System Card
A system card is an accompanying document that a company publishes for a new AI product. It states what the system is intended for, how well it performed in tests, and where it typically makes mistakes.
When a company releases a new computer program that writes texts or recognizes images on its own, it often provides an accompanying document. This document is called a system card. It states what the program is intended for and what it is explicitly not intended for. The developers also describe which tests they carried out beforehand and how the program performed in them. An important part concerns the known weaknesses: tasks where the program regularly gives wrong answers. You can think of a system card like the package leaflet of a medication — effect, area of application, and side effects all on one sheet.
What a system card reveals about trust
Modern AI systems are not transparent from the outside. You only see what goes in and what comes out. Anyone who wants to deploy such a system in a school, a bank, or a hospital needs reliable information, however. Without a system card, all that remains is the manufacturer’s marketing, and that rarely names the limits of the product.
For journalists and regulators, system cards are often the only reliable source. When the news reports that a model achieved 80 percent on a certain test, that figure usually comes from a system card. New laws are also heading in this direction. The European Union’s AI Act requires technical documentation for high-risk applications that looks very similar in content.
There is a catch, though. System cards are almost always written by the manufacturer itself, without independent oversight. The manufacturer decides which tests are included. A test that turned out poorly can simply be left out. So a system card is a starting point for transparency, not a seal of approval.
What such a document consists of
At the beginning comes the description of the system: which inputs it processes, in which languages it operates, in which products it is embedded. Then follow the measurement results. For this, the system is made to solve standardized sets of tasks, such as math problems or programming tasks with known solutions. The success rate is given as a percentage, allowing comparison with other systems.
A second part deals with risks. Here, companies commission teams that deliberately try to coax harmful answers out of the system. This approach is called red teaming, borrowed from military exercises with an opposing force. The system card documents what these teams found and which safeguards were subsequently built in.
The system card is often confused with the model card. A model card describes only the trained model itself, i.e., the computational core. A system card describes the entire product surrounding it: filters, safety rules, user interface. This is the more practically relevant view, because users always interact with the overall system.
System cards in product launches and headlines
Major providers such as OpenAI, Anthropic, and Google publish a system card for every important model, often on the day of its announcement. The documents are freely available online and are usually between twenty and over a hundred pages long. Anyone who wants to know what a new model can really do will find more there than in the press release.
System cards often appear in news articles when they contain something unpleasant. Reports about a model’s dangerous capabilities frequently stem from the risk sections of these documents. Such details regularly make headlines, even though the manufacturer disclosed them itself.
System cards also play a growing role in professional life. Companies purchasing AI now demand them as part of the bidding documents. If the document is missing, it becomes hard to justify why one trusts the system. For pupils and students, it is a good starting point: you learn from a real-world example how a technology describes its own limits.