What is a “system card”?
A system card is a structured public document that describes an AI system’s purpose, capabilities, limitations, training data summary, evaluation results, safety measures and known risks. It extends the earlier “model card” concept (Mitchell et al., 2019) from individual model artifacts to deployed systems that integrate models with tools, guardrails and downstream applications. OpenAI, Anthropic, Google and Meta publish system cards for their flagship releases as a transparency practice.
Typical system card sections
- System overview: purpose, intended use cases, out-of-scope uses.
- Architecture summary: model components, retrieval, tool use, guardrails.
- Training data: source categories, time cutoff, opt-out handling.
- Evaluation: benchmarks, internal safety evals, third-party red-team results.
- Known limitations: hallucination patterns, language coverage, domain gaps.
- Safety mitigations: content filters, output validation, rate limits.
- Misuse risks: documented misuse vectors and mitigations.
- Versioning and change log.
System card vs. related artifacts
- System card vs. model card: model card describes a single model; system card describes the system around the model.
- System card vs. AI Act technical documentation: system card is voluntary transparency; AI Act tech doc is mandatory for high-risk systems.
- System card vs. data sheet: data sheet describes data alone; system card describes the deployed system comprehensive.
Why deploy system cards
- Regulatory readiness: EU AI Act and GPAI obligations require similar information; system cards bridge to compliance.
- Customer trust: enterprise buyers increasingly require transparency documentation.
- Incident response: public system cards speed root cause analysis when issues emerge.
- Internal alignment: drafting the card forces clear documentation of design decisions.
Do: publish a system card for every customer-facing AI deployment; version it with each material change.
Don’t: treat the system card as marketing material — honest documentation of limitations builds more trust than over-claiming.