In short
- Article 10 applies data governance duties to training, validation and testing sets, which is 3 data sets rather than one.
- Article 10(2) names 8 practices: design choices, collection processes and origin, preparation operations, assumptions, availability and suitability, bias examination, bias mitigation, and identification of gaps.
- Article 10(3) sets the substantive standard: relevant, sufficiently representative, and to the best extent possible free of errors and complete in view of the intended purpose.
- Regulation (EU) 2026/1744 deleted Article 10(5) and moved the special category gateway to a new Article 4a, which has applied since 27 July 2026 and now also covers deployers and systems that are not high risk.
- Annex IV point 2 requires the same ground in the technical documentation, so the evidence has 2 homes and must agree in both, from 2 December 2027 for Annex III systems.
Article 10 is the obligation that turns a data pipeline into a documented one, and it is the one where the answer most often is that nobody wrote it down at the time.
The provision is not asking for a data catalogue. It is asking for an account of judgement calls.
01What does Article 10 of the AI Act actually cover?
Three data sets and eight practices.
Article 10(1) applies the duty to training, validation and testing data sets. That plural is doing work: an organisation with immaculate training data documentation and an undocumented test set has satisfied a third of the requirement. Article 10(6) closes the other side, applying the same criteria to testing data where a high-risk system is built without training a model at all.
The 8 practices in Article 10(2) run from the relevant design choices, through data collection processes and the origin of the data, preparation operations such as annotation, labelling, cleaning, updating, enrichment and aggregation, the assumptions made about what the data measures and represents, an assessment of availability, quantity and suitability, examination for possible biases, measures to detect, prevent and mitigate them, and identification of relevant data gaps.
02Which parts of Article 10 can teams usually evidence?
| Practice | Usually evidenced | What is missing |
|---|---|---|
| Preparation operations | Yes | Version of the pipeline that ran |
| Suitability and quantity | Partly | Written assessment, not just counts |
| Origin of the data | Rarely | Licence and lawful basis per source |
| Assumptions | Rarely | What the data was taken to represent |
| Data gaps | Rarely | Named populations not covered |
The last row is the one auditors press hardest, because a stated gap is evidence of examination and an absent gap list usually means the examination did not happen.
03What is the substantive quality standard in Article 10(3)?
Relevant, representative, and as free of errors and complete as possible.
Article 10(3) of Regulation (EU) 2024/1689 requires that training, validation and testing data sets be relevant, sufficiently representative, and to the best extent possible free of errors and complete in view of the intended purpose. Each of those is judged against the intended purpose rather than in the abstract, which means the same data set can satisfy the standard for one use and fail it for another.
That relativity is the reason a change of intended purpose is treated so seriously elsewhere in the Regulation. Repointing a system at a new decision does not only change its risk classification. It can invalidate the data assessment that supported it.
04Where did the special category gateway go?
Out of Article 10 and into Article 4a, with a wider reach.
Regulation (EU) 2026/1744 deleted Article 10(5) of the AI Act and re-enacted the permission as a standalone Article 4a, in force since 27 July 2026. Two things changed with the move.
Article 10(1) now points at paragraphs 2, 3 and 4 of Article 10 together with Article 4a(1), so a citation to Article 10(5) in a policy written before mid 2026 names a provision that no longer exists. And Article 4a(2) widens the personal scope beyond providers of high-risk systems to deployers of those systems and to providers and deployers of other AI systems and models, while stating expressly that it creates no obligation to carry out bias detection.
| Condition | Requirement | Evidence |
|---|---|---|
| Necessity | Strictly necessary for bias detection and correction | Written necessity assessment |
| Alternatives | Other, synthetic or anonymised data considered first | Record of why they were insufficient |
| Access | Strict access control and technical limits on reuse | Access list, 1 named owner |
| Onward sharing | No transmission to other parties | Sharing prohibition in the system record |
| Retention | Deleted once the bias is corrected | Deletion record with a date |
Organisations tend to read the gateway as permission and stop there. It is permission conditioned on a documented chain: why it was necessary, what was tried first, who could see it, whether it left the building, and when it was destroyed. Without the last item the gateway has been used and not closed.
05When does Article 10 start to bind?
2 December 2027 for Annex III, 2 August 2028 for Annex I.
Regulation (EU) 2026/1744 moved the application of the high-risk requirements in Chapter III to 2 December 2027 for standalone systems under Article 6(2) and Annex III, and to 2 August 2028 for AI embedded in products regulated under Annex I. Article 10 sits inside that block, so the data governance duty carries those dates too. Article 4a does not: it has applied since 27 July 2026, which is the one part of this subject that is live now.
The deferral is not slack in the schedule, for the reason set out in our account of which duties the delay left untouched. Origin and assumptions are the two practices that cannot be reconstructed after the fact, and every month of undocumented collection between now and 2 December 2027 is a month that will have to be described from memory.
06Where does the evidence live?
In two places that must agree.
Annex IV point 2 requires the technical documentation to describe the development process, including a general description of the training data sets and information about their provenance, scope and main characteristics. Article 10 requires the governance practice itself. An organisation that maintains a thorough internal data document and a summary technical file that says something slightly different has created the worst possible artefact, because the discrepancy is the finding.
The workable pattern is one source document per system, with the technical file generated from it rather than written separately. It is less elegant than it sounds and it removes an entire class of audit question.
07Frequently asked questions
What does Article 10 of the AI Act require for training data?
Data governance and management practices appropriate to the intended purpose. The 8 practices named in Article 10(2) cover design choices, collection processes and origin, preparation operations such as labelling and cleaning, the assumptions embedded in the data, its availability and suitability, examination for bias, mitigation measures, and identification of data gaps.
Does Article 10 apply to validation and test data too?
Yes. Article 10(1) names training, validation and testing data sets. Teams that document training data thoroughly and treat the test set as an implementation detail have covered one of three. Article 10(6) extends the same criteria to testing data for high-risk systems developed without training.
Can you process special category data to check for bias?
Yes, within limits, but the provision moved. Regulation (EU) 2026/1744 deleted Article 10(5) and re-enacted the gateway as Article 4a of the AI Act with effect from 27 July 2026. Article 4a(1) permits processing of special categories of personal data where strictly necessary for bias detection and correction, subject to conditions including other data types first, technical limits on reuse, strict access control, no transmission to other parties, deletion once the bias is corrected, and a record of why the processing was strictly necessary.
When does Article 10 start to bind?
2 December 2027 for standalone high-risk systems under Article 6(2) and Annex III, and 2 August 2028 for AI embedded in products regulated under Annex I. Article 4a is the exception in this area: it has applied since 27 July 2026, independently of those dates.
What is the most common evidence gap?
Origin. Teams can usually show what preparation was done and what the model scored, and often cannot show where each portion of the data came from, under what licence or lawful basis, and what it was assumed to represent.
08References and method
- Regulation (EU) 2024/1689, Article 10, for the data governance practices in Article 10(2) and the quality criteria in Article 10(3).
- Regulation (EU) 2026/1744 of 8 July 2026, in force 27 July 2026, for the deletion of Article 10(5), the insertion of Article 4a, and the application dates of 2 December 2027 and 2 August 2028 for high-risk obligations. Consolidated text of the AI Act as at 27 July 2026 consulted on 15 September 2026.
- Regulation (EU) 2024/1689, Annex IV point 2, for the corresponding technical documentation requirement covering the development process and the training data sets.
- Regulation (EU) 2016/679, Article 9, for the special categories of personal data referred to by Article 4a.
- The evidence gap observations are editorial, prepared by The Guardrail from publicly documented audit expectations. Practice varies between certification bodies.