Model Cards and Datasheets for Datasets

Model cards were proposed by Margaret Mitchell and colleagues in the paper "Model Cards for Model Reporting", presented at the ACM FAT* conference in 2019. The suggested sections are model details, intended use, factors (such as demographic or environmental conditions), metrics, evaluation data, training data, quantitative analyses, ethical considerations, and caveats and recommendations. The key idea is disaggregated evaluation: reporting performance separately for relevant subgroups rather than as one average score that can hide failures. Datasheets for Datasets, by Timnit Gebru and colleagues, first circulated in 2018 and was later published in Communications of the ACM. It borrows from electronics datasheets. Dataset creators answer questions about motivation, composition, collection process, preprocessing and labeling, uses, distribution and maintenance. Related formats include data statements for language datasets and data cards. System cards, used by Meta and OpenAI since 2022 and prominently for GPT-4 in 2023, document a whole deployed system, including safety testing and mitigations, rather than only the model. Regulation has turned much of this practice into a requirement. The EU AI Act requires technical documentation for high-risk AI systems, set out in Annex IV. Providers of general-purpose AI models must keep technical documentation, give information to downstream providers, and publish a summary of training content using a Commission template. A common misconception is that a model card proves a model is safe. It is a disclosure, not a certification, and a card can be incomplete, outdated or written like marketing. Its value depends on honest limitations and evaluations that someone else can reproduce.