Datasheets for Datasets, Communications of the ACM, 2021
- In 2019, OpenAI released its GPT-2 model in stages, starting with smaller versions, to study misuse before releasing the largest one. · Source: Solaiman et al., Release Strategies and the Social Impacts of Language Models, OpenAI, 2019
- Under the EU AI Act, providers of general-purpose AI models must keep technical documentation and give developers who build on them information about the model's capabilities and limitations. · Source: EU Artificial Intelligence Act, Regulation (EU) 2024/1689, Article 53
Updated:
Questions people ask
What is a model card in AI?
A model card is a short document published with an AI model. It describes what the model is for, how it was trained and tested, how well it performs and where it is known to fail.
What is the difference between a model card and a system card?
A model card describes one model. A system card, used by some labs, covers a whole product built around models, including safety testing, filters and how it is deployed.
Why are AI models released in stages?
Releasing to a small group first lets a lab watch how the model is used and fix problems before millions of people rely on it.
Where can I find a model's model card?
Usually on the lab's website next to the launch announcement, in its technical report, or on the page where the model can be downloaded.
Related terms