Is Exam-Scoring AI High-Risk Under the EU AI Act?

Yes — in almost every form that matters to an edtech company. AI that scores exams, grades tests, evaluates learning outcomes, decides who gets admitted, or monitors candidates for cheating during an assessment is high-risk under Annex III point 3 of the EU AI Act. The obligations attach from 2 December 2027. But there is one line you must not cross: an emotion-recognition feature in an educational setting is not high-risk at all — it is prohibited under Article 5(1)(f), with no compliance route.

This page sets out which education uses are high-risk, where the prohibited line sits, and what a private edtech provider actually has to do.

---

What Annex III point 3 covers

Annex III point 3 designates four families of AI use in education and vocational training as high-risk:

• AI intended to determine access or admission, or to assign people, to education and vocational-training institutions at all levels;
• AI intended to evaluate learning outcomes, including where those outcomes steer the learning process a student receives;
• AI intended to assess the appropriate level of education an individual will receive or be able to access;
• AI intended to monitor and detect prohibited behaviour by students during tests — exam proctoring.

If your system does any of these, it is high-risk regardless of how accurate, fair, or well-intentioned it is. The classification is by use, not by the quality of the model. An automated essay grader, an adaptive test that gates which module a learner reaches next, an admissions-scoring engine, and an AI proctoring tool are all squarely inside point 3.

One narrowing test exists. Article 6(3) lets an Annex III system escape high-risk status if it performs only a narrow procedural task, improves the result of a previously completed human activity, detects decision patterns without replacing human judgment, or does preparatory work — and only one of those needs to apply. But there is a hard backstop: any system that profiles a natural person is always high-risk, and most exam-scoring and proctoring tools profile the candidate. In practice, the 6(3) exemption rarely rescues an education-scoring product. Document the assessment either way and register under Article 49.

---

The line you cannot cross: emotion recognition in education is prohibited

This is the most common and most dangerous misclassification in edtech. AI systems to infer the emotions of a natural person in the areas of workplace and education institutions are prohibited under Article 5(1)(f) — not high-risk. The ban has been in force since 2 February 2025, well before the high-risk regime begins.

So a proctoring feature that flags a candidate as "stressed", "anxious" or "likely cheating" by reading facial expression, micro-emotion or affect during an exam is not a high-risk system you can document and deploy. It sits in the prohibited tier, carrying the top penalty — up to €35 million or 7% of total worldwide annual turnover (Article 99(3)). The narrow exception in 5(1)(f) is for medical or safety reasons, which a cheating-detection feature does not meet.

The practical rule: detecting a behaviour (a second person in the room, the candidate leaving the camera) can be a lawful high-risk proctoring function; inferring an emotion to draw a conclusion about the candidate is a prohibited use. Engineer the feature on the right side of that line, or remove it.

---

What a private edtech provider must do (and the FRIA trap)

If you build and sell exam-scoring or proctoring AI under your own name, you are the provider (Article 16). Your duties include a risk-management system (Article 9), data governance (Article 10), technical documentation to Annex IV (Article 11), logging (Article 12), transparency and instructions for use to your customers (Article 13), human-oversight design (Article 14), accuracy and robustness (Article 15), conformity assessment via the Annex VI internal-control route (Article 43), and registration in the EU database (Article 49). The school or exam board that runs your tool is the deployer (Article 26).

Here is the trap worth correcting: a Fundamental Rights Impact Assessment (Article 27) is not automatically triggered for a private edtech provider. The Article 27 FRIA falls on public-body and public-service deployers, plus deployers of creditworthiness (Annex III 5(b)) and life/health-insurance (5(c)) systems. A private edtech company selling a grading tool is none of those. A state university or public exam authority that deploys your tool may owe a FRIA — but that obligation is theirs as a public deployer, not yours as the private provider. Don't market a FRIA as a universal education obligation; it isn't.

---

Worked example: LernPrüf GmbH

LernPrüf GmbH is a Berlin edtech provider, ~60 employees, ~€9M turnover, selling an online assessment suite to language schools and a few public universities. It has three features. (1) An automated scoring engine that grades written and spoken tests — high-risk under Annex III point 3; LernPrüf is the provider and owes the full Article 16 stack. (2) A proctoring module that flags a second face in frame or the candidate leaving the camera — also high-risk point-3 proctoring; lawful with the provider obligations and Article 14 human review of every flag. (3) A planned "stress meter" that reads facial affect to score exam anxiety as a cheating signal — this is prohibited under Article 5(1)(f). LernPrüf cannot ship it; the only fix is to drop the emotion inference and keep the behavioural detection.

On the FRIA: LernPrüf itself does not owe an Article 27 assessment. Its public-university customers, as public deployers, do — and they will ask LernPrüf for the documentation that lets them complete it, which is exactly what Article 13 instructions for use are for.

---

How Confir helps

Confir's classification logic tests an education use against the Article 5 prohibitions before Annex III, so an emotion-recognition proctoring feature is flagged as prohibited rather than mistakenly scoped as "high-risk to document". For genuine point-3 uses it assembles the provider stack — Articles 9 to 15, the Annex IV documentation pack, and the Article 43 conformity route — and tells a deployer whether an Article 27 FRIA applies to them. The engine is deterministic and rule-based: the same intake produces the same finding, each cited to the Article or Annex point that fired.

Related guides

• Annex III point 3: education AI systems
• Student admission AI and the EU AI Act
• Exam monitoring and proctoring AI
• Emotion recognition in education is prohibited
• Article 5 prohibited practices
• How Article 6 classifies high-risk AI