EU AI Act high-risk education: what Annex III lists and who signs
What the EU AI Act's Annex III, point 3 lists for education, which duties sit with NUADU or the institution, and what 'EU AI Act-ready' does and does not mean.
The EU AI Act's high-risk education list includes AI that evaluates learning outcomes, and NUADU's AI Transparency Notice applies that classification to its own AI assessment capabilities when they are used in institutional settings. The phrase EU AI Act-ready on NUADU's pages points to a workflow that follows from it, not to a badge. This piece sets out what Annex III lists for education, what the Act asks of high-risk systems, who carries each duty under NUADU's Institutional Agreement, and what ready does and does not mean.
Is AI in education high-risk under the EU AI Act?
Some of it is. Annex III is where the Act lists high-risk uses by area, and point 3 of that annex covers education and vocational training at all levels. It names four uses, and one of them, evaluating learning outcomes, is the core of AI-assisted marking. Paraphrased, Annex III, point 3 of the AI Act covers AI systems intended to be used for:
- (a) deciding access or admission to education and training institutions, or assigning people to them;
- (b) evaluating learning outcomes, including where those outcomes are used to steer a person's learning;
- (c) assessing the level of education a person will receive or be able to access;
- (d) monitoring students during tests to detect prohibited behaviour.
NUADU's notice takes a position on point 3(b) in two sentences: "AI systems intended to evaluate learning outcomes are high-risk under the Act. This applies to NUADU's AI assessment capabilities when they are used in institutional settings." Everything below starts from that classification. It is NUADU's own statement about its own capabilities, and it is not legal advice about any other system.
Read the Act with your counsel. What follows maps a classification onto a marking workflow; it is not a legal opinion.
AI prepares the evidence; a named educator decides.
What are the EU AI Act human oversight requirements?
Article 14 asks that a high-risk system be designed so that people can oversee it effectively while it is in use. It must be provided so that the people assigned to oversight are enabled, as appropriate and proportionate, to understand what the system can and cannot do, stay alert to over-relying on its output, interpret that output correctly, and decide to disregard, override or reverse it. The Act also asks deployers to assign that oversight to people with the necessary competence, training and authority (Article 26(2)).
In an assessment workflow those people already have job titles. With AI assistance on, the marker reviews a draft score, overrides it where they disagree, and signs the result. The invigilator reviews an integrity signal after the fact, and the institution's academic-integrity process decides. A credential is issued only once a named educator has signed the result. In each case the decision belongs to the person, and the system records who made it.
Human in the loop? In assessment the loop has job titles: marker, invigilator, approver. What human in the loop means for AI in education walks through where each of them sits, module by module.
Two details matter to a reviewer reading Article 14 against a product. First, understanding a system's limits is easier when the limits are written down. NUADU's notice states them: models can mis-score, miss errors or give irrelevant feedback, and no output is definitive without human review. Second, oversight is easier to scope when the AI is. Each AI module can be enabled per assessment, per programme, or per institution — or left off entirely, so a review covers the modules actually running and a module that is off has nothing to oversee.
Who carries which obligation, provider or institution?
Under NUADU's Institutional Agreement, the provider duties listed below sit with NUADU, record-keeping is shared, and human oversight in operation sits with the institution. The notice sets the allocation out article by article:
- NUADU, as provider: risk management (Article 9), data governance (Article 10), technical documentation (Article 11), transparency to deployers (Article 13), accuracy, robustness and cybersecurity (Article 15), conformity assessment (Article 43) and registration in the EU database (Article 49).
- NUADU and the institution together: record keeping and logging (Article 12).
- The institution, as deployer: human oversight in operation (Article 14), using the review, override and sign-off capability NUADU provides.
The shared line is easy to misread. The record exists for every assisted decision, and what it holds — the output, the override and its reason, the approver, the timestamp — is set out in what an AI audit trail records. The oversight line is easy to over-promise. A vendor can build review, override and sign-off into the product; only the institution can staff them with people who have the authority to use them.
The same allocation answers the GDPR's question about automated decisions. In institutional deployments, the notice states that no decision about a named learner is taken solely by automated means; the marker signs every result, and the learner can contest it through the institution's appeals process. Where the data itself is hosted, and under which contract, is a separate GDPR question. The Trust & compliance page covers the DPA, data residency options and sub-processors, and does GDPR require data to be stored in the EU answers the storage question directly.
What does EU AI Act-ready mean, and what does it not?
It means the controls are in the workflow and can be inspected. The notice names them: oversight, logging and documentation built into the platform workflow, with each AI output logged, each override recorded, and a named educator or approver deciding and signing. It then draws the line itself: "EU AI Act-ready" describes these controls, it is not a certification claim, and NUADU does not claim to be "EU AI Act certified" or "compliant".
What remains open is stated with a date. As of 2 October 2026, the notice lists four formal items as in preparation: the risk management system, data governance and technical documentation (Articles 9 to 11); conformity assessment (Article 43); registration in the EU database (Article 49); and machine-readable marking of AI-generated output (Article 50(2)). The AI Transparency Notice is where that status is updated when it changes, and it is the page to cite rather than this one.
Ready, then, means inspectable. It does not mean a completed conformity assessment, and it is not a judgement about any institution's own use. The inspection is the institution's to run, and the material for it is on request: a documentation pack covering methodology, validity evidence, bias review, retention, data flows and sub-processors, and an EU AI Act briefing for assessment.
Frequently asked questions
Is exam proctoring high-risk under the EU AI Act?
It can be. Annex III, point 3(d) lists AI systems intended for "monitoring and detecting prohibited behaviour" of students during tests in education institutions. What NUADU's pages say about its own integrity layer is factual: lockdown browser and identity verification at exam start are the defaults, other layers are opt-in per assessment, and integrity signals go to a named invigilator as surfaces, not findings. Whether a given configuration falls under point 3(d) is a question for counsel.
Does the EU AI Act ban emotion recognition in education?
Yes, with narrow exceptions. Article 5(1)(f) prohibits AI systems that infer the emotions of a person in workplaces and education institutions, unless the system is intended for medical or safety reasons. NUADU's statement about its own platform is separate and plain: it does not infer attention, stress, engagement, confidence or intent from face, voice, gaze or keystrokes, and anomaly signals are objective events. That is a description of the product, not a legal assessment.
Does the EU AI Act apply to higher education?
Yes. Annex III, point 3 applies to educational and vocational training institutions at all levels, so a university that uses AI to evaluate learning outcomes is within the same list as a school. For a university, the practical questions are the ones above: which modules are running, who carries which duty, and who signs.
→ See the AI capabilities page — every module, who decides, what it never does