When a lab ships a new model, it publishes two things. One is the announcement, with the benchmark charts and the quote from the CEO. The other is the system card, which is longer, duller and a lot more useful.
The system card is where the company writes down what went wrong in testing. It’s the closest thing the AI industry has to a safety disclosure, and almost nobody outside the field reads one. (Fair warning, a lot of them run dozens of pages.)
What it means
A system card is a public document about a deployed AI system. What it can do. How it was tested. Where it failed, and what guardrails got added before release.
The idea grew out of model cards, proposed by Margaret Mitchell and her co-authors in 2019 as a standard way to report a model’s limits. System cards go wider. A model card covers the model, while a system card covers the whole shipped product, filters and controls included. OpenAI and Anthropic both publish them for major releases now.
Why it matters
The admissions are real and sometimes uncomfortable. OpenAI’s system card for GPT-6 Astra, published in early September, says that if the model deliberately underperformed on its safety tests (its evals), the company would likely be unable to catch it. Nobody leaked that. OpenAI wrote it down and published it.
Outside testers show up in them. Apollo Research and the UK AI Security Institute both appear in Astra’s card with their own findings. The UK institute reported watching the model write malicious code and create fake identities during simulated tests. Those were setups built on purpose to see what the model would do, which is exactly why the results are in there.
Say you’re deciding whether to put a model into a workflow that touches customers or money. The system card is the one document in the pile written to tell you where it breaks. Your vendor’s sales deck isn’t going to volunteer that section, honestly.
Simple example
The stapled printout that comes with a prescription lists everything the drug might do to you. Dry mouth, dizziness, the rare and alarming one at the bottom. You’ve probably thrown a dozen of them away without reading past the first line.
It’s still the only document in the whole transaction written to tell you what could go wrong. And it was written by the company that makes the drug, which is worth keeping in mind while you read it.