Skip to content

Datasheet for Datasets Template (EU AI Act Art 10)

Template4 November 2026· 6 min read

A datasheet-for-datasets template for EU AI Act Article 10: provenance, representativeness, bias examination 10(2)(f), special-category data under Article 4a.

A datasheet for a dataset is the practice; Article 10 of the EU AI Act, Regulation (EU) 2024/1689, is the duty it documents. "Datasheet for datasets" is a machine-learning convention — a standard record of where a dataset came from, how it was collected, what it contains, and where it falls short. The Act never uses the phrase, but it requires almost exactly that information for high-risk systems: Article 10 sets the data-governance criteria, and Annex IV requires the technical documentation to describe the data used and how it was obtained, selected and labelled. A datasheet is how a provider turns "we used good data" into something an auditor can check. It is a provider artefact — Article 10 sits with whoever develops the high-risk system.

A dataset with no datasheet is a black box. When a regulator, a deployer, or your own future engineer asks "is this data representative, and how do you know?", the datasheet is the answer — or its absence is the finding.


What Article 10 requires of your data

Article 10 requires that training, validation and testing datasets for high-risk systems be subject to appropriate data-governance and management practices. Among them:

  • Provenance and collection — the design choices, data origin, and the collection process; for personal data, the original purpose of collection.
  • Preparation — annotation, labelling, cleaning, updating, enrichment and aggregation operations.
  • Representativeness (Article 10(3)) — datasets must be relevant, sufficiently representative, and to the best extent possible free of errors and complete in view of the intended purpose; and consider the characteristics of the geographical, contextual, behavioural or functional setting in which the system will be used.
  • Bias examination (Article 10(2)(f)) — examination in view of possible biases likely to affect health and safety, negatively impact fundamental rights, or lead to discrimination prohibited under Union law.
  • Bias mitigation (Article 10(2)(g)) — appropriate measures to detect, prevent and mitigate identified biases.

The datasheet is structured to produce a written answer for each of these.


The special-category-data allowance (Article 4a)

You cannot examine a dataset for, say, ethnic-origin bias without processing data about ethnic origin — which runs into GDPR Article 9's restriction on special-category data. The AI Act resolves this narrowly. Since the Digital Omnibus (Regulation (EU) 2026/1744, in force 27 July 2026) the basis is Article 4a, which replaced the deleted Article 10(5) and extends the allowance beyond high-risk providers to other AI systems and models. It permits processing special-category personal data to the extent strictly necessary to detect and correct bias — it creates no duty to do so — but only with safeguards: the bias must not be detectable by processing other (including synthetic or anonymised) data; reuse is technically limited; state-of-the-art security and privacy-preserving measures apply; access is strictly controlled and documented; and the data is not transmitted to third parties and is deleted once the bias is corrected or the retention period ends. The datasheet records the why (no less-intrusive route worked), the safeguards, and the deletion — so the use is defensible, not assumed.


Provenance, representativeness, gaps — the honest section

The most useful part of a datasheet is the part that admits what is missing. A provider that writes "fully representative" with no analysis has documented nothing; a provider that writes "under-represents applicants over 60 and rural postcodes; mitigated by targeted collection; residual gap noted" has produced exactly the evidence Article 10(3) asks for.

Datasheet sectionArticle 10 / Annex IV linkThe honest question to answer
ProvenanceData origin, collection processWhere did every record come from, and is the source lawful?
RepresentativenessArticle 10(3)Of which population — and where does the dataset fall short?
Gaps & errorsArticle 10(3)What is missing, mislabelled or stale, and what did you do?
Bias examinationArticle 10(2)(f)Which groups, which metric, what disparity?
Special-category useArticle 4aWas it strictly necessary, and were the safeguards applied?
LabellingAnnex IV (preparation)Who labelled it, by what scheme, with what quality control?

The template — fields to capture

1. Dataset identification — name, version, the system and intended purpose it feeds; whether training, validation or testing; the Annex III category.

2. Provenance and collection — sources, collection method and dates, lawful basis for any personal data, third-party data terms.

3. Composition — size, features, classes, time span, languages/geographies covered.

4. Representativeness (Article 10(3)) — the target population; representativeness analysis; known under-representation and its effect.

5. Gaps, errors and completeness — missing data, label quality, error rate, staleness; remediation.

6. Bias examination (Article 10(2)(f)) and mitigation (10(2)(g)) — protected attributes examined; metrics and disparities; measures applied; residual disparity. Cross-reference your bias-assessment record.

7. Special-category data (Article 4a) — what was processed, the strict-necessity justification, the safeguards, and the deletion record; the GDPR Article 9 / Article 35 linkage.

8. Labelling and preparation — annotation scheme, labellers, quality control, preprocessing steps.

9. Maintenance — update cadence, versioning, and the trigger to re-examine (new data, drift, retrain). The completed sheet feeds the Article 11 / Annex IV documentation.


Worked example: a provider of a tenancy-scoring model

Wohnscore GmbH, a provider with about 45 employees and roughly €8M turnover, develops a high-risk system that scores rental applications (Annex III, access to essential services). Its datasheet records provenance — three years of anonymised tenancy records from partner letting agencies, with the lawful basis stated. The representativeness section, under Article 10(3), admits the dataset over-represents one metropolitan region and under-represents rural applicants, and documents targeted collection to close the gap. To run the Article 10(2)(f) bias examination across national origin, Wohnscore relies on Article 4a to process the protected attribute strictly for bias correction, restricts access to two analysts, records that synthetic data could not surface the disparity, and deletes the attribute once correction is verified. The labelling section names the annotation scheme and its quality control. The whole datasheet drops into the Annex IV technical documentation, with a GDPR Article 35 DPIA covering the personal-data side. The dataset is no longer a black box — it is a documented, defensible input.


How Confir helps

Confir's data-and-technical-robustness assessment walks an Article 10 dataset through plain-English questions — provenance, the Article 10(3) representativeness analysis, gaps and errors, the Article 10(2)(f) bias examination, and the Article 4a special-category-data conditions — and records the answers as structured evidence that assembles into the Article 11 / Annex IV pack. It cross-maps the data-governance controls to GDPR so the AI Act and data-protection records reference each other. The engine is deterministic and rule-based: the same dataset facts produce the same finding every time.


Manage your EU AI Act compliance in one place

Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.

Start free trial →