Annex IV outlines data requirements

Annex IV of the AI Act requires detailed descriptions of training, validation, and testing data for high-risk AI systems. This includes information about the origin, scope, and characteristics of the data, as well as how it was obtained and processed.

Provenance and data governance defined

The AI Act uses terms like "provenance" and "data governance" instead of "data lineage." These terms require a dated, written trace connected to a specific model version, emphasizing clear documentation of data sources and management practices.

Written practice for data sets

Article 10 of the AI Act mandates that data sets be managed according to appropriate governance practices. This includes tracking data collection methods, preparation steps like labeling and cleaning, and identifying and addressing gaps or deficiencies in the data.

Versioning and updating documentation

Article 11 requires that technical documentation be created before a system is deployed and kept up to date. This means that documentation must evolve with each model version, recording details like data sources, selection methods, and validation procedures for each update.

Provenance versus system logs

Article 12 distinguishes between provenance—documenting where training data comes from—and system logs, which record events during the AI system's operation. Both are important but documented in different parts of the technical documentation.

Documentation retention requirements

Article 18 specifies that technical documentation must remain accessible to national authorities for ten years after a system is commercialized or deployed. This ensures that data links remain readable even if tools or personnel change over time.

Regulation delays and simplified reporting

Regulation (EU) 2026/1744 has delayed the application of certain AI Act provisions. High-risk systems in Annex III must comply by December 2, 2027, and those in Annex I by August 2, 2028. SMEs and start-ups may use a simplified form for compliance, though the exact format is still being developed by the European Commission.