AI Model Card

Purpose

A Model Card is a concise, structured reference document that describes an AI model's intended use, performance characteristics, known limitations, and ethical considerations. It is designed to be scannable — a governance committee member, auditor, or downstream user should be able to understand the model's capabilities and constraints in under five minutes.

Where the AI Impact Assessment evaluates potential harms and the Risk Register tracks ongoing risks, the Model Card captures a point-in-time snapshot of what the model is, what it does well, where it falls short, and what the organization knows about its behavior across different populations and conditions. It is updated at each significant model version change.

Model Cards were originally introduced by Google's model cards paper (Mitchell et al., 2019) and have since become a governance standard adopted by Hugging Face, Meta, and regulatory bodies including the EU AI Act's technical documentation requirements. This template adapts the format for operational governance use — structured for a PM or governance professional to complete, not a machine learning researcher.

Where It Fits in Your Document Pack

The Model Card draws from your TEVV Plan (performance metrics, test methodology), your Bias & Fairness Assessment (fairness evaluation results across demographic groups), your Risk Register (known limitations and risks), and your Human Oversight Plan (recommended use conditions and human review requirements). It feeds into your Monitoring Plan by establishing the performance baselines that post-deployment monitoring tracks against.

What This Template Covers

• Model identification: model name, version, type, development date, and linked governance documents
• Intended use: primary use case, intended users, deployment environments in scope, and explicit out-of-scope uses
• Model architecture summary: model type, training data description, training approach, and infrastructure context — at a level appropriate for governance review, not technical deep-dive
• Performance metrics: key accuracy, precision/recall, and reliability metrics with the test conditions under which they were measured — not aspirational targets, actual results
• Fairness evaluation summary: results across demographic subgroups for in-scope protected characteristics, with pass/fail status against defined thresholds
• Known limitations: conditions under which the model performs poorly, out-of-distribution inputs, edge cases, and known failure modes identified in testing
• Ethical considerations: data provenance, consent and privacy considerations, potential for misuse, and any unresolved ethical questions
• Recommended use conditions: human oversight requirements, minimum review thresholds, deployment environment requirements, and conditions that should trigger model review or suspension
• Version history: tracks significant model updates with performance change notes and the governance review conducted at each version
• Completion guidance page with field-by-field instructions including guidance on what constitutes a meaningful performance metric vs. a vanity metric

Framework Alignment

• EU AI Act Art. 11 & Annex IV — Technical documentation requirements for high-risk AI systems: model description, performance metrics, data governance, and known limitations
• NIST AI RMF — GOVERN 1.7 & MEASURE 2.5 — AI transparency documentation and bias/fairness evaluation records
• ISO/IEC 42001 Clause 8.4 — AI system documentation requirements including intended use, performance characteristics, and limitations
• OECD AI Principles 1.3 — Transparency and explainability: AI actors should provide meaningful information about AI systems, including their capabilities and limitations
• Mitchell et al. (2019) Model Cards for Model Reporting — Original model cards framework adapted for operational governance use

Download