AI Act Article 10 Data Governance: What Training, Validation and Input Data Must Prove
Article 10 governs the data behind every high-risk AI system. The eight governance practices, the quality criteria, the bias examination, the deployer's input-data duty, and the documents that prove each one when someone asks.
Every high-risk AI system rests on a claim about data: gathered transparently, prepared consistently, examined for bias, fit for the job. Article 10 makes that claim testable: the conditions training, validation and testing data sets must meet, the governance practices around them, and the checks owed to the input data you supply. For a mid-market buyer of a CV-screening tool or a claims-triage model, “our data is high quality” becomes documents you can request and file. What follows separates the provider’s duties from yours.
What Article 10 requires, and who it lands on
Article 10(1): high-risk AI systems using techniques involving the training of AI models shall be developed on the basis of training, validation and testing data sets meeting the quality criteria in paragraphs 2 to 5 — three data sets, three jobs, one governance regime.
Three scope points clear the rest. Paragraphs 2 to 5 bind the provider; a buyer’s leverage is demanding the evidence. Personal data is not a precondition: claims narratives, transaction logs and CV text all fall under Article 10. Article 10(6) covers systems that do not train — for a rules-based triage engine, paragraphs 2 to 5 apply only to the testing data sets.
Timing. The Annex III regime applies from 2 December 2027 after the Digital Omnibus deferral; Article 6(1) systems follow in August 2028 — questionnaires ask for provenance and bias evidence today regardless.
The eight practices under Article 10(2)
Paragraph 2 requires practices appropriate for the intended purpose and lists what they must “concern in particular”. Each point is a document you either have or do not:
- (a) Design choices — why this population, window and variable set, in writing.
- (b) Collection processes and origin of data, and the original purpose of collection for personal data.
- (c) Preparation operations: annotation, labelling, cleaning, updating, enrichment, aggregation — who did it, against which guideline.
- (d) Formulation of assumptions, especially what the data are supposed to measure and represent.
- (e) Availability, quantity and suitability of the data the system’s claims require.
- (f) Examination for biases likely to affect health and safety, harm fundamental rights or cause discrimination prohibited under Union law — especially where outputs feed inputs for future operations.
- (g) Measures to detect, prevent and mitigate those biases — discovery without remediation fails the article.
- (h) Data gaps or shortcomings, and how they are addressed.
Article 10(3) and 10(4): relevance, representativeness and error minimisation
Paragraph 3 sets the bar: data sets shall be relevant, sufficiently representative, and to the best extent possible free of errors and complete in view of the intended purpose.
Two clarifications decide audits. “To the best extent possible” is a duty to try and to show your working: documented correction of omissions and missing values, not perfection. Statistical properties attach to people: appropriate properties regarding the persons or groups the system serves, met per data set or in combination.
Paragraph 4 is the setting test: data must reflect the geographical, contextual, behavioural or functional setting of use. Claims triage trained on motor claims in one low-fraud market fails it on travel claims across six countries; a CV screener built on one subsidiary’s engineering hires fails it when the group’s applicant mix differs.
GDPR Article 5(1)(d) runs alongside — accurate personal data, kept up to date and rectified when wrong. Accuracy is record-level; Article 10 is population-level, and Article 26(9) links the deployer’s DPIA so the two never diverge.
Bias examination: Article 10(2)(f)–(h) and the 10(5) conditions
Bias work under Article 10 is a sequence, not a workshop: examine under (f), mitigate under (g), register what remains under (h). The phrase that decides most cases: where data outputs influence inputs for future operations.
- CV screening. A model trained on a decade of hires in a department that recruited men learns the imbalance as signal; its shortlists decide who joins, and next year’s data passes through them. The duty attaches to the data itself.
- Claims triage. Narratives vary by handler — terse dictation versus pages of note. If note length tracks region and the model reads length as severity, the bias entered at annotation, where points (c) and (f) meet.
Article 10(5) — special categories, under strict conditions. Bias testing sometimes needs data the GDPR forbids, so providers may process it only where strictly necessary and only if six conditions hold: no less intrusive route (including synthetic data); reuse limits and pseudonymisation; documented access controls; no onward transfer; deletion once the bias is corrected; reasons in the records of processing. The Digital Omnibus now carries this permission in Article 4a.
When you only supply input data: your side of Article 10
Most mid-market companies are deployers. Paragraphs 2 to 5 are the provider’s, but Article 26(4) is yours: to the extent you control input data, it must be relevant and sufficiently representative in view of the intended purpose.
“Control” draws the line. Fields your systems capture, forms candidates complete, severity codes handlers key in — yours. The vendor’s training corpus is theirs; say so rather than claim credit.
On the ground. In CV screening: mandatory intake fields candidates cannot complete, recruiter free-text notes, demographic data you strip while the role mix skews. In claims triage: severity codes keyed locally, add-ons dropped by ingestion, narratives captured in one language for a multilingual base.
Alongside sit Article 26(5) monitoring against the instructions for use and Article 26(6) logs under your control, kept at least six months — together they turn an input-data assertion into evidence.
The evidence file: documents, logs and what to ask your vendor
Article 10 has no standalone dossier; proof lands in Article 11 technical documentation — origin, preparation, assumptions, biases, gaps — kept ten years under Article 18 inside the Article 17 quality-management system. Logging is the other half: Article 12 obliges automatic recording of events over the system’s lifetime for risk detection, post-market monitoring under Article 72 and monitoring under Article 26(5); biometric systems add a fixed minimum — period of use, database checked, matching input, verifier identity. The parallel duty sits in Article 19; yours is the Article 26(6) log set. Each paragraph becomes a request:
| Article 10 requirement | Question to ask your vendor | Evidence to request |
|---|---|---|
| Origin of data — 10(2)(b) | “Where does the corpus come from, and why?” | Provenance statement: sources, licences, dates |
| Preparation — 10(2)(c) | “Who annotated the data, against which guideline?” | Labelling guideline plus agreement results |
| Relevance and representativeness — 10(3) | “How does the corpus compare with your population?” | Composition report by segment, region, period |
| Bias examination and mitigation — 10(2)(f)–(g) | “Which biases were examined, and what did mitigation change?” | Bias report: method, thresholds, before and after |
| Data gaps — 10(2)(h) | “What is missing, and how is it handled?” | Gap register with mitigations applied |
| Setting fit — 10(4) | “Was the data drawn from settings like ours?” | Validation on your setting, or a stated limitation |
Keep two files. Provider-side: provenance, composition report, bias report, gap register, change notifications. Deployer-side: your input-field inventory with the relevance reasoning, monitoring notes, log-retention setting and DPIA. Name an owner for each; re-open them on any model or purpose change.
How to check your own position
- Confirm the classification first — tier, Annex III sub-point, date, reviewer. Article 10 bites only on high-risk systems.
- Split provider from deployer: evidence to request, or Article 26(4) to discharge.
- Inventory the input fields you control, one line each: why it is necessary, who supplies it.
- Send the seven-item request: provenance, labelling guideline, composition report, bias report, gap register, test-set design, log export.
- Spot-check outcomes you already have across groups you may lawfully examine; divergence where outputs feed inputs is a point (f) finding.
Frequently asked questions
What does Article 10 of the AI Act actually require?
Article 10 requires training, validation and testing data sets meeting paragraphs 2 to 5: eight governance practices — design choices, collection and origin, preparation, assumptions, availability, bias examination, mitigation and gaps (10(2)(a)–(h)); relevance, representativeness, error minimisation and completeness (10(3)); setting fit (10(4)); strict conditions for special-category bias testing (10(5)); and testing data sets only where a system does not train (10(6)).
Does Article 10 bind us when we only buy the system?
Paragraphs 2 to 5 bind the provider. Yours is Article 26(4): where you control input data, it must be relevant and sufficiently representative for the intended purpose, backed by monitoring under Article 26(5) and six-month log retention under Article 26(6). Article 10(5) is often mis-cited as the deployer duty — it governs providers processing special categories for bias detection.
What evidence should we ask our AI vendor for?
Seven items: a provenance statement with sources, licences and original purpose; the labelling guideline with agreement measurements; a composition report by segment; a bias report with method, thresholds and before-and-after results; a gap register with mitigations; a test-set design note on frozen hold-outs; a log export proving Article 12 capabilities.
When do the Article 10 obligations start applying?
Annex III provider and deployer duties, Article 10 included, apply from 2 December 2027 after the Digital Omnibus deferral; Article 6(1) systems follow on 2 August 2028. Penalties and transparency duties already apply, and buyers request data-governance evidence now — discovering a year out that your vendor cannot answer the seven-item request is expensive.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.
Keep reading
AI in healthcare under the EU AI Act: clinical decision support and high-risk duties
11 min
Risk & ClassificationThe Annex III High-Risk AI Systems List: All Eight Use Cases Explained with Examples
11 min
Risk & ClassificationThe Eight Prohibited AI Practices Under Article 5: A Compliance Checklist
10 min