AI Act Article 10 requires providers of high-risk AI systems to govern how their training, validation and testing data is designed, collected, prepared, examined for bias and checked for gaps. The rule asks for documented practices that suit the system's intended purpose, not a single technical test. Teams that record these decisions from the start meet the requirement far more easily.
Key takeaways
- Article 10 applies to high-risk AI systems and covers training, validation and testing data.
- The obligation is a documented governance process, covering design choices, data origin, preparation and gap analysis.
- Bias examination and handling of sensitive data both need explicit safeguards and a clear legal basis.
- Governance fails most often through missing lineage records, unreviewed data changes and unclear ownership.
What does AI Act Article 10 require?
Article 10 requires providers of high-risk AI systems to apply data governance and management practices to their training, validation and testing data sets. Those practices must fit the intended purpose of the system and be written down in a way that someone outside the project can follow.
The obligation sits in Article 10 of Regulation (EU) 2024/1689, which you can read in the official EUR-Lex text of the AI Act. This article is general information and not legal or tax advice. Rui Codex provides technical implementation support only, so the legal reading of your obligations should come from qualified counsel.
The text lists practices that providers must address. They include the design choices made for the data, the collection processes and origin of the data, the preparation operations such as annotation, labelling, cleaning, updating, enrichment and aggregation, and the assumptions the team makes about what the data represents. The list also covers an assessment of data availability, quantity and suitability, an examination for possible biases, and the identification of data gaps and how they will be addressed.
Article 10 also sets quality expectations. Datasets should be relevant, sufficiently representative and, as far as possible, free of errors and complete in view of the intended purpose. The regulation does not give a single metric or threshold for these qualities, so each provider must define what adequate means for its system and justify that choice.
Which data sets does Article 10 cover?
Article 10 covers the training, validation and testing data sets used to develop a high-risk AI system. For systems that do not use training techniques, the requirements apply to the testing data sets. In practice, the boundary is drawn by how the data is used in the development lifecycle, not by where it is stored.
The same source data often appears in several roles. A customer record might feed a training set, a validation split and a production monitoring report. Governance has to follow the data through each role, which makes lineage the core technical requirement.
Data that comes from operational systems is usually the hardest to govern. Records in an ERP, CRM or ticketing tool change daily, use fields with inconsistent meanings, and are often copied through integrations before they reach a model pipeline. Mapping those flows is a systems problem as much as a data problem, and it is where our work on system integrations is directly relevant.
How do you build a data governance process?
You build it by turning each Article 10 practice into a recorded step with an owner, an input, an output and a review point. The process does not need specialised tooling to start. A shared register, versioned dataset snapshots and a short decision log cover most of the requirement for a first release.
Step 1: Fix the intended purpose and the context of use
Describe what the system decides, who uses it, and in which setting it operates. The context matters because Article 10 asks you to consider the geographical, behavioural or functional setting in which the system is used. A model trained on data from one customer segment may behave differently in another.
Step 2: Register every data source with its origin
For each source, record where it came from, under which legal basis it was obtained, the original purpose if it contains personal data, and the transformations applied since. A register entry should let a reviewer reconstruct the path from raw record to training row without asking the original engineer.
Step 3: Record preparation decisions as code and prose
Cleaning rules, label definitions and aggregation logic should live in version-controlled code, with a short written rationale for each non-obvious choice. Prose alone drifts away from the pipeline, and code alone hides the reasoning behind it.
Step 4: Test representativeness and gaps
Compare the data distribution with the population the system will face, and list the groups or conditions that are thin or absent. Each gap needs a recorded response: collect more data, restrict the intended purpose, add human review, or accept the limitation and document it.
Step 5: Set explicit review triggers
Define the events that reopen the governance record, such as a new data source, a retrained model, a change in the upstream system or a pattern of complaints. Without triggers, the register describes the dataset as it was at launch and nothing more.
For teams that need help with the technical side of these measures, our page on AI Act technical measures describes the implementation support we offer.
Where do training data governance programmes fail?
Most programmes fail at the boundaries between teams, not inside any single step. The main decisions involve trade-offs, and each option has a failure mode that is worth understanding before you commit to it.
- Central catalogue versus embedded records. A central catalogue gives auditors one place to look, but it goes stale when engineers do not update it. Records embedded in pipelines stay current but are harder to search across projects.
- Aggressive filtering versus volume. Removing noisy or incomplete records improves quality, but it can remove the very minority cases that a bias examination needs. Filter rules should be tested against representativeness, not only against error rates.
- Synthetic data versus real coverage. Generated records can fill gaps, yet they reproduce the assumptions of the generator. A synthetic set that looks balanced on paper can hide a gap that exists in the real world.
- Frequent versioning versus review capacity. Snapshotting every change gives a complete history, but reviewers quickly stop reading the logs if the volume is too high. Version on meaningful changes and keep the rest in machine-readable logs.
The common failure pattern is a register written once for the launch review and never updated. The second is a bias analysis run on a sample that no longer matches the production data. Both are preventable with clear owners and review triggers, and both are easy to miss when a project is under delivery pressure.
Can sensitive data be used to test for bias?
Article 10 allows providers to process special categories of personal data for the purpose of detecting and correcting bias, but only under the conditions and safeguards the article sets out. Outside those conditions, the normal rules still apply.
The AI Act text is available on EUR-Lex, and the GDPR, which restricts processing of special categories of personal data, is on EUR-Lex as well. Both sources should be read together, because the AI Act does not replace the GDPR's lawful basis and purpose limitation rules.
In practice, the bias analysis should be planned as its own documented activity. Record which attributes are used, why they are needed for the examination, which technical measures limit access, and when the data is deleted. The analysis should be treated as a separate processing activity, not as a side effect of model development.
Hypothetical examples
Here are two hypothetical scenarios:
Scenario one. A team builds a document classification tool for an internal operations department. Its training data comes from a ticketing system whose categories were changed twice in the past three years. The team discovers the change only when validation results shift, and it resolves the problem by versioning the category mapping and documenting the period each label covers.
Scenario two. A team builds a scoring model that uses records from two regional branches. The records from one branch are far more complete, so the model performs unevenly. The team records the gap, restricts the intended purpose to the well-documented branch for the first release, and schedules a data collection review before expanding its scope.
Neither scenario describes a real organisation. Both show how a register, a clear purpose statement and a review trigger turn a data problem into a documented decision.
Frequently Asked Questions
Does Article 10 apply to every AI system?
Article 10 applies to high-risk AI systems, so it does not automatically cover every AI tool a company uses. Whether a system is high-risk depends on its use case and the classification rules in the regulation, so check the official text before deciding. Even for systems outside scope, recording data decisions helps debugging and audit trails.
Does Article 10 require a specific data quality metric?
No, Article 10 does not prescribe one metric or threshold. It asks that data be relevant, sufficiently representative and as free of errors and complete as possible for the intended purpose. The provider decides what adequate means for its system and documents the reasoning behind that choice.
How often should a training data set be reviewed?
Article 10 does not set a fixed review interval. The schedule should follow the intended purpose, changes in data sources and monitoring results. Set explicit triggers such as a new data source, a retrained model or a pattern of complaints, and record each review.
Is synthetic data allowed under Article 10?
Article 10 does not ban synthetic data, but synthetic records must be documented like any other data source. Generated data can fill gaps, although it can also reproduce the assumptions of its generator. Test it against real-world distributions and record how it was made and why it was used.
How does the GDPR interact with Article 10?
Both apply when training data contains personal data. The GDPR governs the lawful basis, purpose limitation and special category rules, while Article 10 governs data quality and governance for high-risk systems. Good documentation often supports both, but the two obligations should be assessed separately.
Conclusion
Article 10 asks for a governance process that a reviewer can follow, from the intended purpose through data origin, preparation, bias examination and gap handling. The process works best when it starts before the first training run and is kept current through clear triggers and named owners. Teams that treat the register as part of the engineering work, rather than a document written for an audit, find the requirement far easier to meet. If you are planning a high-risk system and want technical support with the data pipeline, you can request a project consultation at the Rui Codex contact page.
Sources
- Regulation (EU) 2024/1689 (Artificial Intelligence Act) - European Union, EUR-Lex
- Regulation (EU) 2016/679 (General Data Protection Regulation) - European Union, EUR-Lex