Model Card

A model card is a concise, standardised document that accompanies an AI model and provides essential information about it. Proposed by Google researchers in 2019, model cards serve as a form of 'nutrition label' for AI models, making key information accessible to users, developers, regulators, and other stakeholders. A typical model card includes: model details (architecture, version, developer), intended use (primary use cases and out-of-scope applications), performance metrics (accuracy, fairness measures across demographics), training data (description of data sources, size, preprocessing), ethical considerations (known biases, limitations, risks), and recommendations (guidance for appropriate use). Model cards are increasingly adopted across the AI industry. Major AI companies publish model cards for their public models, and the practice is becoming an expected standard. The EU AI Act's documentation requirements align closely with model card content, making them a practical tool for compliance.

Model cards provide the documentation needed for responsible AI governance. They create a shared understanding of what a model can and cannot do, helping users make informed decisions about whether and how to use it. This prevents misuse and misaligned expectations. For organisations building custom models, creating model cards instils documentation discipline that improves model quality and governance. The process of filling out a model card forces developers to evaluate performance across groups, identify limitations, and think through ethical implications — activities that improve the model even if no one reads the card. Model cards also support compliance with emerging AI regulations. The EU AI Act requires technical documentation that covers many of the same elements as model cards. Organisations that adopt model card practices now are building the documentation habits that regulations will soon require.