Compliance

Lawful alternatives to emotion recognition for measuring engagement

Emotion recognition is prohibited in EU workplaces and education institutions, and unpopular almost everywhere else. The good news is that it was never the strongest signal available.

8 min read
In short

The most effective lawful alternative is behavioural telemetry: tab-switching, window focus loss, dwell time per section, scroll depth and session abandonment. Research published in 2024 found tab-switching to be the single strongest predictor of disengagement in online learning — outperforming both self-regulation and satisfaction measures — and none of it involves biometric data.

Why you might need an alternative

Three reasons come up, and only one of them is legal.

The law. Since February 2025, Article 5(1)(f) of the EU AI Act has prohibited AI systems that infer emotions from biometric data in workplace and education contexts. No consent exception. Fines to 7% of global turnover.

The culture. Even where it is lawful, telling employees or students that their faces are being analysed goes badly. Student campaigns against proctoring software have made the reputational cost concrete, and objections have included that gaze-based flagging discriminates against neurodivergent people. Works councils in Germany, the Netherlands and the Nordics will generally refuse.

The evidence. Facial emotion recognition has real accuracy limits — cultural variation in expression intensity, the gap between posed and spontaneous expression, degradation with glasses, beards and poor lighting. It works, but it is not the oracle vendors imply.

The seven signals that work

None of these require a camera. All of them are available through native browser APIs.

# Signal What it captures API
1 Tab-switching Learner navigates away from the content Page Visibility API
2 Window focus loss Application loses focus without a tab change blur / focus events
3 Dwell time per section Time actually spent on each slide or segment Intersection Observer
4 Scroll depth and velocity Whether content was read or skimmed Scroll events
5 Interaction latency Response time to prompts and questions Event timestamps
6 Replay and rewind Sections needing a second pass — confusion signal Media events
7 Session abandonment Where people leave, and what preceded it Unload / heartbeat

The seventh is the ground truth that makes the rest useful. Once you know where people actually quit, you can look backwards at what the other six signals were doing in the minutes beforehand, and build a predictive picture of what disengagement looks like before it happens.

What the research actually shows

This is the part that surprises people.

#1
Tab-switching ranked the strongest single predictor of online disengagement
ScienceDirect, 2024
+3.1pt
Accuracy gain from adding facial expression to behavioural signals (91.5% → 94.6%)
OUCI/DNTB online learning study
73.3%
Disengagement prediction accuracy with 40% of the session still remaining
Boote, Agarwal & Mostow, 2021

A 2024 study found "cyberloafing" — switching away from the learning content — to be the single most important predictor of disengagement in online courses, outperforming self-regulation and satisfaction measures. It is a behavioural signal, it requires no camera, and it is the best individual predictor anyone has found.

Meanwhile, multi-signal studies that add facial expression to behavioural data report accuracy moving from roughly 91.5% to 94.6%. That increment is real. It is also small, and it is where all of the legal and cultural risk lives.

If you are diagnosing content rather than people — which is where the commercial value sits anyway — behavioural signals are sufficient.

Building a composite engagement score

A single signal is noisy. Someone who tabs away for eight seconds may be checking a reference; someone who tabs away for eight minutes has gone. The useful construct is a composite.

A workable approach:

  1. Establish a per-learner baseline. People differ enormously in browsing habits. Measure deviation from an individual's own pattern, not against a population average. This is the same principle that makes facial baseline calibration work, applied to behaviour.
  2. Weight by dwell-adjusted attention. Time on section, discounted for time the window lacked focus. Someone with a module open in a background tab is not engaged.
  3. Score interaction quality, not just quantity. Latency to first interaction after a prompt is more informative than a click count.
  4. Flag pattern breaks. A learner who was steady for twenty minutes and has now switched tabs three times in ninety seconds is disengaging. The pattern is the signal, not the individual events.
  5. Aggregate to content level. "Section 4 loses 60% of learners" is actionable, needs no personal data, and avoids the entire surveillance conversation.

That last point deserves emphasis. Most of the value in engagement analytics is content diagnostics, not individual monitoring. Once you aggregate, you often stop processing personal data altogether — which resolves the GDPR question, the works council question and the cultural question simultaneously.

What you lose, honestly

Two things, and it would be dishonest not to name them.

Emotional texture. Behavioural signals tell you that someone disengaged. Facial analysis offers a hypothesis about how it felt — confusion looks different from boredom, which looks different from frustration. If your use case genuinely depends on distinguishing those states, behavioural data will not give it to you.

Passive-viewing sensitivity. For long-form video where a learner sits still and watches, behavioural signals are thin. There is not much to measure when the correct behaviour is stillness. Mitigations exist — periodic interaction prompts, chaptered content, comprehension checks — but they change the content design rather than just the measurement.

For most corporate L&D and e-learning use cases, neither loss is decisive. You wanted to know which module to fix. Behavioural data tells you that, lawfully, without a camera, and without the conversation about whether you are surveilling your staff.

Frequently asked questions

Is behavioural tracking really as good as facial analysis?
For predicting disengagement, the evidence suggests behavioural signals do most of the work. Studies combining facial expression with behavioural signals report classification accuracy improving from about 91.5% to 94.6% — so the facial layer contributes roughly three percentage points. For content diagnostics, where you want to know which section loses people rather than how any individual felt, behavioural signals are sufficient on their own.
Do I still need consent for behavioural tracking?
You need a lawful basis under GDPR, and for employee monitoring you will normally need a DPIA. But behavioural telemetry is not special category data and not biometric data, so the bar is materially lower than for facial analysis. If you aggregate to content level and never identify individuals, you may not be processing personal data at all.
What about eye tracking or gaze detection?
Gaze data is likely to be biometric data under the AI Act's broad definition. Whether it is caught depends on the inference drawn: attention position may be defensible, while inferring an emotional state from gaze patterns would not be. Treat it as a grey area and take advice before deploying in a workplace or education setting.
Can I still use facial analysis anywhere?
Yes. The Article 5(1)(f) prohibition covers workplace and education institutions specifically. Market research, advertising and media testing, consumer UX studies, and genuine medical or safety applications are outside its scope, though GDPR still applies wherever you process biometric data.

This is what Signals does

Behavioural telemetry, content-level reporting, no camera. Built for exactly this constraint.

See how Signals works
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