A model card is a standardised documentation format used to communicate essential information about an AI model, including its intended use, performance metrics, limitations, and ethical considerations. Model cards play a vital role in AI assurance by promoting transparency, supporting informed decision-making, and enabling responsible deployment.

Originally proposed as a transparency tool, model cards help developers, auditors, and users understand how a model was built, what it was trained on, where it performs well, and where it might fail. They are particularly valuable in high-risk contexts where incorrect or biased outcomes can have significant consequences.

A well-designed model card typically includes:

• Model purpose and intended use cases

• Dataset information and training methodology

• Performance metrics across relevant subgroups

• Known limitations and failure scenarios

• Ethical or fairness considerations

• Recommendations for appropriate use and user training

Model cards support assurance in several ways:

• Facilitating audits by providing a standardised summary of model characteristics

• Improving reproducibility by documenting training data and parameters

• Enhancing transparency by clearly stating what the model can and cannot do

• Guiding procurement and deployment decisions through structured risk disclosures

Assurance teams evaluate model cards to ensure they are complete, accurate, and aligned with best practices and regulatory expectations. The EU AI Act and other regulatory initiatives are beginning to formalise the requirement for model documentation, making model cards a valuable compliance tool.

Model cards are also useful in public-sector AI, where transparency and trust are paramount. For example, a model card for an AI system used in social services can help ensure that frontline workers, policymakers, and citizens understand how eligibility decisions are made.

In sum, model cards make the inner workings of AI systems more accessible and accountable. They are a practical, scalable, and adaptable tool for embedding assurance into the AI development and deployment process.