Compliance

Emotion recognition vs behavioural analytics

Vendors in this category use "engagement analytics" to describe two fundamentally different technologies. One is prohibited in EU workplaces. The other is not. Buyers need to be able to tell them apart.

6 min read
In short

Emotion recognition infers an internal affective state from biometric data such as facial expression or vocal tone. Behavioural analytics measures observable actions — tab-switching, focus loss, dwell time — and draws no conclusion about feelings. Under EU AI Act Article 5(1)(f) the first is prohibited in workplace and education contexts; the second is not.

The distinction in one table

Emotion recognition Behavioural analytics
Input Face, voice, physiological signal Browser and interaction events
Hardware Camera or microphone None
Inference Internal affective state Observable action
Example output "Learner appears confused" "Learner replayed this section twice"
Biometric data? Yes No
AI Act Art. 5(1)(f) ❌ Prohibited in workplace/education ✅ Not caught
GDPR Special category, high bar Ordinary personal data, or none if aggregated
DPIA Almost certainly required Usually required for employee monitoring
Works council Frequently refused Routinely approved

The AI Act prohibition hangs on a specific chain: an AI system, that infers emotions, from biometric data, in a workplace or education institution. Break any link and Article 5(1)(f) does not apply.

Behavioural analytics breaks two links at once. It does not infer emotions — it records what happened. And it does not use biometric data — a visibilitychange event is not a physiological characteristic.

That is not a loophole. It reflects what the prohibition is actually aimed at: the inferential leap from someone's body to a claim about their inner state, in contexts where they cannot meaningfully refuse.

The test that settles most cases

Ask what the system would output if it were wrong. A behavioural system that mis-measures says "spent 40 seconds here" when it was 30 — an inaccurate observation. An emotion system that mis-measures says "this person was frustrated" when they were concentrating — an inaccurate claim about someone's mind. The second is the category the law is worried about.

Where the marketing blurs it

Four patterns worth watching for on vendor calls.

"Engagement score" with no stated input. Engagement is an output, not a signal. Ask what goes into it. A number between 0 and 1 tells you nothing about whether a camera was involved.

"Attention tracking" that means gaze. Attention derived from focus and dwell events is behavioural. Attention derived from eye position is biometric, and sits in a genuine grey area.

"Privacy-first because it runs on-device." True and valuable for GDPR. Irrelevant to Article 5. Vendors conflate these constantly, and the conflation is doing a lot of work in this market.

"Emotion AI, but only aggregated." Aggregation happens after inference. If the pipeline inferred emotional states from biometric data to produce the aggregate, the prohibited step already occurred.

Why the behavioural approach also performs better than expected

The assumption that facial analysis is the "real" measurement and behavioural signals are a compromise does not survive contact with the literature.

A 2024 study identified tab-switching as the strongest single predictor of disengagement in online courses, ahead of self-regulation and satisfaction measures. Studies fusing facial expression with behavioural signals report accuracy improving from around 91.5% to 94.6% — meaning the facial layer contributes roughly three percentage points while carrying essentially all of the regulatory risk.

For content diagnostics specifically — which module loses people, and where — behavioural signals are not a fallback. They are the primary evidence, and they aggregate cleanly to a level where no personal data is involved at all.

What to put in the contract

If you are buying in the EU, three clauses are worth insisting on:

  1. A warranty that the deployed configuration does not infer emotions from biometric data in your workplace or education context.
  2. A description of the actual signal set in the schedule, not just the marketing name of the output.
  3. A change-control obligation, so the vendor cannot enable emotion inference in a later release without your written agreement.

A vendor confident in its position will sign all three without much discussion. Reluctance on any of them is informative.

Frequently asked questions

Is engagement an emotion?
Not as the AI Act uses the term. Engagement is a construct that can be derived from emotional inference or from behavioural observation. The legal question is not what you call the output but what you inferred it from. An engagement score derived from facial affect is an emotional inference; one derived from dwell time is not.
What if a product does both?
Then the analysis depends on configuration, and you should insist on contractual clarity. Ask whether the behavioural-only mode is a genuine product mode or a setting that merely suppresses display while emotional inference continues in the pipeline. Guidance has not resolved whether a system with emotion inference as a secondary function is caught, so a conservative reading is prudent.
Does client-side processing change the classification?
No. Where the computation runs affects your GDPR analysis — client-side processing means biometric data is not transmitted — but it has no bearing on whether a practice is prohibited under Article 5. A prohibited practice is prohibited on-device.

Not sure which side of the line a product sits on?

Run it through the checker — eight questions, no email required.

Open the AI Act checker
Get Started

See what completion rates can't tell you

Find out exactly where your content works, where it fails, and what disengagement looks like before people leave.

Request a Demo