Is AI Course Evaluation High-Risk Under the EU AI Act?

Bottom line up front: Using AI to analyse student feedback about a course or instructor is, in most designs, not a "high-risk" system under Annex III of the EU AI Act, because that category targets AI that assesses students — admissions, grading, proctoring — not AI that summarises what students say about teaching. But "not Annex III high-risk" is not the same as "unregulated." Transparency duties, the GDPR, and the Act's general principles still apply, and a handful of design choices can push a feedback tool over the line. This piece reads the actual text of Regulation (EU) 2024/1689 so quality-assurance teams can reason about it rather than guess.

What the Act actually says about education

The EU AI Act takes a risk-tiered approach: a small set of prohibited practices, a defined list of high-risk systems with heavy obligations, limited-risk systems with transparency duties, and everything else as minimal-risk. The high-risk list lives in Annex III, which names eight domains. Education and vocational training is point 3, and it is specific. It covers AI systems intended to be used:

(a) to determine access or admission to educational institutions;
(b) to evaluate learning outcomes, including when those outcomes steer the student's learning process;
(c) to assess the appropriate level of education an individual will receive or can access;
(d) to monitor and detect prohibited behaviour during tests (proctoring).

Read those four carefully. Every one of them is about an AI system acting on a student — deciding their admission, scoring their work, routing their education, or surveilling their exam. The throughline, made explicit in the Act's recitals, is that these systems "may determine the educational and professional course of a person's life" and therefore affect fundamental rights, equity, and access to opportunity. A 2025 governance study testing Annex III against Italian universities frames the relevant case precisely as AI-based learning-outcome assessment — algorithms that grade or place students (Frontiers in Artificial Intelligence, 2025).

A course-evaluation tool does the opposite. It does not assess the student; it gathers and analyses the student's assessment of the teaching. The "subject" being evaluated is the course, the module, or — more sensitively — the instructor. On a plain reading of Annex III point 3, summarising open-text comments into themes, or scoring response quality, falls outside the four listed uses.

Where it gets subtle — and where you can cross the line

Three caveats keep this from being a simple "you're fine."

1. Profiling the instructor is its own risk. Annex III point 4 covers AI used in employment for evaluating candidates or making decisions affecting work relationships — promotion, task allocation, termination. If an institution feeds AI-summarised student feedback into automated or heavily AI-influenced personnel decisions about teaching staff, the staff-facing use can implicate the employment category, independent of the education one. The risk migrates from the student to the lecturer. This is a strong argument for keeping feedback analytics formative and advisory, with human judgement firmly in the loop on any tenure, promotion, or contract decision.

2. The GDPR never went away. Even where the AI Act does not classify a feedback tool as high-risk, Article 22 of the GDPR restricts decisions "based solely on automated processing" that produce legal or similarly significant effects on a person, and grants a right to human intervention, to express a view, and to contest. Open-text comments are personal data — sometimes special-category data, when students disclose health or beliefs. A defensible programme needs a lawful basis, data-protection-by-design, and usually a Data Protection Impact Assessment.

3. Transparency obligations apply regardless of tier. Article 50 of the Act requires that people be told when they are interacting with an AI system unless it is obvious. An AI-moderated feedback interview must make clear to the student that they are talking to an AI, not a human reading in real time.

The timeline — and the 2025 twist

Institutions planning budgets should know the dates have moved. High-risk Annex III obligations were originally set to apply from 2 August 2026 (implementation timeline). In late 2025, EU institutions reached political agreement on a Digital Omnibus that postpones the applicability of high-risk obligations for stand-alone Annex III systems to 2 December 2027 (with product-embedded Annex I systems pushed to 2028), as summarised by Gibson Dunn. Crucially, the Article 50 transparency rules remain on the original 2 August 2026 track. The postponement is provisional until formally adopted and published in the Official Journal, so treat it as a planning assumption, not settled law. The prohibited-practices and general-purpose-AI provisions already applied earlier in 2025.

But doesn't this mean AI feedback tools are basically unregulated?

This is the strongest objection, and it deserves a direct answer: no. The argument that "if it is not Annex III high-risk, anything goes" misreads how the layers stack. A feedback tool can sit outside Annex III and still be bound by GDPR lawful-basis and Article 22 limits, by Article 50 transparency, by national education-data rules, and — when the analysis informs staff decisions — by the employment provisions and by ordinary employment law. The honest framing is not "regulated vs unregulated" but "which obligations bite, and how heavily." A responsible vendor should be able to tell you exactly which apply to their product and show the controls, not wave the question away.

There is also a fundamental-rights case for over-compliance. Even where the law does not mandate human oversight, the consequences of getting teaching feedback wrong — misreading a comment, amplifying a biased pattern, exposing a student's identity — are reasons to build oversight in voluntarily. Annex III draws a legal line; it does not exhaust the ethical one.

A short checklist for quality-assurance teams

Before adopting any AI feature in course evaluation, a QA team can reason about its regulatory footprint with five questions:

Who is the subject? Does the system assess students (admission, grading, placement, proctoring) or teaching? Only the former lands squarely in Annex III point 3.
Does it decide, or does it inform? Is any consequential decision — for a student or a staff member — made solely by the system, or does a human review and own it? Solely-automated significant decisions trigger GDPR Article 22.
Does it touch staff outcomes? If feedback analytics feed promotion, contract, or workload decisions, weigh the employment provisions and employment law, not just the education ones.
Is the interaction disclosed? Article 50 transparency applies regardless of risk tier — students should know they are talking to an AI.
What is the lawful basis, and is there a DPIA? Open-text feedback is personal data, and sometimes special-category data.

Answer those honestly and the classification usually resolves itself — and the answers double as the documentation an auditor will later ask for.

Where Koji fits

Koji for Education was built for the European quality-assurance context, and the regulatory reading above is exactly why its design choices matter:

It analyses teaching, not students. Koji summarises and thematically codes student feedback about courses; it does not grade students, gate admissions, or proctor exams — the uses Annex III point 3 actually targets.
Transparency by default. Students are told they are completing an AI-moderated interview, satisfying the spirit of Article 50.
Human-in-the-loop reporting. Programme- and institution-level reports are decision-support for committees and teaching-and-learning centres, not automated verdicts. That keeps high-stakes staff decisions out of "solely automated" territory under GDPR Article 22 — and it is why we describe Koji as surfacing evidence, never deciding.
GDPR/AVG-aware data handling. EU-appropriate processing, data-protection-by-design, and support for DPIAs are first-class, because open-text feedback is personal data.
Bias-aware, standardised moderation. A single consistent AI moderator reduces the human-moderator variability that itself raises fairness questions.

Teams that also run general user, customer, or staff research can use the same AI interview engine on the main Koji platform, keeping one compliant approach across institutional and operational research.

The EU AI Act is not a reason to avoid AI in course evaluation. It is a reason to choose tools that know exactly where the legal lines are — and stay on the right side of them. If you want to see how bias-aware, GDPR-ready conversational evaluation works in practice, explore Koji for Education.