Definition
System card is a system card is a document an AI lab publishes with a model release describing what it can do, how it was tested, and the safety risks and limits found before launch.
A system card is a report that accompanies an AI model or product. It is the closest thing the industry has to a nutrition label. A good one tells you what the model is for, how it performed in testing, what risks the lab looked for, and where the model still falls short.
The name comes from "model cards," a documentation idea from machine learning research. System cards go further by covering the whole deployed system, including safeguards, rather than only the model's weights.
Contents vary by lab, but a system card commonly covers:
If you build on a model, the system card is where a lab tells you what it did and did not check. You can use it to decide where to add your own tests, what to restrict in production, and which claims to repeat to your own customers.
Reading it well takes some skepticism. A system card is written by the lab that sells the model, so it shows what the lab chose to measure and report. Independent evaluations by outside auditors are a useful cross-check, because they test claims the lab cannot grade for itself. Notice also what is missing: a model released with no safety documentation tells you nothing about its testing, not that none happened.
Before putting a new model in a customer support product, you open its system card, find the section on harmful content and prompt-based attacks, and note the weak spots. Then you write your own tests for those exact cases instead of assuming the lab covered your use.
System cards describe the results of work in AI alignment and often cite reasoning behavior such as a chain of thought. They differ from a benchmark, which measures one skill, because a system card covers risk and limits.
A system card is a document a lab publishes with a model, describing its capabilities, safety testing, known risks and limits.
They are related. A model card documents a model, while a system card also covers the deployed system around it, including safeguards.
SemiAnalysis checked 857 model releases from nine Chinese developers and found a published safety result for 31 of them, and only 9 at launch. If you run Chinese open weights, the safety testing is your job.
OpenAI says Tomek Korbak, Mikita Balesni and Jasmine Wang broke its rules on handling sensitive information. Their letter says the firings chill safety work, and the real fight is over outside auditors' access.