The EU AI Act Now Bans Emotion Recognition in Education. What That Means for "Sentiment Analysis" of Course Feedback

Bottom line up front: Since 2 February 2025, Article 5(1)(f) of the EU AI Act has prohibited — not merely regulated — AI systems that infer a person's emotions from biometric data in workplaces and education institutions. This is the Act's most severe tier: a red line, enforced with fines of up to €35 million or 7% of worldwide annual turnover. If a course-evaluation or "student experience" product infers mood from a webcam image of a student's face, or emotional state from the acoustic properties of their voice, that is now unlawful in an EU university. Sentiment analysis of typed comments generally sits outside the ban — but the boundary is narrower and sharper than most marketing pages admit, and GDPR closes much of the gap the AI Act leaves open.

If you are procuring AI to make sense of student feedback, this is the single compliance fact you cannot afford to get wrong. A high-risk system can be used with conformity assessment, documentation, and human oversight. A prohibited system cannot be used at all.

What Article 5(1)(f) actually says

The EU AI Act lists a small set of practices deemed to carry unacceptable risk. Article 5(1)(f) prohibits placing on the market, putting into service, or using "AI systems to infer emotions of a natural person in the areas of workplace and education institutions," with two narrow carve-outs: systems put in place for medical or safety reasons.

Two definitions do the work. An "emotion recognition system" (Article 3(39)) is an AI system for identifying or inferring the emotions or intentions of natural persons on the basis of their biometric data. "Biometric data" means data resulting from specific technical processing of physical, physiological, or behavioural characteristics — a face, a voice, a gait, a pattern of keystrokes. The prohibition therefore has a precise trigger: emotion or intention inferred from biometric signals, in an education setting. All four conditions must hold at once — an in-scope AI system, biometric data, inference of emotion, and a workplace or educational context — for the red line to apply (Future of Privacy Forum analysis).

The prohibitions took effect on 2 February 2025, ahead of most of the Act, and the top penalty tier under Article 99 — up to €35 million or 7% of global turnover, whichever is higher — attaches to breaches of Article 5 (Article 99).

Why education got a red line

Recital 44 is unusually candid about the reasoning. The legislator points to the thin scientific basis for emotion-inference technology: limited reliability, lack of specificity, and limited generalisability across cultures and contexts, all of which can produce discriminatory outcomes. It adds a second concern specific to classrooms and offices: the power imbalance. Students and employees are not free actors negotiating consent; a system that reads their affect can lead to "detrimental or unfavourable treatment" of individuals or whole groups. The EU decided the combination of dubious accuracy and structural coercion was not worth tolerating in the two settings where people have least room to refuse.

For anyone who has read the research on shifting standards and construct-irrelevant variance in evaluation, the reliability critique will sound familiar. An affect-detection model that mislabels a concentrating student as "confused," or reads a non-native speaker's prosody as "negative," is not a neutral instrument. It is a new source of bias wearing the costume of objectivity.

The line that matters: biometric inference vs text sentiment

Here is where vendors and buyers routinely conflate two very different things.

Text sentiment analysis — scoring the words a student typed as positive, negative, or neutral — is not, on its own, emotion recognition under the Act, because typed text is not biometric data. A model that reads the sentence "the pacing was rushed" and tags it "negative" is analysing content, not inferring emotion from a physiological signal. This kind of scoring falls outside Article 5(1)(f). (Whether it is useful is a separate question — as we argued in why a sentiment score is not insight, "78% positive" tells you almost nothing about what to change.)

Voice-tone and facial-expression analysis are a different matter. If an "AI interview" or proctoring-adjacent tool infers a student's emotional state from the acoustics of their voice or a webcam image, it is processing biometric data to infer emotion — in an education institution. That is squarely within the prohibition. The medium is what moves it across the line: the same claim ("we detect student frustration") is lawful nowhere near the same way depending on whether it comes from parsing words or from measuring a body.

This distinction is easy to miss because product copy blurs it. "Emotion AI," "affective analytics," and "engagement detection" are marketed interchangeably whether they run on text, audio, or video. Under EU law they are not interchangeable at all.

GDPR closes part of the gap the AI Act leaves open

Suppose a tool only does text sentiment and is therefore outside Article 5. It is not therefore unregulated. Inferring emotional or health states can generate special-category data under Article 9 GDPR, a problem we examined in the special-category data hiding in free-text feedback. And where emotion inference is permitted under the AI Act (for example, outside education, or under the narrow medical/safety exception), Article 50 still requires deployers to inform the people exposed to the system that it is operating. The floor, in other words, is transparency even at its most permissive — which connects to the disclosure duty we set out in should you tell students an AI is running their evaluation.

"But surely our tool is exempt" — the counterarguments, read honestly

"We only analyse typed text, so the ban does not touch us." Largely correct — but check what your pipeline actually ingests. If you record spoken responses and run any affect or tone model over the audio, you are processing biometric data. Transcribing speech to text and analysing the words is fine; scoring the voice is not. Ask the vendor precisely which signal the emotion claim is computed from, and get it in writing.

"Ours is for safeguarding, so the safety exception applies." The medical and safety carve-outs exist, but Recital 44 and the regulators are explicit that they are to be read narrowly, with a high bar. Detecting distress to route a student to support is a genuinely hard case; a blanket "wellbeing analytics" dashboard marketed to administrators is not what the exception was written for. Do not assume it covers you.

"Emotion detection makes evaluation richer." Even setting law aside, the evidence on the validity of these systems is weak, and the EU has now judged the accuracy-plus-coercion trade-off unacceptable in classrooms. Richness that you cannot trust, and cannot lawfully deploy, is not richness.

What compliant, useful AI course evaluation looks like

The honest position is not "avoid AI" — it is "analyse what students say, not what their faces or voices supposedly feel." That is the design principle behind Koji for Education. Koji runs AI-moderated conversational interviews that probe beyond a rating, then applies automatic thematic analysis to the content of open-text responses — the reasons, examples, and suggestions a student offers. It does not infer emotional state from biometric data, and where spoken responses are used, the object of analysis is the transcript, not the timbre. Every surfaced theme is traceable to the quotes that produced it, so a human reads evidence rather than an opaque affect score.

That keeps a system on the right side of Article 5(1)(f) while still doing the thing universities actually need: turning hundreds of comments into a defensible, actionable picture of the course. (Teams running general customer and user research face the identical line, which is why the main Koji platform is built on the same content-first interview engine.)

Emotion-recognition marketing will not disappear overnight. But in an EU education institution, the compliant question is no longer "can our AI tell how students feel?" It is "can our AI show us, with evidence, what students are telling us?" Only one of those is still legal — and it was always the more useful one.

Frequently asked questions

Is analysing student course feedback with AI banned under the EU AI Act? No — not as such. Article 5(1)(f) bans inferring emotions from biometric data (face, voice, physiology) in education institutions. Analysing the content of typed or transcribed feedback with AI is not prohibited, though it may be high-risk or subject to GDPR depending on use.

When did the emotion-recognition prohibition take effect? The Article 5 prohibitions applied from 2 February 2025, ahead of most other provisions of the AI Act.

What are the penalties for breaching the prohibition? Breaches of Article 5 sit in the top penalty tier: up to €35 million or 7% of total worldwide annual turnover, whichever is higher (Article 99).

Does text sentiment analysis count as emotion recognition? Generally no, because typed text is not biometric data. Voice-tone or facial-expression analysis does count, because it infers emotion from biometric signals. The distinction turns on the signal, not the marketing label.

Are there exceptions to the ban in education? Only narrow ones, for medical or safety purposes, and regulators say they must be interpreted strictly. General "wellbeing" or "engagement" analytics are unlikely to qualify.

How does Koji stay on the right side of this rule? Koji analyses the content of what students say — thematic analysis of open-text and transcribed responses with quote-level provenance — rather than inferring emotional state from biometric data such as faces or voices.