Model Card

Think of a model card as a model's datasheet — the technical and ethical specification that tells a deployer exactly what they are getting, where it works, where it fails, and what risks to manage. A complete model card covers six domains: (1) **Model details** — architecture, version, training dates, and responsible development team; (2) **Intended use** — primary use cases, in-scope users, and explicitly out-of-scope applications; (3) **Factors** — demographic, geographic, or contextual factors that affect model behavior; (4) **Metrics** — performance benchmarks including disaggregated performance across subgroups; (5) **Evaluation data** — datasets used for evaluation and their limitations; (6) **Ethical considerations and caveats** — known failure modes, bias risks, and recommended mitigations.

The operational value of model cards in an enterprise context extends well beyond transparency theater. During procurement, model cards enable systematic vendor evaluation — comparing models across standardized performance and risk dimensions rather than relying on marketing claims. During deployment, model cards define the operating envelope, enabling AI risk managers to flag when a model is being applied outside its intended scope. During incidents, model cards provide the baseline documentation required for post-mortem analysis and regulator communication.

The regulatory significance is growing rapidly. The EU AI Act's Article 11 technical documentation requirements and Annex IV are essentially a mandated model card for high-risk AI systems. The FDA's guidance on AI/ML-based Software as a Medical Device (SaMD) requires similar lifecycle documentation. Enterprises that have invested in model cards are discovering they already have much of the documentation required for compliance — enterprises that have not face urgent remediation.