Detecting emotions by computer
Software announces the affective state of a customer, a driver or a candidate from a video. The European regulation adopted in 2024 prohibits part of these uses.
What these systems do
An emotion recognition system extracts characteristic points of the face from an image, or parameters of the voice from a recording, and compares them with a model trained on labelled examples. It then produces a label: joy, anger, stress, engagement.
The technical performance is real on test data: the model recovers the labels of the examples it was taught. That is not the question. The question is what those labels are worth on a real person, in a real situation, with no instructions.
The scientific objection
The review published in 2019 by Lisa Feldman Barrett and co-authors, covering several hundred studies, concludes that the same facial configuration accompanies different states and that the same state is expressed in varied ways. The correspondence is too loose to ground an inference about an individual.
Physiological measures do no better. Skin conductance and heart rate inform well about the level of activation, badly about its nature: no sensor distinguishes anger from enthusiasm.
The market's retreats
Several players have backed off. An American publisher of automated video interviews, widely used in recruitment, announced in 2021 that it was abandoning facial analysis in its assessments, judging its contribution insufficient in view of the objections raised.
An older precedent concerns aviation security: a US public report published in 2013 concluded that the behavioural observation programme for passengers did not identify dangerous people better than chance, and recommended limiting its funding.
The European prohibition
The European regulation on artificial intelligence, adopted in 2024, places among prohibited practices systems intended to infer a person's emotions in the workplace and in educational establishments, save for medical or safety reasons. That text's prohibitions came into application before the rest of the framework.
The choice of scope is telling: it targets contexts where the person observed cannot escape observation. An employee does not leave their post, a pupil does not leave their classroom, and consent there is largely formal.
The distinction the debate lacks
A signal too noisy to say anything about an individual can remain informative about a population. An average variation of activation in a room informs about what is being shown there; the same measurement on one viewer informs about nothing.
That is the difference between a statistical use and a decisional use. The strongest objections target the second, when an uncertain inference is used to take a decision that commits someone: a hiring, a mark, an inspection.
Key points
The correspondence between facial configuration and inner state is too loose to ground an individual inference.
Physiological sensors measure activation, not the nature of the emotion.
The 2024 European regulation prohibits these systems at work and in education, outside medical or safety reasons.
This article describes documented mechanisms and disagreements. It recommends no position, no party and no vote, and commits to no reading other than that of the work it cites.
And you, where do you stand?
The Civix test situates your convictions on six dimensions and matches you with one of the twelve profiles, without ever telling you what to think.
Start my test