MODEL CARDS
Nine short sections
The 2019 paper that proposed model cards suggests sections from intended use and evaluation data to caveats and recommendations, so readers can see where a model fits.
Check a releaseWHAT A RELEASE SHOULD TELL YOU
Read the card,
not the launch.
Every new AI model arrives with an announcement.
Check a release
LOOK FORthe limitations
- Intended use first
- Tested on what?
- Known limits, stated
- Three questions for any release
The useful part is usually quieter: the documentation that says what the model is for, how it was tested and where it falls short. Researchers call this a model card, and knowing its sections makes any release easier to read.
Updated
THE SHORT ANSWER
A good release
shows its working.
Documentation turns a launch into something you can check.
Three places to look.
EU RULES
Documentation by law
Since 2 August 2025, providers of general-purpose AI models in the EU must keep technical documentation, have a copyright policy and publish a summary of training content.
What the rules ask
TRANSPARENCY
Scores went down
An independent index scores major model developers on 100 indicators. The average rose to 58 in 2024, then fell to 40 in 2025.
See the numbersA QUICK RELEASE CHECK
What came with
the model?
Pick what a release came with to see how much you can check for yourself. The bar shows how much documented detail that item usually gives a reader.
Based on the model cards paper (Mitchell et al., 2019), NCSC secure AI guidelines and the European Commission's AI Act Q&A. General information, not legal advice.
TWO NUMBERS WORTH KNOWING
Openness rose,
then slipped back.
One figure on how much developers disclose,
one on where EU rules draw the line for the largest models.
40/100
Average developer score out of 100
The Foundation Model Transparency Index averaged 37 in October 2023 and 58 in May 2024, then fell to 40 in 2025. Training data, training compute and real-world use were the least disclosed areas.
TRANSPARENCY INDEX · 2025
10²⁵ FLOP
Training compute that triggers extra duties
Models trained with more than this amount of computation are presumed to carry systemic risk and face extra duties, including model evaluations and serious incident reporting. The Commission says such training currently costs tens of millions of euros.
EU AI ACT · SYSTEMIC RISK
The transparency index found that information about data, such as copyright, licences and personal information, stayed among the most opaque areas across editions.
FOUR QUESTIONS TO ASK
Before you rely on
a new model.
None of these needs technical knowledge. They are the same questions the documentation frameworks were written to answer.
What is it meant for?
Look for the primary intended uses and users, and anything the developer says is out of scope.
What was it tested on?
Check which evaluation data and metrics were used, and whether results are broken down by group or condition.
What are the known limits?
The NCSC asks developers to be clear about known limitations and potential failure modes. If none are listed, that is worth noticing.
Which version is this?
Note the version or date. A changed model can behave differently, so results from an older version may not carry over.
WHAT EU RULES ASK
Open weights,
different duties.
Under the EU AI Act, every provider of a general-purpose AI model must publish a training content summary and have a copyright policy. Most must also keep technical documentation for authorities and for businesses that build on the model.
The Commission says some documentation duties do not apply when a model is released under a free and open-source licence with its weights published, but this exception does not cover models with systemic risk.
Questions
Questions people ask
What is a model card?
A short document that accompanies a trained machine learning model. It was proposed in a 2019 paper to set out intended uses, how the model was evaluated, and its caveats, so it is not used where it is poorly suited.
Where can I find one?
On one large public model hub, the model card is the README file shown on each model's page, with metadata at the top such as the licence. Developers also publish them on their own sites.
Does every AI release come with documentation?
Not always in public. An independent index of major developers found the average transparency score fell from 58 out of 100 in 2024 to 40 in 2025.
What does EU law require since August 2025?
From 2 August 2025, providers of general-purpose AI models must draw up technical documentation, put in place a copyright policy and publish a summary of training content. Models with systemic risk carry extra duties.
Does an open-weights release need less documentation?
Some documentation duties do not apply to models released under a free and open-source licence with weights published, according to the Commission. The training content summary and copyright policy still apply, and the exception does not cover models with systemic risk.
Where the facts come from
- Mitchell et al. (2019): Model Cards for Model Reporting, arXiv
- Hugging Face Hub documentation: Model Cards
- Wan et al. (2025): The 2025 Foundation Model Transparency Index, arXiv
- Stanford CRFM: Foundation Model Transparency Index, May 2024
- European Commission: General-purpose AI models in the AI Act, questions and answers
- NCSC: Guidelines for secure AI system development