Model Card — Standardized AI Model Documentation for Transparency

Definition

A model card is a structured documentation artifact that provides concise, standardized information about a machine learning model: its intended use, performance characteristics, evaluation results across demographic groups, training data summary, limitations, and ethical considerations. Introduced by Google researchers in 2018, model cards have become the de facto standard format for communicating AI model capabilities and limitations to downstream users, deployers, regulators, and the public.

A standard model card includes: model description and intended use cases; out-of-scope uses (how the model should not be used); factors that affect model performance; evaluation metrics and results, including disaggregated performance across demographic subgroups; training data summary; ethical considerations and caveats; limitations and known failure modes; and recommendations for safe deployment.

Under the EU AI Act, GPAI model providers must publish technical documentation about their models including capabilities, limitations, and appropriate use cases — requirements that model cards directly satisfy. For high-risk AI systems, the technical file under Annex IV includes many of the same elements that model cards document, making model cards a valuable component of the broader documentation infrastructure.

Why it matters operationally

Model cards matter because they are the primary mechanism through which AI model transparency is operationalized for downstream audiences. A deployer considering whether to use a model for a specific application needs to know: what the model was trained to do, how it performs on the specific demographic context of the deployment, what its known failure modes are, and what uses are explicitly not recommended. Without a model card, deployers must either conduct their own evaluation (costly) or deploy without adequate information (risky).

For organizations developing AI models, publishing model cards signals transparency and responsible practice. For organizations deploying third-party models, requiring model cards from vendors is a due diligence practice that surfaces the information needed to make informed deployment decisions and satisfy EU AI Act deployer obligations.

Regulatory framework

Framework	Model card requirements
EU AI Act — GPAI	GPAI model providers must supply technical documentation including capabilities, limitations, and appropriate use contexts. Model cards directly satisfy this requirement.
EU AI Act — Annex IV	Technical file elements for high-risk systems overlap significantly with model card content.
ISO/IEC 42001	Annex A controls on AI system documentation include model cards as a model lifecycle documentation format.
NIST AI RMF	The NIST AI RMF Playbook references model cards as a transparency practice in the Govern function.

How Zertia evaluates it

Zertia evaluates model card adequacy as part of the AI Model Audit. The audit assesses whether model cards exist for AI systems in scope, whether their content is accurate and current, whether they cover disaggregated performance metrics relevant to the deployment context, and whether they meet the transparency documentation requirements of the EU AI Act for GPAI providers and the Annex IV technical file requirements for high-risk system providers.

[AI Model Audit] · ISO 42001 Certification