Margaret Mitchell

This paper proposes that every trained machine learning model should be released with a 'model card' providing standardized documentation of its intended use cases, performance benchmarks disaggregated across different demographic groups, ethical considerations, and known limitations. The authors argue that without such documentation, models are routinely deployed in contexts their creators never intended, leading to failures that disproportionately affect marginalized communities, and that evaluation metrics averaged across populations can mask significant disparities in performance. The model card framework draws inspiration from nutrition labels and safety data sheets, aiming to create a common language for communicating about a model's capabilities and risks that is accessible to both technical and non-technical stakeholders. The paper includes detailed examples of model cards for two commercial systems, a face detection model and a toxicity classifier, demonstrating how the framework surfaces important information about differential performance across skin tones and dialects. The proposal has been widely adopted across the industry, with model cards now standard practice at organizations including Google, Hugging Face, and Meta, making it one of the most practically impactful contributions to responsible AI development.