Proctoring and AI assessment: what the AI Act requires from a training organisation
Editorial lead: The Kit Certif team
A candidate sits the final assessment of a course from their living room, webcam on. Software films them, records their screen and automatically flags “suspicious behaviour” to the trainer. Another tool grades case studies and suggests an exit level. For many distance-learning providers in France, these features are now routine. The EU Artificial Intelligence Act (AI Act) nevertheless places them in a specific category, with a timetable that changed in July 2026. Here is what already applies, what arrives on 2 December 2027 and what the French data protection authority (CNIL) has expected since 2023.

Education and vocational training: a high-risk area
Annex III of Regulation (EU) 2024/1689 lists the areas in which an AI system is classified as high-risk. Its point 3, “Education and vocational training”, covers four uses, “at all levels”:
| Annex III point | Use covered | Example in a training organisation |
|---|---|---|
| 3(a) | Determining access, admission or assignment to education and vocational training institutions | Automated screening of applications to a selective course |
| 3(b) | Evaluating learning outcomes, including to steer the learning process | Automated grading of a final assessment, adaptive path deciding the next module |
| 3(c) | Assessing the appropriate level of education a person will receive or be able to access | Placement test assigning a learner to a level group |
| 3(d) | Monitoring and detecting prohibited behaviour during tests | Proctoring with automated cheating detection |
A training organisation does not need to be a school or university to be covered: the text refers to education and vocational training institutions. Assessing learning outcomes is central to the French Qualiopi quality framework (indicator 11); as soon as it is handed to an AI system, it enters this scope.
There is a safety valve. Article 6(3) provides that an Annex III system is not high-risk when it does not pose a significant risk, for instance if it performs a narrow procedural task, improves the result of a previously completed human activity or prepares an assessment without replacing it. But the same article states that an Annex III system is always high-risk when it performs profiling of natural persons. The provider must document this assessment before placing the system on the market; ask for it in writing.
What has been prohibited since February 2025
Article 5 of the AI Act lists prohibited practices, applicable since 2 February 2025. Its paragraph 1, point (f), prohibits the use of AI systems to infer the emotions of a natural person in the workplace and in education institutions, except for medical or safety reasons.
For a training organisation the consequence is direct: a feature that claims to spot a candidate’s stress, anxiety or nervousness to infer cheating has no place in your exam set-up. Breaching the Article 5 prohibitions can lead to administrative fines of up to EUR 35 million or 7% of total worldwide annual turnover (Article 99(3)). Read your tools’ documentation with this in mind: words such as “engagement”, “attention” or “emotional state” should trigger a question to the vendor.
The new timetable: 2 December 2027
Regulation (EU) 2026/1744 of 8 July 2026, known as the “digital omnibus on AI”, published in the Official Journal on 24 July 2026 and in force since 27 July 2026, amended Article 113 of the AI Act. Sections 1 to 3 of Chapter III, which carry the requirements and obligations for high-risk systems, now apply:
- from 2 December 2027 for systems classified as high-risk under Annex III, including education and vocational training;
- from 2 August 2028 for systems linked to the products of Annex I.
The delay is real, but it changes nothing on two points: the Article 5 ban on emotion inference already applies, and the tool contracts you sign today will still be running in December 2027.
Your obligations as a deployer from that date
The software vendor is the provider; the organisation using it under its own authority is the deployer. Article 26 sets out the obligations of deployers of high-risk systems. The most concrete ones for a training organisation:
- Use the system in accordance with its instructions for use (paragraph 1). A tool designed to flag anomalies to a human must not be used to invalidate an exam automatically.
- Assign human oversight to people who have the necessary competence, training and authority (paragraph 2). The trainer or learning manager who decides on alerts must know what the tool detects, and what it cannot do.
- Ensure input data is relevant where you control it (paragraph 4): marking schemes, instructions, expected answers.
- Monitor operation and suspend use where there is a risk, informing the provider and the authority (paragraph 5).
- Keep the automatically generated logs under your control for at least six months, unless other rules apply (paragraph 6).
- Inform people: deployers of Annex III systems that make or help make decisions about natural persons inform them that they are subject to the system (paragraph 11). If the tool is used on your own employees, workers’ representatives and the employees concerned are informed before it is put into service (paragraph 7).
Article 86 adds a right to explanation: a person subject to a decision based on an Annex III system, with legal or similarly significant effects they consider adverse, can obtain clear explanations from the deployer about the role of the system in the decision. A learner whose exam was invalidated after an automated alert fits this case exactly. Breaches of deployer obligations fall under the EUR 15 million or 3% of worldwide turnover ceiling (Article 99(4)).

What the CNIL already expects
The AI Act does not replace the GDPR. On remote proctoring, the CNIL published a recommendation adopted on 8 June 2023 (deliberation no. 2023-058), applicable to any type of online exam or certification, organised by public or private bodies. Three points stand out:
- remote proctoring should be an option offered to candidates rather than an obligation, with an in-person alternative, except in specific cases (health crisis, organisations where distance learning is at the core of their model);
- any automated analysis of candidates’ behaviour should be excluded; automated analysis is limited to the candidate’s environment and to situations that justify it;
- candidates must be informed as early as possible of the proctoring conditions, and tools should be easy to install and uninstall.
In other words, automated cheating detection based on eye movements or facial expressions is already a GDPR problem, well before the 2027 deadline.
What research shows
The stakes are not only legal. In an analysis published in 2021 in Philosophy & Technology, Simon Coghlan, Tim Miller and Jeannie Paterson examine online exam supervision technologies in light of academic integrity, fairness, transparency, privacy, autonomy and trust (Coghlan, Miller and Paterson, 2021). They encourage institutions considering these tools to carry out prior inquiries and put governance and review processes in place to justify their use. That is precisely the logic of the human oversight the AI Act requires.
Checklist for a training organisation
- List the AI tools involved in admission, placement, assessment or exam monitoring.
- Remove any emotion-inference feature.
- Ask each vendor whether it classifies its system as high-risk and, if not, for the documentation of its Article 6(3) assessment.
- Keep decisions human: a tool alert triggers a review by an authorised person, never an automatic sanction.
- Offer an in-person alternative and inform candidates at enrolment (welcome booklet, terms, exam notice).
- Train the people providing human oversight and keep a record of that training.
This last point echoes Article 4 of the AI Act, rewritten in July 2026: providers and deployers take measures to foster the development of AI literacy among their staff, without having to guarantee a specific individual level. For your trainers and administrative teams, the AI literacy course (AI Act, Article 4) from Formations Obligatoires (in French) lasts 7 hours online and issues a personal certificate of completion, useful to document these measures. For day-to-day use of generative AI in your organisation, see our guide generative AI in a training organisation, and for inspections of distance learning, the article on inspections under an assumed identity.
Take action
Review your assessment tools before renewing a subscription: a contract signed in 2026 often runs beyond December 2027. To make sure your assessment methods, learner information and regulatory monitoring hold up at a Qualiopi certification audit, the Complete Kit Certif brings together the indicator sheets and evidence templates, updated for the framework applicable from 1 November 2026.
Frequently asked questions
+Is a proctoring tool used by a training organisation a high-risk AI system?
If it uses AI to monitor and detect prohibited behaviour during exams, it falls under point 3(d) of Annex III of Regulation (EU) 2024/1689, which covers vocational training at all levels. Since Regulation (EU) 2026/1744, the related obligations apply to Annex III systems from 2 December 2027.
+What is already prohibited today?
Since 2 February 2025, Article 5 prohibits the use of AI systems to infer the emotions of a person in the workplace and in education institutions, except for medical or safety reasons. A module claiming to detect a candidate's stress or nervousness to flag cheating must be avoided. The maximum fine for prohibited practices is EUR 35 million or 7% of worldwide turnover.
+Does the training organisation that buys the tool have obligations, or only the vendor?
Both. The vendor is the provider; the organisation using the system under its own authority is the deployer. From the Annex III deadline, the deployer must use the system according to its instructions, assign human oversight to competent and trained people, keep the logs under its control for at least six months and inform the people concerned.
+Does the CNIL allow remote proctoring of online exams?
Yes, with conditions. In its recommendation of 8 June 2023 (deliberation no. 2023-058), it considers that remote proctoring should be an option offered to candidates rather than an obligation, with an in-person alternative except in specific cases, and that any automated analysis of candidates' behaviour should be excluded.