Research

The limits of facial emotion recognition

We sell a product with a facial analysis mode, which makes this an awkward page to write. It is also the page most likely to save you from buying the wrong thing — from us or anyone else.

9 min read
In short

Facial emotion recognition performs well on posed, frontal, well-lit expressions and considerably worse on spontaneous ones. Reported engagement classification accuracy of around 91.5% from facial data alone rises to 94.6% when behavioural signals are added. The underlying assumption — that discrete emotions map reliably to facial configurations — is scientifically contested.

Why a vendor is publishing this

Emotuit has a facial analysis configuration, so we have a commercial interest in you believing this technology works. We are publishing the limitations anyway, because the alternative is you discovering them after purchase — and because a vendor who only publishes strengths should be read accordingly. That includes us on every other page.

The headline numbers

91.5%
Engagement classification accuracy from facial expression alone
OUCI/DNTB online learning study
94.6%
The same task with behavioural signals added
OUCI/DNTB online learning study
+3.1pt
The entire contribution of the facial layer
Derived

Two things are worth noticing about those numbers.

First, 91.5% sounds excellent — until you see that adding non-biometric behavioural signals gets you to 94.6%. The facial layer contributes roughly three percentage points on top of data you can collect from browser events.

Second, and more importantly: accuracy against what? These figures measure agreement with a labelled dataset. The label is usually either a posed expression or a human annotator's judgement. Neither is the same as the person's actual internal state, which is unobservable. Every accuracy figure in this field is really a measure of agreement with a proxy.

Posed versus spontaneous

This is the single largest gap between marketing claims and field performance.

Most benchmark datasets consist of posed expressions — participants instructed to display anger, surprise, disgust. Posed expressions are exaggerated, prototypical, frontal and well-lit. Classifiers do very well on them.

Spontaneous expressions, produced by someone genuinely reacting to a training module, are quite different: subtler, briefer, frequently asymmetric, often blended, and regularly absent entirely. A person can be deeply frustrated and display almost nothing.

This is why baseline calibration and change-over-time measurement matter more than absolute classification. Tracking that a person's affect shifted during a segment is far more defensible than asserting what they felt.

The contested science

The technology rests on an assumption from Ekman's Facial Action Coding System: that a small set of basic emotions map to recognisable, largely universal facial configurations.

That assumption has been seriously challenged. A major review led by Lisa Feldman Barrett concluded that the mapping between facial configurations and emotional states is far weaker and more context-dependent than the basic-emotion model implies — people scowl when angry considerably less often than the model predicts, and scowl for many reasons other than anger.

The practical implication is not that the technology is useless. It is that the inference is probabilistic and context-dependent, and confident individual-level claims are not supported.

Regulators have noticed. The ICO has publicly warned about biometric technologies claiming to detect emotional states, citing the weak scientific footing and the discrimination risk when such systems do not work as claimed.

Demographic variance

Performance is not uniform across populations. Documented sources of variance include:

Per-user baseline calibration mitigates some of this, because it measures deviation from that individual's own neutral rather than a population norm. It does not eliminate it.

Physical failure modes

More mundane, and in practice they cause most day-to-day error:

Condition Effect
Poor or backlit lighting Landmark detection degrades sharply
Glasses, especially with glare Eye-region features obscured
Facial hair Mouth-region action units obscured
Off-axis camera angle Laptop cameras look up; models expect frontal
Low-resolution webcam Subtle action units unresolvable
Face partially out of frame Detection fails intermittently
Masks Lower face entirely unavailable

None of these are exotic. A learner on a laptop in a dim room wearing glasses is an ordinary case, and it is close to the worst case for this technology.

What it is genuinely good at

Having spent seven sections on limitations, the honest other side:

Relative change within a person. Did this individual's affect shift between segment A and segment B? With a per-user baseline this is reasonably reliable, and it is the measurement most engagement use cases actually need.

Aggregate patterns across many people. Individual error partially cancels at scale. "This segment produced more negative affect across 400 viewers" is far more defensible than any single-person claim.

Detecting presence and gross state changes. Whether someone is there, whether they turned away, whether something changed markedly.

Timing. Even where the emotion label is uncertain, when the affective change occurred is often accurate, and for content diagnostics timing is frequently the whole answer.

What this means for buyers

Ask what the accuracy figure is measured against. Posed dataset or spontaneous field data? The difference is large and the answer is rarely volunteered.

Be sceptical of individual-level claims. Aggregate and relative measurement is defensible. "This learner was frustrated at 14:32" is not.

Never make consequential decisions about individuals from it. Not performance management, not academic integrity, not hiring. The error rate is not low enough and the failure modes are not evenly distributed.

Check whether you can lawfully use it at all. In EU workplace and education contexts, emotion inference from biometric data has been prohibited since February 2025. This question comes before the accuracy question.

Ask what you lose without it. In our case: roughly three percentage points of engagement classification accuracy, and emotional texture — the ability to distinguish confusion from boredom. For most content diagnostics, behavioural signals answer the question without any of the above.


The summary we would give a prospect on a call: facial emotion recognition is a real technology with real uses, oversold by most of its vendors, resting on a contested scientific premise, and — for the specific job of working out which part of your training loses people — largely unnecessary.

Frequently asked questions

Is facial emotion recognition accurate enough to use?
It depends entirely on what you use it for. For detecting relative change in a single person over time — did this segment produce more negative affect than that one — it is reasonably reliable. For asserting an individual's emotional state at a moment, it is not. Most vendor marketing implies the second while the evidence supports the first.
Why do vendors quote such high accuracy figures?
Because those figures usually come from benchmark datasets of posed expressions — actors instructed to display anger, joy and so on. Posed expressions are exaggerated and prototypical, and classifiers do very well on them. Spontaneous expressions in real settings are subtler, and performance drops.
Does per-user baseline calibration fix the accuracy problem?
It fixes one specific and important problem: measuring against a person's own neutral face rather than a population average, so naturally downturned features are not read as sadness. It does not fix the deeper issue of whether facial configurations map reliably to internal states at all.
Should I avoid facial emotion recognition entirely?
In EU workplace and education contexts you have no choice — it is prohibited. Elsewhere it is a legitimate tool for aggregate, relative measurement in market research and media testing. What it should not be used for is consequential decisions about individuals.

The configuration that avoids all of this

Signals measures engagement from behavioural telemetry alone — no camera, no emotion inference, no accuracy ceiling to argue about.

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