Emotion recognition in the workplace: the AI Act ban

Emotion recognition in the workplace and in education is prohibited under the AI Act, except where the system is intended for medical or safety reasons (Article 5(1)(f)). That prohibition has applied since 2 February 2025. Outside work and education, emotion recognition is not prohibited, but it is high-risk (Annex III point 1(c)), and since 2 August 2026 whoever deploys the system must inform the people exposed to it (Article 50(3)). Fatigue detection that does not infer emotions falls outside the definition of emotion recognition (Article 3(39)).

This page covers emotion recognition and biometric categorisation. The full overview of every prohibition is at prohibited AI practices under Article 5 AI Act. This page also accompanies the explanation of Article 50 AI Act, with the provider and deployer split per paragraph. For the contested assumptions behind emotion recognition, read our overview of the Dutch DPA report.

When is emotion recognition in the workplace prohibited?

This is the point organisations most often miss. Emotion recognition in the workplace and in education is a prohibited practice under Article 5(1)(f). The prohibition covers placing such a system on the market, putting it into service for this purpose and using it, so it reaches both the supplier and the employer or school that deploys it. An employer that infers the mood of employees from facial expressions or voice, or a school that does the same with pupils, falls under that prohibition. Transparency changes nothing here: informing does not make a prohibited practice permitted.

The only exception is a system intended to be put in place for medical or safety reasons. That exception is narrow: it concerns the purpose of the system, not an employer's general interest in productivity or customer satisfaction. An infringement can be fined up to 35 million euro or 7 percent of worldwide annual turnover (Article 99(3)).

A prohibition also applies to certain categories of biometric categorisation. Inferring race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation is prohibited under Article 5(1)(g), except for labelling or filtering lawfully acquired biometric datasets and in law enforcement. Other forms of biometric categorisation may be high-risk under Annex III; that is a separate dimension alongside the transparency duty, not a replacement for it.

What counts as emotion recognition, and what does not?

An emotion recognition system is an AI system for the purpose of identifying or inferring emotions or intentions of natural persons on the basis of their biometric data (Article 3(39)). Recital 18 gives as examples happiness, sadness, anger, surprise, disgust, embarrassment, excitement, shame, contempt, satisfaction and amusement.

According to the same recital, two things fall outside it:

• Physical states such as pain or fatigue. A system that detects fatigue in professional drivers or pilots to prevent accidents is not emotion recognition. Such a system is already outside the definition and therefore does not need the safety exception, as long as it does not infer emotions.
• The mere detection of readily apparent expressions, such as a frown, a smile or a raised voice, unless they are used to infer emotions.

What applies outside work and education?

Outside the workplace and education, emotion recognition is not prohibited. Think of analysing customers' voices in a customer call. Such a system is high-risk (Annex III point 1(c)); the high-risk obligations for the Annex III route apply from 2 December 2027. In addition, the Article 50(3) disclosure duty set out below already applies, as does the GDPR.

What exactly does Article 50(3) require?

Article 50(3) requires deployers of a system for emotion recognition or biometric categorisation to inform the natural persons exposed to it about the operation of the system. In addition, they must process the personal data in accordance with the GDPR and other applicable law.

The disclosure must be given at the latest at first exposure, and must be clear and accessible. So it is not a formality somewhere in a privacy statement, but information that actually reaches the person concerned at the moment the system is applied to them.

Who must inform?

The duty sits with the deployer: the organisation that uses the system under its own responsibility. Not the person being scanned, and under this provision not the provider of the system either.

Note the link with the GDPR: biometric data can be special categories of personal data (Article 9 GDPR). The transparency duty under the AI Act therefore comes on top of the requirements the GDPR already sets for processing that data.

Which exception applies to the disclosure duty?

There is an exception to Article 50(3) for systems permitted by law to detect, prevent or investigate criminal offences, with the corresponding safeguards. Outside that domain, the disclosure duty applies in full.

This exception is narrow and tied to a legal basis. Anyone relying on it must be able to show that the use is permitted by law and that the safeguards are observed.

The assumptions behind emotion recognition

There is a substantive reason to be strict here. According to the AI and Algorithms Netherlands report of July 2025 by the Dutch Data Protection Authority, emotion recognition rests on contested assumptions, carries high error rates and brings discrimination risk. The objective presentation of subjective and unreliable data is misleading: a system that presents an emotion as fact creates an appearance of certainty that the technology does not deliver.

That is precisely why the legislator prohibited these uses in the workplace and in education, and why transparency elsewhere is not enough to remove the underlying risk.

What should you do now?

Organisations that keep to this order avoid the classic mistake: building a tidy disclosure for a use that is not allowed at all.

Want to determine which obligation applies to your own use case? Walk through the Article 50 decision tree: per role and use case the paragraph, the steps from the guidelines and examples of good notices.