Model Card

A model card is a structured document that accompanies an AI model, providing essential information about what the model does, how it was built, what it is good at, and where it falls short. Think of it as a nutrition label for AI.

Why model cards exist

AI models are not self-explanatory. Without documentation, users cannot know what data a model was trained on, what biases it might carry, how it performs on different demographics, or what tasks it was designed for versus tasks it should not be used for. Model cards make this information transparent and accessible.

The concept was introduced by researchers at Google in 2018 and has since become an industry standard, particularly for models released publicly.

What a model card includes

• Model details: Name, version, architecture, developer, release date.
• Intended use: What tasks the model is designed for, and importantly, what it should not be used for.
• Training data: Description of the data used to train the model, including any known gaps or biases in that data.
• Performance metrics: Benchmarks showing how well the model performs on standard tasks, broken down by relevant categories.
• Limitations: Known weaknesses, failure modes, and scenarios where the model performs poorly.
• Ethical considerations: Potential harms, fairness assessments, and bias evaluations.

Who reads model cards

• Engineers use them to determine if a model suits a particular technical requirement.
• Product managers use them to understand limitations before committing to a model.
• Compliance teams use them to assess risk and regulatory alignment.
• Researchers use them to compare models and identify areas for improvement.

Model cards in practice

Major platforms like Hugging Face require model cards for all hosted models. Cloud providers like Google, Microsoft, and Amazon publish model cards for their AI services. Responsible AI teams increasingly treat model card completion as a mandatory step before deployment.

Limitations of model cards

They are only as good as the effort put into them. Some model cards are thorough and genuinely useful; others are superficial checkboxes. They also represent a snapshot in time and may not reflect changes in model behaviour after updates.