EU AI Act Art. 10: Preparing High-Risk AI Training Data by 2 December 2027

By George Curta · Published March 17, 2026 · 8 min read · EU AI Act Compliance

The EU AI Act imposes data governance obligations on providers of high-risk AI systems under Article 10. For any training dataset that contains personal data, anonymization supports compliance — removing PII before it ever enters the training pipeline. anonym.plus processes training datasets entirely offline, keeping your data inside your infrastructure.

Who Is Affected by EU AI Act Art. 10

Article 10 applies to providers of high-risk AI systems — organizations that develop, train, or deploy AI systems listed in Annex III of the EU AI Act. These include:

• AI systems for biometric identification and categorization
• AI used in critical infrastructure (transport, energy, water)
• Educational and vocational training AI
• AI in employment decisions (hiring, HR management, worker monitoring)
• Essential private and public services (credit scoring, insurance risk assessment)
• Law enforcement AI
• Migration, asylum, and border control AI
• AI in administration of justice

Organizations that fine-tune foundation models (GPT-4, Claude, Llama) on their proprietary datasets for these purposes are also covered.

What Article 10 Requires for Training Data

Article 10 mandates that training, validation, and testing data must:

1. Be relevant, representative, and free from errors for the intended purpose
2. Have appropriate statistical properties for the AI's use case
3. Take into account biases that could lead to prohibited discrimination
4. Be subject to documented data governance practices — covering origin, collection methods, preprocessing, and known limitations
5. Personal data in the data sets must be covered by the documented data-governance practices under Art. 10(2)(b), including the original purpose of the data collection; special categories of personal data only under the conditions for bias detection and correction in Art. 4a(1) (formerly Art. 10(5))

Art. 10 does not forbid personal data in training data. Where a data set contains it, the data-governance record must cover it — including the original purpose of its collection (Art. 10(2)(b)) — and the GDPR applies to it in full, including the need for a lawful basis for the training. Special categories of personal data may be processed only where strictly necessary for bias detection and correction, under the safeguards of Art. 4a(1).

Anonymization as Part of Art. 10 Data Governance

Removing personal data from training datasets before the AI training pipeline begins is a practical way to support the Art. 10 data-governance duties:

• Anonymized training data is not personal data (GDPR Recital 26). No GDPR lawful basis required for training. No data subject rights apply to the dataset. No DPA needed for processors handling the dataset.
• Reduces the personal-data part of the Art. 10 data-governance record — the personal data detected in the training data is removed; questions of provenance and the original purpose of collection remain (Art. 10(2)(b)).
• Data governance documentation is simplified — you document that PII was removed, what entity types were detected, and what tool was used.

Training Data Formats Supported by anonym.plus

For large datasets above these limits, process files in batches using anonym.plus batch mode (Pro plan). All processing is 100% offline — training data never leaves your infrastructure.

Which PII to Remove from Training Data

For EU AI Act compliance, prioritize removing:

• Direct identifiers: names, email addresses, phone numbers, national IDs, passport numbers
• Quasi-identifiers: dates of birth, job titles, postal codes, rare combinations of demographic attributes
• Special categories (Art. 9 GDPR): health data, racial/ethnic origin indicators, religious beliefs, political opinions, union membership, sexual orientation
• Financial data: IBANs, credit card numbers, account numbers
• Location data: precise GPS coordinates, home addresses, frequently visited places

anonym.plus detects all of these through 340+ built-in entity types. The European Auto-Detect preset (confidence 0.85) is the recommended starting point for training data preparation.

Documenting Compliance for Art. 10

After anonymizing your training datasets, document the following in your AI system's technical documentation (required under Art. 11):

• Data sources and collection methods
• PII removal method: anonym.plus v[x.x], Replace operator, European Auto-Detect preset, confidence threshold 0.85
• Entity types detected and replaced
• Date of processing and dataset version
• Any residual risks identified and mitigations applied

anonym.plus creates a processing history entry for each file, including entity counts, operator used, and timestamp — supporting this documentation requirement.