Why assessment is the sensitive point
Training organisations and companies that train their teams increasingly use artificial intelligence. It helps write exercises, rephrase feedback and mark a quiz. One use, however, draws the particular attention of the European legislator: the assessment of learners.
Regulation (EU) 2024/1689 of 13 June 2024, known as the AI Act, lays down harmonised rules on artificial intelligence in the European Union. It classifies certain uses as "high risk". A high-risk AI system is a system subject to the regulation's heaviest requirements, set out in its Chapter III.
The classification of an assessment tool does not depend on its technology. It depends on its intended purpose and on the role it plays in the decision. This guide explains how, with the text in hand.
Provider or deployer: who bears which obligation
The AI Act allocates obligations according to the role of each actor. Two definitions in its Article 3 matter here.
The provider is the person who develops an AI system, or has it developed, and places it on the market or puts it into service under its own name or trademark. This is usually the publisher of the assessment platform.
The deployer is the person who uses an AI system under its own authority, except in the course of a personal non-professional activity. A training organisation that uses a tool bought from a publisher is in principle a deployer.
The distinction has a practical consequence. An organisation that develops its own assessment tool, or distributes it under its own brand, takes on the role of provider. It then bears the documentation required by Article 6(4).
What Annex III, point 3, covers
Annex III of the AI Act lists the areas where an AI system is presumed to be high risk. Its point 3 covers education and vocational training. It targets four families of systems, at all levels of education.
| Annex III point | AI systems covered |
|---|---|
| 3(a) | determining access, admission or assignment of persons to educational and vocational training institutions |
| 3(b) | evaluating learning outcomes, including to steer the learning process |
| 3(c) | assessing the appropriate level of education a person will receive or be able to access |
| 3(d) | monitoring and detecting prohibited behaviour of students during tests |
Point 3(b) is the broadest. It covers systems intended to evaluate learning outcomes, "including when those outcomes are used to steer the learning process". A platform that adapts a learner's path based on their results therefore also falls within its scope.
Point 3(d) covers remote exam proctoring. A tool that analyses a candidate's webcam or screen to detect cheating falls under this point.
The Article 6(3) derogation
Being listed in Annex III is not enough to make a system high risk. Article 6(3) provides for a derogation. A system referred to in Annex III is not high risk where it does not pose a significant risk of harm to the health, safety or fundamental rights of persons.
The text specifies that this absence of risk may result from "not materially influencing the outcome of decision making". The derogation applies where one of the following four conditions is met.
| Condition of Article 6(3) | What the system is intended to do |
|---|---|
| (a) | perform a narrow procedural task |
| (b) | improve the result of a previously completed human activity |
| (c) | detect decision-making patterns or deviations, without replacing or influencing the human assessment without proper human review |
| (d) | perform a preparatory task to an assessment relevant to Annex III |
The logic is the same across all four conditions: AI may prepare, assist or check, as long as a human keeps the decision. The regulation does not look at the tool in isolation. It looks at who decides the mark or the validation of learning outcomes.
The exception with no derogation: profiling
The same paragraph closes a door. An AI system referred to in Annex III "shall always be considered to be high-risk where the AI system performs profiling of natural persons". None of the four conditions can then apply.
To define profiling, the AI Act refers to Article 4 of Regulation (EU) 2016/679, the GDPR. Profiling there means automated processing of personal data that evaluates personal aspects of a person, for example to analyse or predict their performance at work, reliability or behaviour. A tool that segments learners according to supposed traits, or predicts their success, falls within this area.
Reading a concrete use against the text
The cases below illustrate our reading of the regulation. They do not replace the analysis of a specific tool, which depends on the intended purpose declared by the provider and on its actual use.
| Use | Reading under the AI Act |
|---|---|
| The trainer marks, then an assistant rephrases the feedback | falls under condition (b): improving a previous human activity |
| An assistant prepares a grid that the trainer reads, corrects and validates | falls under condition (d), if the human review is real |
| A tool marks a closed-answer quiz on its own | may fall under condition (a), a narrow procedural task |
| A tool awards a mark or validates learning outcomes without human review | likely high risk, under point 3(b) |
| A tool decides on a candidate's admission to a course | likely high risk, under point 3(a) |
| A proctoring tool detects behaviour during an exam | likely high risk, under point 3(d) |
| A tool profiles learners | always high risk |
The decisive point is whether the human review is real. A one-click validation, without reading the learner's work, is not a human review in the sense intended by the text. The derogation assumes that the decision genuinely remains that of a person.
The obligations that come with these rules
For the provider that rules out high risk
Article 6(4) does not leave the derogation without a trace. A provider that considers that a system referred to in Annex III is not high risk documents its assessment before placing it on the market or putting it into service. It provides it to national competent authorities on request.
The provider also remains subject to the registration obligation in Article 49(2). Regulation (EU) 2026/1744 provides for simplifying the content of this registration, without removing it.
The timetable after the 2026 omnibus
Regulation (EU) 2026/1744 of 8 July 2026, known as the digital omnibus on AI, changed the timetable. Sections 1, 2 and 3 of Chapter III, which carry the high-risk rules, apply from 2 December 2027 to systems under Annex III. The original date of application was 2 August 2026.
| Deadline | What applies |
|---|---|
| 2 February 2026 | deadline set for the Commission's guidelines on Article 6 |
| 2 December 2027 | high-risk rules for Annex III systems, including education and training |
| 2 August 2028 | high-risk rules for systems under Annex I |
Article 6(5) requires the Commission to publish guidelines on the implementation of the article. They must come with a list of examples of high-risk and non-high-risk use cases. They are the first document to consult, alongside the text, to classify a tool.
What it means for you
For a training organisation
The period until 2 December 2027 leaves time to map the tools used in assessment. For each one, the useful question is simple: who pronounces the mark or the validation of learning outcomes? If the answer is an identified trainer who reads the learner's work, the tool is likely to fall within the derogation.
An organisation that buys an assessment tool usually acts as a deployer, not as a provider. It is in its interest to ask the provider for its position on classification and for the documentation required by Article 6(4).
For a company that trains its teams in-house
The same rules apply as soon as a tool evaluates learning outcomes in a vocational training setting. An assessment that conditions an internal qualification deserves particular attention, as it may affect the employee's career path.
For an educational software publisher
The declared purpose of the product carries great weight. A tool presented as an aid to the marker is not classified like a tool presented as an autonomous marker. The documentation of the assessment must exist before placing on the market.
What we recommend
Keep the human in the decision. It is both the best teaching practice and the condition for the derogation. AI prepares, the trainer decides.
Rule out any profiling of learners. No derogation covers it: such a use moves the tool into high risk.
Train teams in the use of these tools. AI literacy falls under Article 4 of the AI Act, rewritten in 2026; see our analysis of Article 4 of the AI Act. For the quality framework of training organisations, see our article on the 2026 Qualiopi framework.
Our artificial intelligence training courses build this regulatory framework in from the design of each programme. For an overview of uses in our fields, see our guide to AI in energy engineering offices. To build a programme suited to your teams, see our training courses or write to us.
