System card

In short

A system card is a report an AI company publishes when it releases a model, describing what the model can do, the safety tests it went through, the risks found and the safeguards put around it before launch.

Also called: System cards, Model system card
Topic: Safety, ethics and risk
Level: Beginner
First described: Proposed by Meta AI researchers in "System-Level Transparency of Machine Learning", February 2022¹.

System cards grew out of model cards, which document a single model. A system card looks at the whole system: the model plus the rules, access controls and abuse monitoring wrapped around it in a product².

Meta researchers proposed the name in February 2022. Their paper said earlier documentation such as model cards had started the work, and that system cards would show how the different parts of an AI system interact and how data is used¹.

OpenAI published a system card for GPT-4 in 2023. It covered risks such as convincing but false text and help with harmful requests, compared an early version of the model with the safer launch version, and said its mitigations were "limited and remain brittle in some cases"².

Major labs now publish them with new models, and they can be very long. Anthropic's card for Claude Opus 4 and Sonnet 4 in May 2025 ran to 123 pages and covered tests under its Responsible Scaling Policy, reward hacking, agentic safety and an alignment assessment³. Anthropic lists system cards for each Claude release on one page⁴.

System cards are written by the companies themselves. They show what a developer chose to test and disclose, not an independent audit.

Example

The Claude Opus 4 system card reported that, in a made-up test where the model was about to be replaced and had found emails showing the engineer was having an affair, it tried blackmail in 84% of runs, in a set-up designed to leave it no other way to avoid replacement³.

Why it matters

System cards are often the most detailed public account of how risky a new AI model is. Many of the headlines about AI models behaving badly in tests come from them.

Questions people ask

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

A model card describes one trained model. A system card covers the model plus the policies, access controls and monitoring around it, as OpenAI's GPT-4 system card puts it². In practice the names overlap: Anthropic lists its earlier Claude model cards on its system cards page⁴.

Who checks what goes in a system card?

The company writes it. Some cards include results from outside testers: Apollo Research, for example, tested OpenAI's o1 before release and its findings appeared in the o1 system card⁵. But the card still reflects what the developer chose to measure and publish³.

Where can I read AI system cards?

On the developers' websites. Anthropic publishes a list with a card for each Claude model⁴, and OpenAI publishes cards such as the GPT-4 system card².

Related terms

Tools that use it

Sources

Back to the glossary