A system card is a comprehensive documentation artifact that describes an AI system's capabilities, limitations, intended uses, safety evaluations, and ethical considerations. It serves as the primary transparency mechanism through which AI developers communicate how their system works, how it was tested, and how it should (and should not) be used.

What a System Card Includes

• Model Overview: Architecture, training data summary, parameter count, training compute, and key design decisions.
• Intended Use Cases: What the system is designed to do, target users, and expected deployment contexts.
• Out-of-Scope Uses: Explicitly listed use cases the system is not designed for and should not be used for.
• Performance Metrics: Benchmark results across relevant tasks, disaggregated by demographic groups where appropriate.
• Safety Evaluations: Results of red-teaming, adversarial testing, bias audits, toxicity evaluations, and jailbreak resistance testing.
• Limitations and Risks: Known failure modes, biases, hallucination rates, and contexts where the system performs poorly.
• Mitigation Strategies: What safety measures were implemented (RLHF, content filtering, guardrails) and their effectiveness.
• Data Practices: High-level description of training data sources, filtering, and any personal data considerations.

Notable Examples

• GPT-4 System Card: Published by OpenAI alongside the technical report, documenting extensive red-teaming and safety evaluations.
• Claude Model Card: Anthropic's documentation of Constitutional AI training and safety characteristics.
• Gemini Technical Report: Google DeepMind's documentation of capabilities and safety testing.

Why System Cards Matter

• Regulatory Compliance: The EU AI Act requires comprehensive documentation for high-risk AI systems.
• User Trust: Transparent documentation helps users make informed decisions about using AI systems.
• Accountability: Creates a public record of known risks and mitigation efforts.

System cards have become a standard industry practice for responsible AI deployment, evolving from the earlier concept of model cards proposed by Mitchell et al. (2019).