Not legal advice. This article summarises official EU texts as of 6 November 2026; check the current texts and take advice on your own deployment.
The EU AI Act treats online proctoring under two different regimes, and buyers often blur them. Monitoring students during tests is a high-risk use, allowed with obligations. Using AI to infer students' emotions in an education institution is prohibited outright, and has been since 2 February 2025. Which regime a given proctoring feature touches depends on what it infers, not on how advanced it sounds.
This piece goes deeper on that one question. For a broader overview of AI Act obligations for proctoring, see our EU AI Act compliance guide for online proctoring.
Two regimes: prohibited practices and high-risk systems
The AI Act (Regulation (EU) 2024/1689) sorts AI uses into tiers. Two of them matter for proctoring:
- Prohibited practices (Article 5). These cannot be placed on the market or used at all, apart from narrow exceptions. Emotion inference in workplaces and education institutions is one of them.
- High-risk systems (Article 6 and Annex III). These are allowed, but providers and deployers carry obligations around risk management, logging, human oversight and transparency.
A single proctoring product can contain features in both categories. That is why a vendor's statement that "our product is AI Act compliant" is not an answer. The useful question is feature by feature.
The prohibition: inferring emotions in education
Article 5(1)(f) prohibits the placing on the market, putting into service for this purpose, or use of AI systems "to infer emotions of a natural person in the areas of workplace and education institutions", with an exception for medical or safety reasons. Under Article 113, Chapters I and II, which contain this prohibition, have applied since 2 February 2025.
The European Commission published guidelines on prohibited AI practices in February 2025. They are not binding (authoritative interpretation belongs to the Court of Justice), but they are the most detailed official reading available, and several points in them bear directly on proctoring:
- It covers deployers, not just vendors. Because the article prohibits the "use" of such systems, the guidelines say the ban applies to providers and deployers alike. A university using a tool that infers emotions is in scope even if the vendor built it.
- Online exams are included. The guidelines read "education institutions" broadly: public and private, all levels, and with no limitation by environment, online, in person or blended. They also say the prohibition applies to candidates during admissibility processes.
- It is limited to inferences from biometric data. The guidelines tie the prohibition to the definition of an emotion recognition system in Article 3(39), which is based on biometric data. Inferring emotions from written text is outside it. Inferring them from facial expressions, voice, body posture or the way someone types is inside it.
- Emotions are read widely, but observable expressions are not emotions. Detecting that someone is smiling, or that their head turned, is not emotion recognition. Concluding that they are happy, angry or anxious is. Physical states such as pain or fatigue are not emotions.
- The rationale is the power imbalance. The guidelines cite Recital 44's concerns about the scientific basis of emotion recognition and note that students are in a particularly vulnerable position.
The guidelines include examples that read as if written for proctoring. An education institution using eye tracking while examining students online, to detect if unauthorised material is used, is not prohibited, because it does not infer emotions. If the same system were also used to detect emotions such as emotional arousal and anxiousness, it would fall within the prohibition. Using emotion recognition to infer students' interest and attention is prohibited. (See section 7.2 of the full guidelines document.)
The high-risk regime: Annex III, point 3(d)
Separately, Annex III lists high-risk uses. Point 3, education and vocational training, includes in sub-point (d): "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."
Two things follow:
- The classification follows the intended purpose. Point 3(d) describes what the system is for, not how or where it computes. A proctoring system that runs entirely on the candidate's device is still a system intended for monitoring students during tests. On-device processing has real data protection benefits, but it is not an exit from Annex III.
- There is a narrow derogation, and it is not a loophole. Article 6(3) says an Annex III system is not high-risk where it does not pose a significant risk of harm, for example because it performs only a narrow procedural or preparatory task, but it also says an Annex III system that performs profiling of natural persons is always high-risk. A provider relying on the derogation must document why.
For a university or certification body, the main obligations as a deployer sit in Article 26. They include using the system according to the provider's instructions, assigning human oversight to people with the competence and authority to exercise it, monitoring operation, keeping the logs under your control for at least six months, using the provider's information to carry out your GDPR data protection impact assessment, and informing people subject to decisions the system helps make. Article 27 adds a fundamental rights impact assessment for deployers that are bodies governed by public law or private entities providing public services, which will include many universities. Article 14 requires providers to design high-risk systems so that people can understand, monitor, override and stop them.
Feature by feature
The table below maps common proctoring features to the two regimes. It is a reasoned reading of the texts and guidelines above, not a legal verdict; the right answer for a specific product depends on what it actually infers and how its output is used.
| Feature | Could it touch the Article 5(1)(f) ban? | Annex III, point 3(d)? | Notes |
|---|---|---|---|
| Face present / absent | Unlikely: presence is not an emotion | Yes, as part of test monitoring | Detecting a face is an observable state |
| Multiple faces | Unlikely | Yes | Counting people infers no emotion |
| Gaze or head direction | Not if used to detect looking at unauthorised material | Yes | The Commission's own example; it crosses the line if it scores attention or anxiousness |
| Tab or window focus | No: not biometric | Yes | Browser events, not body data |
| Audio level or a second voice | Not if it detects sound or speakers | Yes | Inferring stress or nervousness from the voice would be emotion inference |
| "Stress", "nervousness" or "suspicious expression" scores | High risk of falling under the ban | Yes | Inferring an emotional state or intention from the face or voice in an education setting is what the prohibition targets |
| Keystroke timing | Not if used to detect scripted or pasted input | Yes | The guidelines treat typing patterns as biometric when used to infer emotions |
| Age estimation | Not emotion inference | Depends on use | Other AI Act provisions on biometric systems may be relevant; ask the vendor how it is classified |
| Likeness comparison to a reference photo | Not emotion inference | Depends on use | Identity checks raise GDPR special-category questions; ask the vendor and your DPO |
The pattern is consistent: features that detect observable facts (a face, a second voice, a lost focus, a pasted answer) sit in the high-risk regime. Features that claim to read an inner state (stress, nervousness, guilt, attention, intent) are where the prohibition bites.
Timeline as of 6 November 2026
| Date | What applies |
|---|---|
| 2 February 2025 | Chapters I and II, including the Article 5 prohibitions such as emotion inference in education (Article 113) |
| 27 July 2026 | The Digital Omnibus on AI, politically agreed on 7 May 2026, entered into force |
| 2 August 2026 | General date of application for most of the rest of the Act |
| 2 December 2027 | High-risk obligations for Annex III systems, including exam proctoring, as amended by the Omnibus (Commission) |
| 2 August 2028 | High-risk obligations for Annex I (product-embedded) systems |
Two cautions. First, the postponement moved the high-risk deadline; it did not touch the prohibition, which already applies. Second, some earlier commentary, including an earlier version of our own overview guide, cites 2 August 2026 for the high-risk obligations. That was the original date and is no longer correct for Annex III systems. The Commission's AI Act Service Desk timeline reflects the amendments.
Questions to put to your proctoring vendor
- Does any feature output an emotional state, mood, stress, nervousness, attention, engagement or intent score, for any candidate, in any configuration?
- For each signal based on the face, voice or typing, what exactly does it detect, and what does it not infer?
- How do you classify the product under Annex III, point 3(d)? If you rely on the Article 6(3) derogation, can we see the documented assessment?
- What will you provide to support our Article 26 duties: instructions for use, logs, information for our DPIA, and a description of human oversight?
- Can a human reviewer see the evidence behind every flag and override any outcome?
- If you offer age estimation or face comparison, what is each used for, and how is it classified?
- Can these features be switched off per exam?
How ProctorSafe fits
ProctorSafe does not infer emotions. It has no stress, nervousness, attention or intent score, in any configuration. It detects observable events: face presence and multiple faces, gaze direction, tab and window focus, screen sharing, audio levels and a second voice, and how answers reached the page. Where identity checks are enabled, the candidate's face is compared with their own reference photo. Proctored exams use no age estimation. Detection runs in the candidate's browser, and every signal lands on a timeline a human reviewer can check and contest, with the trust score supporting rather than replacing that review. Our GDPR-compliant proctoring overview describes the data flows, and AI cheating detection explains how answer-provenance signals work.
For the human side of oversight, see our piece on bias, fairness and human-in-the-loop review in AI proctoring. To see the review timeline for yourself, try the candidate demo.