Model card

In short

A model card is a short document published alongside an AI model that explains what it is for, how it was trained and tested, how well it performs for different groups of people, and where it should not be used.

Also called: Model cards, Model documentation
Topic: Safety, ethics and risk
Level: Beginner
First described: Mitchell, Wu, Zaldivar, Barnes, Vasserman, Hutchinson, Spitzer, Raji and Gebru, "Model Cards for Model Reporting", posted October 2018 and published at the FAT* conference in 2019¹.

Think of it as a nutrition label for an AI model. The idea came from a 2018 paper by Margaret Mitchell, Timnit Gebru and colleagues, who argued that released models should come with documentation of their intended uses and their performance in different conditions¹.

Their proposal stressed testing results broken down by group, such as by age, sex or skin type, so that people could see if a model worked worse for some groups than others¹. The paper included sample cards for a smiling-face detector and a toxic-comment classifier¹.

Model cards are now standard practice. On Hugging Face, a popular public hub for AI models, the model card is the README file in every model's page. It should cover intended uses and limitations, training details, the datasets used and evaluation results². Google DeepMind publishes model cards for its Gemini and Gemma models³.

A model card describes one model. A system card goes wider, covering the model plus the safety measures and policies around it in a product⁴.

Some of this is now required by law. The EU AI Act obliges providers of general-purpose AI models to keep technical documentation covering training, testing and evaluation results, and to give information on capabilities and limitations to companies building on the model⁵.

Example

Google's model card for its open Gemma 4 model has sections on training data, benchmark results, safety evaluations, intended uses and known limitations⁶.

Why it matters

A model card is often the only place a company says plainly what its model should not be used for. Checking it before relying on a model can save you from using it in the wrong setting.

Questions people ask

What is the difference between a model card and a system card?

A model card documents a single trained model. A system card documents the whole system around it, including policies, access controls and monitoring, as OpenAI's GPT-4 system card explains⁴.

Are model cards required by law?

Not under that name. But the EU AI Act requires general-purpose AI model providers to keep technical documentation and share capability and limitation information with downstream developers, which covers much of the same ground⁵.

Where can I find a model card?

On Hugging Face, every model's page shows its card, written as a README file². Large labs publish theirs on their own websites, such as Google DeepMind's model card pages³.

Related terms

Tools that use it

Gemma 4Google's open-weight model family in four sizes for reasoning and agents on local hardware

Sources

Back to the glossary