Exam proctoring and AI in education: when the EU AI Act makes it high-risk
Proctoring, auto-grading and admissions AI all sit in Annex III point 3 of the AI Act. This guide explains the high-risk duties that fall on institutions and edtech vendors alike.
Exam proctoring, automated grading and admissions screening sit in the high-risk tier of the EU AI Act: Annex III point 3 covers AI that evaluates learning outcomes, steers learning, decides who enrols, or monitors candidates during tests. Run proctoring software and you are a deployer of a high-risk system with duties biting from December 2027 — while the edtech vendor carries the heavier provider obligations on top of GDPR.
What the AI Act exam proctoring rules actually say
The AI Act does not name products; it names purposes. Annex III point 3 pulls AI into the high-risk category where it evaluates individual learning outcomes, steers learning, assesses admission eligibility, or monitors and analyses candidate behaviour during tests. That last limb is where invigilation lives: a tool watching a webcam feed, flagging eye movement or detecting a second voice is analysing behaviour during an assessment — precisely the pattern the draughtspeople had in view.
Two points calm the alarmist readings of proctoring software EU AI Act law. First, classification attaches to intended purpose, not the camera: a timed quiz with no behavioural analysis is not high-risk because a webcam is on; a service scoring “suspicion” from video is. Second, high-risk is a burden, not a ban — lawful once the provider has done its conformity work and the deployer meets its duties. What changes is the paper trail: risk management, oversight, logging, candidate transparency, and a way to challenge the outcome.
The timetable matters: prohibited practices and AI literacy applied since February 2025, transparency duties follow in August 2026, and the Annex III high-risk obligations take effect on 2 December 2027 — a 2026 procurement cycle should already be writing those duties into its contract.
Education AI high risk classification in the EU: four tool types
Most institutions are not buying “one AI” but assembling tools that straddle the Annex III line differently:
| Tool type | Annex III point | High-risk? | Main duties |
|---|---|---|---|
| Remote invigilation / proctoring (behavioural monitoring, anomaly flagging during tests) | Point 3 — monitoring and analysing candidate behaviour during tests | Yes | Deployer: oversight of flags, candidate notice, logging, appeal path. Provider: conformity, CE marking, documentation |
| Automated grading and scoring (short-answer marking, score prediction) | Point 3 — determining or evaluating individual learning outcomes | Yes, almost always — the Article 6 filters are hard to claim when the score drives the result | Deployer: human review of disputed marks, input checks, student notice. Provider: accuracy and data-governance evidence |
| Adaptive learning and steering (content sequencing, personalised study pathways) | Point 3 — steering the process of learning | Frequently yes — borderline cases turn on whether it shapes the outcome | Deployer: pathway oversight, personalisation transparency. Provider: risk management, logging |
| Enrolment and admissions screening (applicant ranking, dropout risk scoring) | Point 3 — assessing eligibility for admission | Yes, where it influences admission decisions | Deployer: applicant notice, human decision-maker with authority to depart, appeal route. Provider: conformity, registration |
Classification does not turn on model sophistication but on whether the output bears on a person’s educational trajectory — a mark becomes a grade, a grade becomes an offer, and that chain is what makes Annex III point 3 so wide.
For AI grading systems compliance requirements: if the output is the mark, treat the system as high-risk and document that reasoning; if it is a suggestion a teacher can override, record the override rate as evidence the human remains in charge. “It seemed fine” will not survive a market-surveillance request — a one-paragraph rationale in your register will.
The narrow research and development carve-out before deployment
Vendors and university teams often reach for the research exemption to trial proctoring on live cohorts — a narrower door than it appears. The Act exempts systems used solely for scientific research, so prototyping on last year’s consenting students is research. The moment the system is put into service for real examinations — applied to candidates who did not enrol for a study — the research framing expires and Annex III point 3 applies in full.
A second filter sits in the classification article: an Annex III system drops out of the high-risk tier only for narrow procedural tasks, improving a previously completed human activity, or preparatory work that does not materially influence the outcome. Proctoring that merely opens a camera check before a human reviews the footage can argue into that filter; proctoring where the flag becomes the outcome cannot. The question to ask back: does a human decide the case, or does the flag decide the case?
Three practical consequences follow:
- Trials on live candidates are deployments — human review of every output, informed candidates, a written classification note.
- The carve-out does not suspend GDPR — lawful basis and a DPIA were needed from day one.
- In-house development can flip your role — a tool put into service under your own name may make you the provider.
AI in admissions: EU compliance from shortlist to offer
Admissions is where the AI Act collides most sharply with the calendar — clearing rounds do not wait for regulatory timetables. Annex III point 3 captures AI assessing admission eligibility, so applicant ranking, entry-test scoring and predicted-dropout models sit inside the high-risk tent wherever they shape who gets an offer.
The deployer duties concentrate in three places. Human oversight must be meaningful: a named admissions officer with authority to depart from a ranking, with a record of when they did. Candidates must be told they are subject to a high-risk system — an express duty for public-law institutions, and silence is a reputational risk in a trust-based sector. Input data must be relevant and representative: a model trained on one cohort and applied to another is a compliance failure, not a fairness debate.
Build the appeal route before the first application cycle: human reconsideration on request, a deadline, a written outcome — covering the explanation right affected persons hold where AI is used in individual decisions. A university that can produce that file has a defence; one that cannot faces market-surveillance and data-protection complaints together.
Student monitoring: AI Act obligations for the institutions running proctoring
Once classification is settled the work is almost entirely deployer-side: a university contracting an invigilation vendor makes the vendor the provider and the university the deployer, and the obligations split cleanly.
- Use the system as instructed. Running a “suspicion threshold” the vendor deliberately left switched off moves the failure onto the institution.
- Assign human oversight to trained people. A named assessment integrity lead must review flags with the competence to overrule them, and be able to explain how a flag is produced and its failure modes: low light, accents, shared accommodation, assistive technology.
- Check your input data. Two candidates sharing a flat with one webcam between them is not an accident; it is a predictable input condition to plan for.
- Keep the logs long enough to support an appeal and answer a market-surveillance request.
- Warn candidates and give them a route out: this is an AI system, here is what it looks for, here is how to request human review, here is the deadline.
- Run a fundamental rights impact assessment where you are a public body — for a university it doubles as the ethical case: necessity, proportionality, alternatives.
An institution holding all six artefacts can answer any regulator’s first letter the week it arrives.
Where GDPR meets the AI Act: behaviour and biometrics
Behavioural data is personal data, always. Keystroke patterns, gaze tracking and tab-switching logs describe identifiable students under observation — processing needs a lawful basis, and systematic monitoring during tests ordinarily requires a DPIA. Extend the existing one rather than writing a second document.
Biometric data needs a separate answer. Comparing a live face to an enrolment photograph is biometric processing, and where it uniquely identifies a person it falls into the special-category regime with a much harder lawful basis. Ask the vendor what is compared, where, and how long templates are kept.
Automated decisions are already a GDPR issue. Grades or admission decisions based mainly on automated processing engage GDPR’s automated-decision rules independently of the AI Act — and the human-review machinery you build for deployer duties largely satisfies them.
Two traps remain: retention — exam footage is evidence until the appeal window closes and a liability afterwards, so set the deletion date at procurement; and special-category triggers, where emotion inference or flags inferring disability push into sensitive-data territory neither the vendor’s marketing page nor your contract mentions.
What edtech vendors owe as providers, and how to run a learning platform AI Act risk assessment
For the software house the compliance weight sits on the other side of the contract: a vendor building an assessment, proctoring or admissions tool for the EU is a provider of a high-risk AI system, and the AI Act for schools ultimately runs through the product it buys. The provider duty set is the familiar stack — lifecycle risk management, data governance, technical documentation, automatic event logging, accuracy and cybersecurity targets, safe-use instructions, workable human-oversight design, conformity assessment (the internal-control route for Annex III education tools), CE marking and EU database entry.
Two provider tasks catch smaller edtech firms off guard:
- Instructions of use are a legal artefact: intended purpose, limits, oversight measures and the performance figures CE marking rests on — “see our help centre” will not do when an appeal panel asks what a flag means.
- Model updates can reopen conformity: material changes to scoring logic can require the assessment to be revisited, so build the refresh into the release process rather than treating CE marking as a one-off task.
For the buyer, the procurement risk assessment is straightforward: the classification rationale and its Annex III point, CE marking and registration status, the instructions-of-use document, the logging specification, and — where a fundamental rights assessment applies — a data sheet on behavioural and biometric processing. A vendor that hands over five documents in week one has done the work; one that promises them “post-signature” has not. Finally, configuring, retuning and marketing under your own name can cross you from deployer to provider — write in the contract who classifies, documents, notifies candidates and keeps the logs.
How to check your own position in three minutes
The free exposure scan asks twelve questions about your AI usage and governance and returns an exposure score, your ranked exposures with the exact article that fired, and your first three moves mapped to the dates that apply. Confirm your email and the full report unlocks.
Frequently asked questions
Is exam proctoring illegal under the EU AI Act?
No — proctoring is high-risk under Annex III point 3: the vendor owes conformity, CE marking and registration; the institution owes deployer duties (oversight, notice, logging, appeal route). Annex III obligations bite from December 2027, while literacy and prohibited-practice duties are already live.
Does a simple online quiz with a webcam count as high-risk?
Classification attaches to intended purpose, not the camera: a quiz with no behavioural analysis does not enter Annex III point 3 because a webcam is on — the moment the system evaluates candidate behaviour, the analysis begins.
Are we a provider if we build our own invigilation tool?
Develop the tool, set its purpose and put it into service under your own name and you can become the provider — inheriting documentation, conformity, CE marking and registration duties. Classify that with legal counsel before the first live exam.
What is the first document an inspector will ask for?
The classification rationale — why this tool sits where it does under Annex III point 3 — then the instructions of use and your candidate-notice copy. One current paragraph of reasoning in the register is the difference between a five-minute conversation and an escalation.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.
Keep reading
AI in healthcare under the EU AI Act: clinical decision support and high-risk duties
11 min
Risk & ClassificationThe Annex III High-Risk AI Systems List: All Eight Use Cases Explained with Examples
11 min
Risk & ClassificationThe Eight Prohibited AI Practices Under Article 5: A Compliance Checklist
10 min