EU AI Act for Schools: What “High-Risk” Means and What to Do About It
Your board wants AI in classrooms this academic year. Your DPO is uneasy. The EU AI Act has been in force since August 2024, and parts of it have been binding since 2 February 2025. Two of those parts already apply to most primary and secondary schools, even though most school leaders I speak with are still treating the Act as a 2027 problem.
This piece is what you actually need to know. I work with primary and secondary schools (sometimes called K-12 in North American materials) on EU AI Act compliance and built Lumination’s free risk assessment tool. What follows is the obligations that already bind you, the high-risk uses you need to recognise, twelve real tools classified against the Act, and a seven-step plan for this academic year.
What’s Already Binding (and What’s Coming)
Two parts of the EU AI Act have bound schools since 2 February 2025. The high-risk obligations that cover most education uses follow on 2 December 2027.
| Date | What kicks in | What it means for a school |
|---|---|---|
| 2 Feb 2025 | Article 4 (AI literacy) + Article 5 prohibited practices, including emotion recognition in education | Already enforceable. If a tool you use infers emotion from pupils’ faces or voices, you are non-compliant today. If staff use AI, the school must take measures to support their AI literacy, in line with their experience and the context of use. The Act does not require a guaranteed level of literacy for each person. |
| 2 Aug 2025 | General-purpose AI model obligations | Mostly affects vendors. Indirect impact: vendors will start passing compliance documentation through to you. |
| 2 Dec 2027 | High-risk AI (Annex III) obligations, including all education uses | Set by the Digital Omnibus in Article 113 of the Act. If your school uses AI for admissions, grading, level assessment or exam monitoring, those uses carry the full high-risk obligations from this date. Public schools also face an Article 27 Fundamental Rights Impact Assessment. |
| 2 Dec 2026 | New prohibitions on AI systems that generate child sexual abuse material or non-consensual intimate images, and machine-readable marking of AI-generated content (Article 50(2)) for generative systems placed on the market before 2 Aug 2026 | Added by the Digital Omnibus. The marking rule covers AI tools your school already uses, and it falls on their providers. |
| 2 Aug 2028 | Embedded high-risk AI (Annex I, products) | Penalty regime: up to €35 million or 7% of global turnover for prohibited-practice violations. |
The literacy obligation is the one most often missed. Article 4 has applied since February 2025, and the Digital Omnibus kept it as a duty on every organisation that deploys AI, schools included. The European Commission explains what it expects in its AI Literacy Q&A. We covered the framework options for getting staff there in our AI Literacy Frameworks Compared article.
The Four High-Risk Uses You Need to Recognise (Annex III, Point 3)
Annex III, point 3 of the Act covers education and vocational training. It lists four uses, each describing a different kind of decision an AI system might make about a pupil. If any AI use in your school fits one of these, it is classified as high-risk and triggers the high-risk obligations, which apply from 2 December 2027.
3(a) Determining access or admission
“AI systems intended to be used to determine access or admission or to assign natural persons to educational and vocational training institutions at all levels.”
Plain English: AI that picks who gets in. Most common in selective independent schools, grammar-school streams, and competitive intake schemes. Rare in standard state school admissions, but worth checking whether any algorithmic ranking is used.
3(b) Evaluating learning outcomes
“AI systems intended to be used to evaluate learning outcomes, including when those outcomes are used to steer the learning process of natural persons in educational and vocational training institutions at all levels.”
This is the widest category. It covers two things: AI that scores work and AI that adjusts what a pupil studies next based on assessed performance. Automated essay grading is high-risk. So is any adaptive learning platform that materially steers a pupil’s pathway. Before buying one, it also pays to check which AI tutor designs actually help pupils learn.
3(c) Assessing appropriate level of education
“AI systems intended to be used for the purpose of assessing the appropriate level of education that an individual will receive or will be able to access.”
Plain English: AI that decides which class, set, or programme a pupil goes into. Includes AI-suggested ability grouping or setting-decision support tools.
3(d) Monitoring prohibited behaviour during tests
“AI systems intended to be used for monitoring and detecting prohibited behaviour of students during tests in the context of or within educational and vocational training institutions at all levels.”
Plain English: AI exam proctoring. Covers any AI-monitored online exam tool that flags behaviour. Watch this category closely because some proctoring products also trigger Article 5(1)(f) emotion recognition rules.
12 AI Tools Used in Schools, Classified Against the Act
Most coverage of the EU AI Act for education stays abstract. The table below classifies twelve tools commonly used in primary and secondary schools today, with the specific provision driving each row. This is a working reference for triage, not legal advice.
| Tool | Typical school use | Most likely classification | Provision driving it |
|---|---|---|---|
| Classroom “engagement analytics” using facial expression analysis | Cameras inferring pupil emotion or attention | Prohibited | Article 5(1)(f). Repackaging as “engagement” does not change classification. |
| Gradescope (auto-grading), any automated essay scoring with determinative output | AI assigns or strongly suggests final marks | High risk | Annex III §3(b). Triggers Article 27 FRIA for state schools. |
| ProctorIO, Honorlock, Examus, Mercer Mettl proctoring | AI behavioural monitoring during online exams | High risk + possibly prohibited | Annex III §3(d) + biometric data rules. Article 5(1)(f) if features analyse emotion or “stress indicators”. |
| DreamBox, Century Tech, Squirrel AI adaptive learning | Adaptive engine that steers pupil learning path based on assessed performance | High risk | Annex III §3(b), where the system materially influences learning outcomes. |
| AI for admissions screening or setting allocation | Algorithmic ranking of applicants or pupils | High risk | Annex III §3(a) or §3(c). |
| ChatGPT, Claude, Gemini, Microsoft Copilot (in class) | General-purpose chat used in lessons; pupil queries; teacher drafting | Limited risk | Article 50(1) transparency. Becomes high-risk if used to determine final grades. |
| Khanmigo (Khan Academy) | AI tutor in maths and other subjects, mastery-based progression | Limited risk | Article 50(1). Could become high-risk if used to gate progression decisions. |
| Google Classroom AI Practice Sets, Microsoft Reading Coach | Reading practice, formative feedback, suggestions | Limited risk | Article 50 transparency. |
| MagicSchool, Eduaide, Diffit, Curipod | Teacher-side lesson planning, rubric drafting, level-adjusted texts, slide generation | Minimal risk | Not in Annex III. Article 6(3)(d) preparatory-task framing applies. |
| Brisk Teaching | Feedback drafting, lesson generation, draft scoring (teacher reviews) | Case by case | Article 6(3) exception if teacher independently assesses; Annex III §3(b) if scores become final. |
| Turnitin AI Detection (Originality) | Plagiarism scoring flagged after teacher review | Debatable | Article 6(3)(c) exempt if used as a flag; high-risk under §3(b) if it determines pass or fail. |
| Administrative AI (timetabling, room booking) | Operations only, no pupil-decision impact | Minimal risk | Not in Annex III. |
The clearest pattern is that brand does not determine classification, use does. The same MagicSchool deployment is minimal-risk for a teacher writing rubrics and high-risk if a school uses it to score pupil work without teacher review.
Audit by use case, not by tool name. Our free EU AI Act risk assessment tool walks through this same logic for any tool you want to check (linked below in the checklist).
Three Myths That Get Schools in Trouble
When I work with schools on compliance, three sentences come up repeatedly. Each one is wrong in a specific way.
“It’s just a chatbot, so it’s minimal risk.”
Limited risk, not minimal. Article 50(1) requires you to tell pupils they are interacting with AI. The fix is a one-line classroom disclosure, but ignoring it is still a compliance gap.
“The teacher reviews everything, so we’re fine.”
Sometimes true. Article 6(3)(d) exempts AI that performs a preparatory task before a separate human assessment. But if the teacher rubber-stamps the AI’s score (changes it <5% of the time, signs it without reading), the exception does not apply.
“We disabled the high-risk feature, so we’re fine.”
Classification is by intended use, not active configuration. If the feature exists in the product and could be used as intended, the obligations attach unless you can demonstrate the use case cannot occur in your deployment.
Naming these three sentences as myths is half the compliance work. The other half is changing the practice they hide. The 2 December 2027 start date gives you time to prepare, and the classification logic stays the same.
The Obligation Most Schools Haven’t Heard Of: Article 27 FRIA
There is one obligation in the Act that applies specifically to public bodies. Most state schools are public bodies. Most school leaders have not heard of it.
Article 27 requires a documented Fundamental Rights Impact Assessment, abbreviated as FRIA. It must cover:
- How the AI integrates into your processes
- How often the system will be used and over what period
- Which categories of pupils or staff are affected
- What specific harms could occur
- How human oversight is implemented
- What mitigation measures are in place
- What complaint procedures are available to affected persons
Independent and private schools may also be in scope as “private entities providing public services” depending on how their delivery is structured under national law. The safe assumption: if you are deploying any Annex III §3 high-risk tool, scope a FRIA process before the December 2027 deadline.
What to Do This Academic Year
These are the seven moves that get most schools to a defensible compliance position before the high-risk obligations apply on 2 December 2027. The literacy and prohibition obligations apply today, so the first few steps do not wait. Sequence matters.
- Run an AI inventory. List every AI tool currently in active use across teaching, admissions, behaviour management, and admin. Shadow AI is endemic. Ask teachers what they actually use, not what was procured.
- Classify each use case against Annex III, point 3. Use the free EU AI Act risk assessment tool for the case-by-case work.
- Audit any proctoring or “engagement analytics” tool for Article 5(1)(f) exposure. This is the highest-urgency item because the prohibition is already in force. If facial expression or “stress indicator” analysis is in the product, get legal advice before next use.
- Run a staff AI literacy baseline. Article 4 has been binding since February 2025. The free AI literacy self-assessment for educators gives you the snapshot.
- Draft or refresh your AI use policy. The free AI Policy Generator walks through the eight sections a policy needs in 2026.
- For state schools: scope your Article 27 FRIA process. Identify which tools will need one. Don’t try to do them all at once. Build the FRIA into procurement, not into a year-end scramble.
- Appoint a compliance owner. Most schools name an existing Director of Digital Learning, IT Director, or DPO. The role is coordination, not a new hire.
UK Schools: Does the EU AI Act Apply to You?
If your school is in the UK, the EU AI Act does not directly apply to your domestic operations post-Brexit. But three things mean most UK schools should align with it anyway.
- You have EU pupils, EU staff, or EU vendor relationships. Any AI tool processing data about EU students or staff still touches EU law somewhere upstream.
- UK AI policy is converging. The UK Department for Education’s generative AI guidance and the forthcoming AI in schools standards borrow heavily from the EU classification logic. Building EU-aligned policies positions you for the direction of UK regulation.
- It’s the best available framework. UK schools that adopt the EU AI Act risk classification voluntarily gain a structured way to defend AI use to their boards, inspectors, and parents.
Treat the EU AI Act as the operating standard until UK-specific guidance hardens. UK GDPR already does most of the data protection work.
FAQ
When does the EU AI Act start applying to schools?
Some obligations are already in force. Article 4 (AI literacy) and Article 5 prohibited practices (including emotion recognition in education) have applied since 2 February 2025. High-risk obligations under Annex III, point 3, apply from 2 December 2027, the date set by the Digital Omnibus on AI (Regulation (EU) 2026/1744). High-risk AI embedded in products covered by Annex I follows on 2 August 2028.
Is ChatGPT illegal in EU schools?
No, but its use creates obligations. As a limited-risk system under Article 50, pupils must be told they are interacting with AI. If it is used to determine grades or progression, that specific use case becomes high-risk under Annex III, point 3(b). The tool is allowed; the use case is what gets classified.
Are AI proctoring tools banned in EU schools?
Most are classified as high-risk under Annex III, point 3(d). Some are also prohibited under Article 5(1)(f) if they analyse facial expressions or “stress indicators” as emotional signals. Audit the specific tool’s feature set carefully and do not rely on vendor marketing claims.
Does a state school need to do a FRIA?
Yes. Most state schools qualify as “bodies governed by public law” and must complete an Article 27 Fundamental Rights Impact Assessment before deploying any Annex III, point 3 high-risk AI tool. The requirement applies from 2 December 2027, when the high-risk obligations start. Independent schools may also fall in scope if they provide public services under national law; check with counsel.
What if our school only uses AI for lesson planning?
Teacher-side AI used for lesson planning, rubric drafting, and feedback drafting is typically minimal risk under Article 6(3)(d) preparatory-task framing. It still falls under the Article 4 literacy obligation and should be covered in your AI use policy. Staff still need to understand what the AI does and where it can fail.
Does the EU AI Act apply to UK schools?
Not directly post-Brexit. But UK schools with EU pupils, EU vendors, or pan-European structures typically align with EU AI Act standards voluntarily, because UK Department for Education guidance is moving in the same direction. See the UK section above for the full picture.