Datasheet for datasets

Datasheets for datasets, proposed by Gebru and colleagues, borrow the idea of electronics datasheets: every dataset should come with answers to a fixed set of questions. Who created it and why? What does each instance represent, and are any groups missing or overrepresented? How was the data collected, with what consent, and was any personal or sensitive information included? What cleaning or labelling was done, and which uses would be inappropriate?

The practice makes data quality and representativeness reviewable, which is exactly what the EU AI Act asks of training, validation and testing data for high-risk systems (Art. 10): relevant, representative, examined for possible biases and governed by documented practices. It also supports data provenance and the lawful basis analysis required by data protection law.

Example: a hospital releasing an imaging dataset documents that all scans came from two devices in one region, warning model builders that performance elsewhere must be tested.