Article 10 · Chapter III: High-Risk AI Systems
Data and Data Governance
High risk
Summary
High-risk AI systems trained on data must use training, validation, and testing datasets that meet quality criteria — relevance, representativeness, accuracy, completeness — and incorporate appropriate data-governance and management practices, including bias detection and mitigation. Permits processing of special-category personal data strictly to detect and correct bias, with safeguards.
Key obligations
- Document data sources, collection, labelling, cleaning, aggregation
- Examine datasets for bias and correct as far as possible
- Ensure relevance, representativeness, accuracy and completeness
- Limit processing of sensitive data to bias mitigation, with technical safeguards