EU AI Act and Education: The Annex III(3) Guide for Schools, Universities and EdTech Providers

If your institution or product decides who gets in, how students are graded, or whether they cheated, the EU AI Act probably treats it as high-risk AI. Education is one of the eight areas listed in Annex III, and the list is more specific than most people expect. This guide walks through what is in scope, what is banned, who carries which duties, and when they start.

What Annex III(3) covers

Point 3 of Annex III, titled "Education and vocational training^[1]", lists four types of AI system, each when used in educational and vocational training institutions at all levels:

1. Access and admission. Systems intended to determine access or admission, or to assign people to institutions or programmes.
2. Evaluating learning outcomes. Including where those outcomes are used to steer the learning process. This is the clause that reaches automated grading and some adaptive-learning engines.
3. Assessing the appropriate level of education. Systems that assess the level of education an individual will receive or be able to access.
4. Monitoring prohibited behaviour during tests. In practice, exam proctoring and cheating-detection tools.

The official text is available on the AI Act Service Desk.

What is not high-risk (usually)

Annex III is deliberately narrow. A timetable optimiser, a library search tool, a generic writing assistant or an admissions chatbot that only answers FAQs does not determine access, grade outcomes or police tests. Whether a tool falls inside depends on its intended purpose, not its underlying technology. The Article 6(3) exception can also pull a system out of high-risk where it only performs a narrow procedural task or improves the result of a human activity already completed, but it never applies if the system profiles natural persons. Our high-risk classification guide explains the test.

What is banned: emotion recognition in education

Separately from Annex III, Article 5(1)(f) prohibits AI that infers emotions of people in workplaces and education institutions, except for medical or safety reasons. A proctoring tool that flags "suspicious" behaviour on screen is a high-risk question. A tool that claims to read students' emotions or attention from their faces is a prohibited-practice question, and prohibitions have applied since February 2025. If your product does both, separate the features and check each against the right article.

Who owes what

EdTech vendors are usually providers. They must meet the Chapter III requirements: risk management (Article 9), data governance (Article 10), technical documentation (Article 11), logging, transparency, human oversight (Article 14), accuracy and robustness (Article 15), a quality management system (Article 17) and conformity assessment before CE marking.

Schools and universities are usually deployers. Under Article 26 they must use the system according to its instructions, assign competent human oversight, monitor operation, keep the logs under their control and inform affected people. Public bodies and certain other deployers must also carry out a fundamental rights impact assessment under Article 27 before first use.

Role drift is the trap. A university that substantially modifies a vendor's tool, or puts its own name on it, can become the provider under Article 25. See our procurement guide on Article 25.

When do the duties start?

The Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026) moved the application date for Annex III high-risk obligations to 2 December 2027, with Annex I systems following on 2 August 2028, according to Abreu Advogados' summary of the amendments. The same summary reports that real-world testing outside sandboxes now extends to Annex III systems, which matters for EdTech pilots. Transparency duties under Article 50 and the Article 5 prohibitions are not delayed. See our implementation status post for the full timeline.

Delay is not the same as time saved: conformity assessment capacity is thin (see the notified body bottleneck), and procurement cycles in education are long.

A practical checklist

• Inventory every AI-enabled tool used in admissions, assessment, learning platforms and exam delivery, including those bought by individual faculties.
• Classify each against the four Annex III(3) limbs and against Article 5(1)(f).
• Switch off or replace anything that infers student emotions.
• Ask vendors for their conformity plan, technical documentation summary, logging capabilities and instructions for use before renewing contracts.
• Name human overseers with the authority to override automated grades or integrity flags, and give them training.
• Prepare the FRIA if you are a public body: include affected groups such as students with disabilities and non-native speakers.
• Plan appeals. Students affected by a high-risk decision have a right to an explanation under Article 86; build a route for it.

FAQ

Is a plagiarism or AI-text detector high-risk? If it is used to detect prohibited behaviour during tests, it is within limb 4. Used on coursework outside test conditions, it is a closer call that depends on intended purpose and how its output is used; document your reasoning.

Is adaptive learning high-risk? Only if it evaluates learning outcomes and uses them to steer the learning process in the way limb 2 describes. A content recommender that does not evaluate outcomes is likely outside.

Do non-EU EdTech companies have to comply? Yes, if their system's output is used in the Union. See our extraterritorial scope guide.

This article is general information, not legal advice.