What it asks for. High-risk AI systems that are trained on data must be developed using training, validation and testing data sets that meet quality criteria. Data sets must be relevant, sufficiently representative and, to the best extent possible, free of errors and complete in view of the intended purpose, and subject to appropriate data governance and management practices.
In practice. Providers document design choices, data collection processes and origin, the original purpose of any personal data, preparation steps such as labelling and cleaning, the assumptions the data is meant to represent, and an assessment of availability and suitability. They examine data for possible biases that could affect health, safety or fundamental rights, apply measures to detect, prevent and mitigate them, and record data gaps and how they are addressed. The data must reflect the geographical, contextual and behavioural setting of use. Special category data may be processed for bias detection only under strict safeguards set out in the article.
Common gaps. Typical gaps are missing data lineage, no documented bias examination, and training data that does not represent the population or region where the system is actually deployed.