Model cards and documentation

Lens adapted to this topic: The case for model cards and better implementation

The mainstream account treats model cards as a practical transparency framework: they make intended use, evaluation conditions, subgroup performance, limitations, and risks more visible to users and governance teams. Their value depends on relevant metrics, candid disclosure, and ongoing maintenance rather than on the document’s existence alone.

• The original model-card framework aims to reduce inappropriate use by documenting intended contexts, evaluation procedures, performance across relevant demographic or other subgroups, and limitations.
• Model cards can support governance by recording intended use, risk ratings, training details, metrics, evaluation observations, recommendations, and model-version changes.
• Documentation is not only a developer task: effective cards may require contributions from developers, sociotechnical specialists, and project organizers with knowledge of scope and deployment.
• Interactive model cards are proposed as a way to make documentation more understandable and useful to non-expert analysts who use models without formal ML training.
• The existence of incomplete cards indicates a need to improve documentation practice, not necessarily to abandon the model-card framework.

Lens adapted to this topic: The limits of model cards in real-world deployment

The critical account accepts that model cards can provide useful evidence but argues that they are often incomplete, self-reported, and disconnected from deployment conditions. A card may describe tests chosen by a vendor without establishing how the model will behave with local users, tools, permissions, escalation procedures, or tolerated harms.

• Documentation practice often falls short of established software-documentation standards, which can impede accountability and contribute to inappropriate or harmful model use.
• A large-scale analysis found uneven informativeness, with limitations and evaluation among the least consistently completed sections; this challenges the assumption that publishing a card ensures meaningful transparency.
• A model card reports what a vendor measured under selected conditions, but does not by itself establish safety or suitability for a particular organization’s users, tools, permissions, and rollback arrangements.
• Critics argue that model cards can drift from a decision-support checklist toward a launch or marketing document, especially when important training-data, evaluation, or out-of-scope-use information is absent.
• Interactive formats may improve access and interpretation, but studies identify continuing challenges involving understandability, skepticism, trust, safety, and different stakeholder needs.