Examples
Worked examples
- Is an instance
An LLM provider's system card disclosing the moderation classifier, output filters, and pre-deployment red-team findings.
- Is an instance
A system card for an enterprise copilot product describing the orchestration of retrieval, tool-use, and answer-generation steps.
Counter-examples
Looks similar, but isn't
- Not an instance
A model-only card for a publicly released checkpoint.
- Not an instance
Marketing material describing a product's features.
Editorial commentary
A system card documents the full deployed sociotechnical system built around one or more underlying models — not the model’s weights and training data (that is a model card‘s job), but everything wrapped around it in production: prompt routing between models, output moderation and safety classifiers, retrieval augmentation, tool/function-calling scope, rate limits, and user-facing controls. It answers “what happens when I actually use this product,” where a model card answers “what is this trained artefact.”
System cards were popularised by frontier-lab releases — OpenAI’s GPT-4 System Card (2023) and successors, Anthropic’s Claude model/system cards, Google’s Gemini model cards — as the standard companion document to a major model launch, typically covering red-teaming results, dangerous-capability evaluations, and deployment-specific mitigations that a model-level document alone cannot capture. In this site’s four-tier documentation hierarchy (algorithm card to model card to use card to system card), the system card is the broadest and most product-facing of the four: an algorithm card exists before training; a model card describes one trained artefact; a use card describes one deployer’s specific application of that artefact; a system card describes the assembled product a user actually interacts with, which may combine several models, classifiers and guardrails at once.
What it is not
A system card is not a substitute for a model card — it typically assumes the reader already knows, or can look up, the underlying model’s training details, and instead focuses on deployment-time behaviour and safety testing. Where a lab publishes only one combined document at model launch, that document is functioning as both, but the two questions it answers remain distinct.
References
- OpenAI, ‘GPT-4 System Card’ (2023)
- Procope et al., ‘System-level transparency of machine learning’ (Meta AI research note, 2022)
Also known as
AI system card
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="System card"
vocab-term-identifier="https://casrai.org/dictionary/term/system-card" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/system-card",
"name": "System card",
"identifier": "https://casrai.org/dictionary/term/system-card",
"description": "A documentation artefact describing an AI-enabled system in its production configuration, including the constituent models, the pre- and post-processing pipeline, safety filters, monitoring, and operational guardrails.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/ai-ml-research-outputs#set",
"url": "https://casrai.org/dictionary/term/system-card",
"sameAs": [
"AI system card"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T02:22:50",
"dateModified": "2026-08-22T15:54:48",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
