Filed under — EU AI Act · Article 10 · Data Governance · High-Risk AI · Compliance
Prefer this source on Google →EU AI Act Article 10: Data Governance Requirements Explained
Article 10 mandates training, validation, and testing data governance for high-risk AI. Learn what documentation you need and how to prove compliance before December 2027.
Free check · no account
Check your own system against what this article describes.
EU AI Act Article 10: Data Governance Requirements Explained
Updated 2 October 2026 — correction. Earlier versions cited the wrong paragraphs of Article 10, said the Act requires test data to be kept separate from training data, and said systems already in production must comply by 2 December 2027. The guide now cites Article 10(2) for governance and bias examination and Article 10(3) for data quality, and follows Article 111(2) as amended by the Digital Omnibus for systems already on the market.
Updated 31 July 2026 — deadline changed. The Digital Omnibus (adopted by the European Parliament on 16 June 2026 and the Council on 29 June 2026) deferred the Annex III high-risk obligations from 2 August 2026 to 2 December 2027. The Article 50 transparency obligations were not deferred and still apply from 2 August 2026. This article has been corrected accordingly. If your AI system is classified as high-risk under the EU AI Act, Article 10 is non-negotiable. It mandates specific data governance practices for training, validation, and testing datasets — and these obligations apply from 2 December 2027. The statutory maximum is €15 million or 3% of global annual turnover, whichever is higher — and for SMEs, whichever is lower.
Most teams assume "we have data lineage" equals compliance. It doesn't. Article 10 requires documented design choices, bias mitigation steps, and statistical properties of every dataset used to train or validate a high-risk system.
This guide walks through what Article 10 actually requires, which systems it applies to, and how to document compliance before the deadline.
What Article 10 Requires
Article 10 applies to high-risk AI systems listed in Annex III (e.g., HR screening tools, credit scoring, biometric identification, critical infrastructure management), and to those classified as high-risk through Annex I. It mandates that training, validation, and testing data meet specific quality criteria (Regulation (EU) 2024/1689, Article 10):
| Requirement | What It Means | Documentation You Need |
|---|---|---|
| Relevant, sufficiently representative, free of errors and complete — the last two "to the best extent possible" (Art. 10(3)) | Data must reflect the real-world use case without systematic gaps | Dataset composition report showing demographic/geographic coverage |
| Appropriate statistical properties (Art. 10(3)) | Data must have sufficient volume, variance, and balance for the task | Statistical summary: sample size, class distribution, variance metrics |
| Examination for biases (Art. 10(2)(f)–(g)) | You must actively search for and document biases that could lead to discriminatory outcomes | Bias audit report with mitigation steps (e.g., resampling, fairness constraints) |
| Data governance and management (Art. 10(2)) | Formal processes for data collection, labeling, storage, and versioning | Data governance policy document + audit trail of dataset versions |
Article 10 does not prescribe specific statistical tests or bias metrics. That's intentional — the regulation is technology-neutral. But it does require you to document your choices and explain why they're appropriate for your system's risk profile.
Who Article 10 Applies To
Article 10 obligations fall on providers of high-risk AI systems — the entity that develops the system or has it developed and places it on the EU market under their name or trademark.
If you're a deployer (an organization using a high-risk system developed by someone else), Article 10 compliance is the provider's responsibility. Your own duty is narrower: to the extent you control the input data, you must ensure it is relevant and sufficiently representative for the system's intended purpose (Article 26(4)). It is still worth asking the provider for evidence of its Article 10 work, especially if you're in a regulated sector (finance, healthcare, public services).
If you're a startup or scale-up building your own AI, you are the provider. Article 10 applies in full.
The Five Data Governance Practices Article 10 Demands
1. Dataset Design Choices Must Be Documented
Why did you choose this dataset? What real-world population or scenario does it represent? What are its known limitations?
Example: If you're building an AI-powered resume screener (Annex III, category 4), your training data must represent the actual applicant population you'll encounter. If your dataset is 80% male CVs from tech roles, and you deploy the system to screen healthcare applicants, the dataset appears not to meet Article 10(3) and 10(4), which require data that is sufficiently representative and that takes into account the setting in which the system is intended to be used.
What to document:
- Dataset source and collection methodology
- Geographic, demographic, and domain coverage
- Known gaps or underrepresented groups
- Rationale for dataset selection
2. Statistical Properties Must Be Appropriate
"Appropriate" means sufficient for the task's risk level and complexity. A high-risk credit scoring model needs more rigorous statistical validation than a low-risk content recommendation engine.
What to document:
- Sample size and how it was determined
- Class distribution (e.g., 60% approved loans, 40% rejected)
- Feature variance and correlation analysis
- Train/validation/test split ratios and methodology
If your dataset is imbalanced (e.g., 95% negative class), document why that reflects reality and what steps you took to prevent the model from ignoring the minority class (e.g., stratified sampling, class weighting, SMOTE).
3. Bias Examination Is Mandatory
Article 10(2)(f) explicitly requires examining datasets "in view of possible biases" that are likely to affect health and safety, have a negative impact on fundamental rights, or lead to discrimination prohibited under Union law (on grounds such as race, sex, age or disability). Article 10(2)(g) then requires appropriate measures to detect, prevent and mitigate the biases found.
This is not optional. You must actively search for bias, document what you found, and explain your mitigation strategy.
Practical steps:
- Slice your dataset by protected attributes (if available) and measure performance disparities
- Use fairness metrics (e.g., demographic parity, equalized odds, calibration) appropriate to your use case
- Document any disparities found and the remediation steps taken (e.g., rebalancing, fairness constraints, post-processing)
- If protected attributes are not in your dataset, document proxy analysis (e.g., ZIP code as a proxy for race in US credit data)
- If detecting and correcting bias requires special categories of personal data (such as ethnic origin or health), the Act allows their processing only exceptionally, where strictly necessary and under strict conditions. That rule was Article 10(5); since the Digital Omnibus it is Article 4a(1) (Regulation (EU) 2026/1744)
Example: A hiring AI trained on historical data may learn that "gaps in employment" correlate with rejection — but if women are more likely to have employment gaps due to parental leave, the model encodes gender bias. Article 10 requires you to detect and mitigate this.
4. Data Governance Processes Must Be Formalized
Article 10(2) requires "data governance and management practices" appropriate for the system's intended purpose — not just good intentions, but documented processes. It lists what they must cover, from design choices, data collection and data preparation to the identification of data gaps.
Minimum documentation:
- Data collection policy (who can add data, under what conditions)
- Labeling guidelines and quality control (inter-annotator agreement scores, label audits)
- Data versioning and lineage (which model version was trained on which dataset version)
- Access controls and audit logs (who accessed training data, when, and why)
If you retrain your model on new data, you must repeat the Article 10 analysis for the updated dataset. One-time compliance is not sufficient.
5. Testing Data Must Be Representative — and Should Be Independent
The quality criteria of Article 10(3) — relevant, sufficiently representative, and to the best extent possible free of errors and complete — apply to testing datasets as they do to training and validation datasets. For a system that is not trained on data, Article 10(6) applies the data requirements to the testing datasets only.
Article 10 does not itself require test data to be kept separate from training data. But the Act defines testing data as data used for "an independent evaluation" of the system before it is placed on the market (Article 3(32)). A model evaluated on the data it was trained on has not had that independent evaluation, so keeping the two apart is both basic ML hygiene and the reading the definition supports.
What to document:
- How you ensured test data independence (e.g., temporal split, stratified holdout)
- Why your test set represents real-world deployment conditions
- Test set performance broken down by subgroups (to detect disparate impact)
Common Article 10 Compliance Gaps
Most teams building high-risk AI have some data governance practices. But few have the documentation Article 10 demands. Here are the most common gaps:
- No bias examination documentation — teams run fairness metrics but don't document findings or mitigation steps
- No dataset design rationale — teams use "whatever data we had" without documenting why it's appropriate
- No versioning or lineage — teams retrain models but can't trace which dataset version produced which model version
- No statistical justification — teams don't document why their sample size, class balance, or feature set is sufficient for the risk level
- No formal governance policy — data practices exist informally but aren't written down or auditable
How to Document Article 10 Compliance
Article 10 compliance is proven through technical documentation (required under Article 11, with the data sheets listed in Annex IV, point 2(d)). At minimum, you need:
-
Dataset Specification Document — for each dataset (training, validation, test):
- Source, collection date, and methodology
- Size, structure, and statistical properties
- Known limitations and gaps
- Bias examination results and mitigation steps
-
Data Governance Policy — organization-wide:
- Data collection and labeling standards
- Versioning and lineage tracking
- Access controls and audit procedures
- Retraining and re-evaluation triggers
-
Model Card or Technical Documentation — per model:
- Which datasets were used (with version hashes)
- Why those datasets are appropriate for the use case
- Test set performance overall and by subgroup
- Residual risks and monitoring plan
These documents must be maintained and updated throughout the system's lifecycle. If you retrain, you update the documentation. If you discover a new bias, you document it and your response.
Article 10 and the December 2027 Deadline
Article 10 obligations apply from 2 December 2027 for Annex III high-risk AI systems, and from 2 August 2028 for those classified as high-risk through Annex I.
A system already placed on the market or put into service before that date is not caught by the high-risk requirements unless its design changes significantly from that date on (Article 111(2), as replaced by the Digital Omnibus). A significant design change brings Article 10 with it. The exception is systems intended to be used by public authorities: their providers and deployers must comply by a separate deadline set in Article 111(2), whether or not the design changes.
If you're launching a new high-risk system after 2 December 2027, Article 10 compliance is required before you place it on the market.
The operative date is 2 December 2027. The statutory maximum for breach of the high-risk operator obligations is €15 million or 3% of global turnover under Art. 99(4) — and for SMEs, whichever is lower. The €35 million / 7% tier under Art. 99(3) applies only to the prohibited practices in Article 5, which is a different thing entirely.
This article is for informational purposes only and does not constitute legal advice. Consult a qualified EU AI Act attorney for compliance guidance specific to your system.
Related dispatches