model card

basicwidespread

A model card is a short, public data sheet for an AI model: it describes its purpose, training data, performance on tests, and what it should not be used for.

In plain terms

A model card is like a drug package leaflet. You do not get the exact recipe of the active ingredient, but rather what the preparation is for, who it is not recommended for, and what side effects to expect. There is one important difference: the package leaflet is approved by an authority, while the model card is usually not checked by anyone externally.

For a language model, this means: who created it, what kind of data it was trained on, what tasks it was designed for, how it performed in tests, and where it tends to make mistakes. A good card also states what not to use it for.

How does it differ from advertising material? Advertising only lists strengths, while a model card should also discuss weaknesses and risks. This does not make it an independent audit: the developer writes the card themselves.

How it works

Before releasing the model, the developer runs tests: measuring accuracy, typical errors, and whether the model treats different human groups, languages, and situations equally well.
It organizes results into a document with a roughly fixed structure: model data, intended use, nature of training data, measurement results, limitations, and ethical considerations.
The card is placed next to the model, for example on the download page or the company website. On major model-sharing sites, this is the first thing you see about a model.
For new flagship models, large companies often publish a more detailed document, a system card. It includes safety evaluations, such as whether the model can be induced to produce harmful content, or how it behaves when carrying out tasks autonomously as an agent.
When a new version is released, the card must be updated because the old numbers are no longer valid for the new model.

An everyday example

Your company wants to introduce a customer service chatbot and is choosing between two models. The model card of one reveals that it was barely tested in Hungarian and is explicitly not recommended for legal advice. That is worth more than any advertisement.

Who is affected

entrepreneur, executiveBefore deploying an AI tool, the model card tells you whether the model was intended for your task at all and where its weak points are.

developerWhen choosing a model, this is the first filter, and if you publish your own model, you are expected to write one too. → Code

lawyer, citizenOne tangible tool for transparency: this shows what the developer has committed to and acknowledged regarding the model's limitations. → Law

everyoneIf you want to know how much you can trust a chatbot's answers, here are the creator's own admissions about errors.

A common misconception

MythIf a model has a model card, it has been reviewed by independent experts, so it is safe.

In realityThe card is written by the developer about their own model, so it is a self-declaration, not a certificate. It is a useful starting point, but its quality and honesty vary greatly by company.

Limits and risks

There is no unified, mandatory format, making it difficult to place and compare the cards of two models side by side.
Test scores were generated under laboratory conditions; on your own tasks, the model may perform worse.
Model cards mostly describe training data only in broad terms; exact sources are rarely revealed.
The model cards of many freely downloadable models are incomplete, outdated or entirely empty. If you see one, that is a warning sign in itself.

Similar, but not the same

system card – A system card describes not only the model itself but the entire system built around it, including safety filters and evaluations, so it is generally longer and more detailed. Terminology is not uniform: some companies call this more extensive document a model card as well.
dataset card – The datasheet documents the dataset used for training; the model card documents the model built from it.
performance test (benchmark) – The benchmark is the measurement itself and its result, while the model card is the document that shares these results along with other information.

Related concepts

Broader

transparency, AI documentation, responsible AI, AI governance

Narrower

intended use, security assessment

On the edge

system card, dataset card, performance test, training data, bias, language model

A brief history

The term was introduced by Margaret Mitchell, Timnit Gebru, and their co-authors – mostly Google researchers – in their paper 'Model Cards for Model Reporting'. It was published in autumn 2018 and presented at a conference in early 2019. Since then, the genre has become standard practice on model-sharing platforms and at major AI companies.

Where things stand

By the end of 2026, major developers typically release new models with model or system cards, although the card sometimes appears late. This is also the default documentation on model-sharing platforms. The EU AI Act has required technical documentation for general-purpose model providers since August 2025. However, this is not the same as a public model card: it is intended for authorities and companies integrating the model, and open-source models are partially exempt. The detail level of public cards therefore remains very uneven.

Status: 2026 November

Also known as

model card · model card · model card · model card · model cards · modelladatlap · system card · system card

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.