The Act lands on a sector that has deployed AI rapidly, largely without governance frameworks, and in contexts involving the most structurally vulnerable users that exist in any legal system: children. The result is one of the Act's most operationally consequential applications — and one of its least understood.
Education appears explicitly in Annex III of the Act as a high-risk category, which means certain AI uses in schools and universities carry the same tier of regulatory obligation as diagnostic algorithms in hospitals. How institutions, vendors, and regulators navigate that reality will shape not only compliance costs but the future architecture of AI in European classrooms.
Actual prohibitions
Before engaging with high-risk compliance, the Act's outright prohibitions — which became enforceable on 2 February 2025 — establish hard limits that apply regardless of any risk assessment or conformity process.
The most unequivocal of these is the prohibition on emotion recognition in educational contexts. Any AI system that attempts to infer students' emotional states through facial analysis, voice monitoring, or keystroke patterns is classified as unacceptable risk and is banned. The prohibition is not conditional on harm being demonstrated; the Act treats emotion inference in classrooms as inherently incompatible with student dignity and the conditions of free learning. Several EdTech products that reached the European market in the early 2020s — tools that claimed to measure student engagement or attention by analysing webcam footage — are, under this framework, unlawful.
The prohibition on social scoring is equally applicable to education. An AI system that evaluates students across multiple behavioural dimensions over time and uses those scores to determine access to resources, opportunities, or treatment within or across educational contexts is prohibited. The Act's language — "evaluation or classification of natural persons over a certain period of time based on their social behaviour or personal characteristics" — captures precisely the kind of longitudinal student profiling that some learning management systems have quietly built into their analytics layers.
The Act also prohibits AI systems that exploit the vulnerabilities of persons due to their age, with the explicit effect of materially distorting their behaviour in a manner that causes harm. Children, as a legally recognised vulnerable group, receive heightened protection under this provision. An AI-powered learning app that uses variable reward schedules, social pressure mechanics, or personalisation designed not to improve learning outcomes but to maximise time-on-platform at the expense of the child's wellbeing falls within this prohibition — the same logic applied to manipulative AI in consumer contexts applies with additional force in education.
The high-risk category
Annex III lists four specific education use cases that qualify AI systems as high-risk, all carrying full compliance obligations from August 2026.
-
The first — and most consequential in scope — covers AI systems used to determine access to or admission into educational institutions at any level. This covers university admissions algorithms, automated screening of school enrolment applications, and any AI system that influences whether or not a student gains access to a programme.
-
The second covers AI systems used to evaluate learning outcomes, including when those outcomes are used to steer the learning process. This captures automated grading systems, adaptive learning platforms that adjust curriculum based on assessment data, and any system whose outputs directly shape what a student is taught next.
-
The third covers AI systems that assess the appropriate level of education a student should receive — placement testing, streaming decisions, differentiated learning pathway assignments.
-
The fourth covers AI systems used to monitor and detect prohibited student behaviour during tests: in other words, automated proctoring.
Each of these high-risk use cases demands the same core compliance stack: a quality management system, risk management documentation, data governance frameworks, technical documentation including training data specifications, human oversight mechanisms with documented authority to override AI outputs, bias and discrimination monitoring, post-deployment incident reporting, and registration in the EU AI database. Public educational institutions deploying these systems must additionally complete a Fundamental Rights Impact Assessment under Article 27 of the Act.
EdTech vendors
For companies building AI products for European schools and universities, the Act introduces obligations that are both more specific and more onerous than most EdTech vendors anticipated when they entered the market.
Compliance costs for an EdTech product in the high-risk category — an AI admissions screening tool, an automated grading platform, an adaptive learning system — run from €250,000 to €350,000 in initial implementation, with €80,000 to €150,000 annually thereafter. The largest single cost driver is the quality management system, which must document not only how the AI functions but how it is developed, tested, updated, and monitored across its deployment lifecycle. Bias detection and fairness testing — which in the education context means demonstrating that the system does not produce systematically different outcomes for students by race, gender, socioeconomic background, or disability status — adds €35,000 to €50,000 to the compliance build. Parental consent mechanisms, student data protection governance (which must satisfy both the Act and the GDPR, with additional constraints applicable to minors under 16), and explainability infrastructure account for a further €25,000 to €35,000.
These are not one-time investments. Post-market monitoring — continuous tracking of real-world performance against the fairness, accuracy, and bias standards established at conformity assessment — costs €40,000 to €80,000 per year. For a startup with 30 people building an AI-powered assessment tool, this is a structural cost that does not diminish as the company grows; it scales with deployment.
The DIGITALEUROPE estimate, published in March 2026, placed the aggregate annual cost of EU AI Act compliance for European technology companies at €3.3 billion per year — with initial implementation costs between €320,000 and €600,000 for any small company whose product falls into a high-risk Annex III category. EdTech falls directly within that range.
Universities as deployers
Universities occupy a dual role under the Act. As institutions that develop AI tools — building automated grading systems, student support chatbots, or research models — they can be providers with full provider obligations when those systems are deployed in live settings. As institutions that procure and use commercially developed AI tools, they are deployers, with a distinct but substantial set of obligations.
The AI literacy requirement under Article 4 of the Act became enforceable on 2 February 2025. It requires all providers and deployers of AI systems to ensure that their staff — anyone operating, overseeing, or making decisions based on AI systems — has a sufficient level of AI literacy, calibrated to the role and the risk level of the systems in use. For a university, this means not merely informing staff that AI tools exist, but ensuring that the people responsible for admissions, assessment, and student support have the operational competence to understand the AI systems they deploy and the authority to override them when necessary. The obligation is active and ongoing; it is not satisfied by a single training session.
For high-risk AI systems deployed by universities, Article 26 deployer obligations apply from August 2026. These include: using AI systems only within their designated intended purpose; assigning human oversight to qualified, trained staff with documented authority to intervene; maintaining automatically generated system logs for a minimum of six months; suspending use and notifying the provider when risks to student rights are identified; and registering usage in the EU database where required for public body deployers. The deployer obligation cannot be discharged by pointing to the vendor's conformity assessment. If a university deploys an AI proctoring system in a manner inconsistent with the provider's instructions — for instance, by using it in contexts for which it was not validated — the university bears independent liability as deployer.
The proctoring question illustrates this concretely. The University of Amsterdam's dispute with students over the Proctorio surveillance software — resolved under GDPR before the Act's obligations became applicable — would today be analysed against a different and more demanding standard. Under the Act, remote proctoring using biometric monitoring is a high-risk AI system. Its deployment requires a documented risk assessment, human review of AI flags before any academic consequence is applied, student notification prior to use, and compliance with GDPR's heightened requirements for biometric data.
The research exemption
Universities that develop AI systems are, under certain conditions, exempt from the Act's requirements while that development remains within a genuine research context. Article 2(8) provides that AI systems developed exclusively for scientific research and development, before being placed on the market or put into service, fall outside the Act's scope. The purpose of the exemption is to protect academic freedom and not obstruct fundamental research.
The exemption's edges are sharper than the text suggests. It requires both elements simultaneously: the AI system must be developed and deployed exclusively for research purposes. A university that develops an automated examination-marking system in its computer science department and then deploys it for actual student examinations — even within the university, even without commercialisation — has moved outside the research exemption and activated the Act's high-risk obligations. The system is "put into service" the moment it operates on real students in a real assessment context. The test environment and the operational environment are not the same, and the Act treats their distinction as legally significant.
A June 2025 analysis from researchers at the University of Tubingen warned that the AI research community substantially underestimates how many common research practices trigger Act obligations — particularly at the boundary between experimental AI development and real-world deployment in research settings that involve human participants. The research exception also does not apply to prohibited practices: even a research institution cannot deploy an emotion inference system on students under the guise of research, because the prohibition in Article 5 contains no research carve-out.
GPAI in the classroom
The proliferation of large language models in educational settings — students using ChatGPT to draft essays, faculty using AI writing assistants, universities deploying institution-wide AI tools — creates a distinct regulatory question that the Act addresses through its general-purpose AI (GPAI) framework rather than through the high-risk classification.
A student or professor using a commercially available GPAI model for drafting, research assistance, or document summarisation is not operating a high-risk AI system under the Act. The classification turns on intended purpose: a commercial LLM has a general-purpose intent, not an education-specific one. However, the Act's Article 50 transparency obligations apply from August 2026 to all deployers of generative AI systems: any institution deploying an AI chatbot or assistant for student-facing services must inform users that they are interacting with an AI system. Outputs generated by AI systems must be machine-readable and technically detectable as artificially generated.
The GPAI question becomes more complex when a university builds a custom AI tool on top of a foundation model — for instance, deploying an AI tutoring assistant that uses a fine-tuned LLM to answer student questions about curriculum content. If that tool is deployed in a manner that evaluates students, steers their learning pathway, or functions as a de facto teaching authority, the question of whether it crosses into high-risk territory depends on the degree to which its outputs materially influence decision-making about individual students. A chatbot that answers factual questions is not the same as one that monitors engagement and flags students for intervention. The former requires transparency disclosures; the latter may require a full conformity assessment.
The student as subject
Students are not merely incidental beneficiaries of the Act — they are its primary intended protected class in the education context, and the regulation creates specific rights that affect how AI is used around them.
Article 86 provides that students subject to decisions made on the basis of high-risk AI systems — including admissions decisions, learning pathway assignments, and graded assessments where AI played a material role — have the right to obtain a clear and meaningful explanation of the AI's role in that decision. Before the Act, a student rejected by a university whose admissions algorithm found them unsuitable had no formal legal mechanism to understand why. That changes with August 2026 enforcement. The right is exercisable when a decision produces significant effects on the student's educational trajectory, and the deployer bears the obligation to provide the explanation, not the system's developer.
What constitutes a meaningful explanation in educational practice is not yet established. A technically accurate saliency map of the features that contributed to a rejected admissions decision is not the same as an explanation accessible to an 18-year-old applicant without technical training. The Act's requirement is for explanations that are "clear and meaningful" — language that sets a qualitative standard without defining it, and that courts and national competent authorities will eventually interpret through enforcement decisions.
The prohibition on exploiting children's vulnerabilities creates obligations that extend well beyond the classroom into the consumer EdTech market that children access from home. An AI tutoring app that targets children, uses personalisation designed to maximise engagement at the expense of learning quality, or collects and monetises data about children's learning difficulties is subject to the Act when operating in the EU — regardless of where the company is incorporated.
Academic integrity
One area where the Act creates obligations that institutions are only beginning to process concerns the detectability of AI-generated academic content. Article 50 of the Act requires that AI-generated outputs be machine-readable and technically identifiable as artificially generated from August 2026. The intent is to protect information integrity at the societal level. In practice, for universities, it creates both a tool and a dilemma.
The AI detection tools currently available — Turnitin's AI detection module, GPTZero, Copyleaks — operate with meaningful false positive rates, identifying human-written text as AI-generated and vice versa. The European University Institute has explicitly declined to deploy them, citing the risk of unjust academic integrity determinations based on unreliable detection. If Article 50's watermarking and technical detectability requirements produce effective, standardised, machine-readable signals in AI-generated text, that technical infrastructure could substantially improve detection accuracy — reducing institutions' dependence on probabilistic classifiers and their attendant false positive burden. Whether the technical standards required for implementation will be available by August 2026 remains uncertain.
Academic integrity is, in this sense, an area where the Act's long-term effect may be transformative but its short-term implementation is incomplete. Universities are navigating a transition period in which AI-generated content is ubiquitous, detection is unreliable, legal obligations to ensure detectability are approaching, and the technical standards required to fulfill those obligations do not yet exist.
The structural tension
Education presents the AI Act with a challenge that differs structurally from healthcare or defence. In healthcare, AI regulation builds on decades of medical device law; there are existing conformity assessment pathways, established notified body infrastructure, and regulatory cultures accustomed to high-stakes compliance. In defence, the exemption — however imperfect — removes most military AI from the Act's immediate scope. In education, neither condition applies.
European schools and universities are asked to comply with a high-risk regulatory framework while operating on public funding, without dedicated compliance infrastructure, without established guidance on how to fulfil obligations like the right to explanation or the FRIA for educational AI, and while facing a growing commercial AI product ecosystem that is developing faster than any regulatory body can assess. The Commission published updated ethical guidelines for AI in education in March 2026, reinforcing obligations around transparency and data protection. National-level policy coherence is uneven: Ireland published a national framework for generative AI in higher education in late 2025; France committed €20 million to AI tools for teachers with a national plan anticipated in 2026; Germany invested €20 million over four years in digital education infrastructure. These are not equivalent levels of preparation, and the Act's enforcement will not wait for institutional readiness to converge.
The deeper question is one that no regulation resolves by itself: what AI should be used for in education, and what it should not. The Act addresses the legal architecture of that question — who is responsible when AI influences educational decisions, what rights students have, what is prohibited — but it cannot substitute for a pedagogical and institutional consensus about the role of AI in teaching and learning that European education systems are still in the early stages of forming.
References
- EU AI Act, Annex III, Point 3: Education and Vocational Training. Official text, August 2024.
- Klarvo — EU AI Act Compliance Platform. "EU AI Act for Education & EdTech: Compliance Guide." January 2025.
- EY Tax Alerts. "AI Act: Prohibited AI Practices Become Applicable." December 2025.
- Higher Education Authority (Ireland). "Generative AI in Higher Education Teaching & Learning Alignment." December 2025.
- FeedbackFruits. "What is the EU AI Act? A Comprehensive Overview." February 2025.
- SwissCyberInstitute. "The EU AI Act: Implications for Ethical AI in Education." February 2026.
- SoftwareSeni. "Budgeting for EU AI Act Compliance: Cost Models for SMB Tech Companies." January 2026.
- The Parliament Magazine / DIGITALEUROPE. "€31bn Cost on Europe's Innovators: Why the AI Act Is Backfiring." March 2026.
- Rockefeller Institute of Government. "The European AI Act and Its Implications for New York State Higher Education." November 2025.
- Harvard University IT. "EU AI Act Prohibited Use Cases." February 2025.
- KPMG Law Germany. "AI Act: This Applies to AI in Universities and Research." March 2025.
- arXiv. Wernick, A., Meding, K. "Beware! The AI Act Can Also Apply to Your AI Research Practices." June 2025.
- Health Management. "EU AI Act Research Exemptions Under Scrutiny." March 2026.
- Academic Journals. Oxford GRURINT. "In Search of the Lost Research Exemption: Reflections on the AI Act." October 2025.
- Global Centre for Human Rights. "How is ChatGPT Regulated by the EU AI Act: Reflections on Higher Education." September 2024.
- AI Act Explorer. Article 50: Transparency Obligations.
- DDG Avocats (French Digital Law). "Transparency of AI-Generated Content: In-depth Legal Analysis of the Draft Code of Practice." April 2026.
- EUI (European University Institute). "Guidelines for the Responsible Use of Artificial Intelligence for Research." 2024.
- DAAD Brussels. "Explainable AI in Education: Fostering Human Oversight and Shared Understanding." 2025.
- Humane AI / University of Amsterdam. "UvA Students Against Online Proctoring." 2021–2024.
- EU AI Act, Article 5: Prohibited AI Practices; AI Act Service Desk guidance.
- European Commission. "Guidelines on the Ethical Use of Artificial Intelligence and Data in Teaching and Learning." March 2026.
- LinkedIn / Vlad Petcu. "Higher Education AI Policy Developments Across Europe." February 2026.
- Future Market Insights. "AI in K-12 Education Market Analysis 2025–2035." September 2025.
- Latham & Watkins. "EU AI Act: Ban on Certain AI Practices and Requirements for AI Literacy." January 2025.