In Depth
Model cards, proposed by Margaret Mitchell et al. at Google in 2019, are structured documentation for machine learning models. They provide essential information including the model's purpose and intended use cases, training data description, performance metrics disaggregated by relevant groups, known limitations and failure modes, ethical considerations, and recommended usage guidelines.
A well-written model card serves multiple audiences: developers need technical details about architecture and performance, product managers need to understand capabilities and limitations, compliance teams need risk assessments, and end users need to know what the model can and cannot do. Model cards create shared understanding and set appropriate expectations across stakeholders.
Model cards have become standard practice in the AI community. Hugging Face requires model cards for all uploaded models. Major AI labs publish model cards for their flagship releases. The practice is evolving from voluntary documentation to regulatory requirements, with the EU AI Act mandating similar documentation for high-risk AI systems. For organizations deploying AI, maintaining current model cards is a foundational governance practice that supports transparency, accountability, and compliance.