System Card
How it works
A System Card is a structured report, usually published when a model or system is released. It typically contains: (1) an overview of the model and system, with deployment context; (2) capabilities and limitations; (3) safety evaluations across risk categories, often within formal processes (e.g., OpenAI's Preparedness Framework or Anthropic's Responsible Scaling Policy); (4) external red-teaming and expert engagement (the GPT-4 System Card used more than 50 experts); (5) a description of identified risks; (6) mitigations at the model level (e.g., safety training, RLHF) and at the product and system level (policies, access controls, monitoring); (7) residual risks and the deployment decision. The document is descriptive — it does not modify the system but documents its behaviour and safety-assurance process.
Problem solved
A Model Card documents the characteristics of a single model, but a deployed AI system also includes non-model components — usage policies, access controls, moderation, tools and monitoring — whose behaviour and risks a model card alone does not capture. The System Card addresses the need for transparency and accountability at the whole-system level, including mitigations and residual risks that only emerge after deployment.
Components
The scope of the document and a description of the model and the system around it (non-model components, deployment context).
What the system can and cannot do, including dual-use capabilities.
Results of safety testing across risk categories, often within a Preparedness Framework or Responsible Scaling Policy.
Engagement of external experts to adversarially test the system before deployment.
Identified risks: bio, cyber, persuasion, privacy, disinformation and others.
Mitigations at the model level (safety training, RLHF) and at the system level (policies, access controls, monitoring).
Risks remaining after mitigations and the rationale for the deployment decision.
Implementation
System cards are voluntary and published by the provider itself; the scope is sometimes deliberately narrowed (e.g., the GPT-4 System Card explicitly excluded image capabilities and fine-tuning), which limits completeness.
There is no single standard; different providers report different risk categories and metrics, which hampers comparability across systems.
Evolution
Introduces the model-reporting card — the foundation that the System Card extends to the system level.
Popularises the term "system card"; a system-level analysis with non-model mitigations and red-teaming (more than 50 experts).
A system card for a multimodal model; evaluations under the Preparedness Framework and third-party dangerous-capability assessments.
Safety evaluations of a reasoning model, external red-teaming, the Preparedness Framework and deliberative alignment.
Anthropic publishes system cards for its Claude models alongside Responsible Scaling Policy evaluations and red-teaming.