System cards and model cards: what to read before you adopt a model
Vendors publish system cards and model cards with every major model. Here is which sections matter to a business buyer, and a short checklist for reading them.
When a lab ships a new model it usually publishes a document describing it. These go by two names: a system card (broader, covers safety testing of the model as deployed) and a model card (shorter, a structured fact sheet). Google describes its model cards as "simple, structured overviews of how an advanced AI model was designed and evaluated" [4]. The documents are written by the vendor, so they are evidence, not an audit. They are still the cheapest due diligence available.
What the real ones contain
The three examples below are openly available and show how much the format varies.
OpenAI's GPT-5 system card (August 13, 2025) has sections on model data and training, safety evaluations (disallowed content, jailbreaks, prompt injection, hallucinations), red teaming and external assessments, and a Preparedness Framework review [1].
Anthropic's Claude Opus 4 and Sonnet 4 system card (May 2025) covers training data, safeguards results, agentic safety including prompt injection against computer use, and an alignment assessment, and states the safety level each model was deployed under [2].
Meta's Llama 3.1 model card is shorter: model information, intended use, hardware and software, training data, benchmark scores, responsibility and safety, and ethical considerations and limitations [3].
The sections that matter to a buyer
Intended use and out-of-scope use. Meta's card states Llama 3.1 is "intended for commercial and research use in multiple languages" and lists uses outside scope [3]. If your use case is not in the intended list, ask why.
Evaluations. Look for the evaluation method, not just the score. OpenAI's card reports hallucination rates measured with an LLM-based grader that agreed with human reviewers 75% of the time [1]. That is useful context for how far to trust the headline figure.
Limitations and known failure modes. These are often the most decision-relevant paragraphs.
Safety testing. Check for prompt-injection results if you plan to give the model tools [1][2].
Data handling. Anthropic's card says its models were trained partly on data from users "who have opted in" [2]. OpenAI's says training data includes information users or trainers provide or generate [1]. The card describes training, not how your own data is handled, so read the vendor's data-retention documentation as well.
A reading checklist
Question | Where to look | Red flag |
|---|---|---|
Is my use case in scope? | Intended use | Silent on your domain |
How was it measured? | Evaluation sections | Scores with no method or comparison |
What goes wrong? | Limitations | Section missing or generic |
Can it be attacked via inputs? | Prompt-injection and jailbreak tests | No agentic or tool-use testing |
Whose data trained it? | Training data | No statement on user data |
Does the card match the model I will call? | Model name and date | Card covers a different version |
Last, record the card's version and date next to the model you adopt. Models change, and a card written for one snapshot may not describe the next.
Sources
- GPT-5 System Card OpenAI, 2025-08-13. cdn.openai.com/gpt-5-system-card.pdf
- System Card: Claude Opus 4 & Claude Sonnet 4 Anthropic, 2025-05-22. www-cdn.anthropic.com/07b2a3f9902ee19fe39a36ca638e5ae987bc64
- Llama 3.1 Model Card Meta (llama-models repository), 2024-07-23. github.com/meta-llama/llama-models/blob/main/models/llama3_1
- Model cards Google DeepMind, 2025-11-21. deepmind.google/models/model-cards
.jpg){kind=link}