AI Act Article 10: what to ask about training and testing data

Article 10 concerns data and data governance for high-risk AI systems. Before commissioning work, establish which datasets the project uses and what evidence connects them to the intended use. A supplier's accuracy figure alone does not explain who or what was represented in its test.

This guide concerns the material an implementation team can prepare for review. Whether Article 10 applies to your system and how its requirements should be met need a system-specific assessment. Sources were checked on 11 October 2026.

What the provision covers

For systems developed using techniques that train AI models with data, Article 10 addresses training, validation and testing datasets. It includes provenance, preparation, assumptions, suitability, potential bias and gaps. Quality is assessed against the intended purpose and relevant operating context. For high-risk systems developed without those training techniques, paragraph 6 applies the specified requirements to testing data. The amended provision also refers to Article 4a. Article 10

The practical starting point is a dataset inventory. It should distinguish the information used to develop the system from the examples used to evaluate it. Ask who can explain each dataset's origin and permitted use. Questions about personal data belong in the legal and data-protection review as well.

A good aggregate score can hide a missing group of cases

Consider a fictional application assessed on clearly typed forms. The intended workflow also receives incomplete forms and documents written in different styles. A strong result on the first set says little about the missing material.

The buyer should be able to see which cases the evaluation includes and which it leaves out. If the project team has no examples for a relevant situation, the report should name the gap. Filling a table with an invented success rate would make the purchasing decision worse.

For this fictional project, we would ask the process owner to identify the kinds of inputs staff actually encounter. The evaluation plan could then separate clear forms from incomplete ones and record the reviewer responsible for the expected result. This preparation would define the evaluation; the system would still need to pass it.

Evidence that survives a change of dataset

A useful handover explains the source of the examples, the selection criteria and the version assessed. It also records corrections. If a reviewer changes the expected result, the team should be able to establish which earlier scores used the old answer.

This table is an engineering preparation aid. It is not a complete legal checklist, and maintaining it does not establish compliance.

What Syntalith can prepare

Syntalith can scope a data inventory, an evaluation plan and a report separating observed results from untested uses. For the fictional forms project, that could mean documenting the available sample and proposing the missing categories for review before integration begins. The customer supplies appropriate access and someone able to assess the task's correct outcome.

The proposal should identify who provides independent subject expertise and whether obtaining further representative material is included. It should also say what happens when the available data is insufficient. A model test on synthetic examples cannot stand in for that missing evidence.

Describe the workflow and the data you already have, without sending personal records in the first message. We can discuss the technical scope and refer to service pricing. Start with the classification preparation guide if the project's risk assessment is still unresolved.