Data quality and governance (Article 10): training, validation and test data

Short answer: Article 10 of the AI Act states that high-risk systems trained with data must use training, validation and test data that meet quality criteria. The data must be relevant, sufficiently representative, as error-free as possible and complete for the intended purpose. Providers must also apply a form of data governance that makes the choices around that data traceable.

Why data is central

The quality of an AI system is largely determined by the data it was trained on. Biased or incomplete data produces biased or unreliable outcomes — precisely in high-risk applications with consequences for people. Article 10 therefore anchors data quality as a standalone obligation, not a side effect of model design.

The quality criteria

Article 10 requires that datasets, as appropriate for the intended purpose:

• Are relevant and representative of the people and situations the system is applied to.
• Are as error-free as possible and complete, with appropriate statistical properties.
• Take into account the specific geographical, contextual, behavioural or functional setting in which the system is used.

Complete error-freeness is not an absolute requirement — the standard is "as error-free as possible" given the purpose. That forces a reasoned trade-off, not perfection.

Data governance and bias

Beyond the data itself, Article 10 calls for governance practices: documentation of design choices, the provenance and collection of data, assumptions, and an examination of possible bias. Where necessary, providers may process special categories of personal data, solely to detect and correct bias, under strict safeguards. These choices must be traceable — they connect to the broader AI governance framework.

Link to documentation and risk

The data choices under Article 10 belong in the technical documentation (Annex IV) and are input for the risk management system: bias in data is a risk you must identify and mitigate. Together they form part of the high-risk obligations overview.

What to do

• Record the provenance of every dataset: source, collection method and assumptions.
• Test representativeness against the target group and the use setting.
• Examine bias systematically and document the findings.
• Correct where needed and justify the choices, including any processing of special data.
• Keep data governance current when the data or purpose changes.

A model is no more reliable than the data beneath it; Article 10 makes that data testable.