
Since 2 February 2025, employers in the European Union can no longer use an AI system that infers their employees' emotions in the workplace, except for medical or safety reasons. That is Article 5 of the EU AI Act. Reading anger in a voice, scoring stress from tone, judging a server's cheerfulness from their face: all of it is covered. Spotting a driver's fatigue, answering a question spoken out loud or noticing that someone is smiling is not. This article gives general guidance under EU law, not legal advice: for your own situation, talk to your data protection officer or an employment lawyer.
Friday, 11:20 pm, in the storeroom
The last ticket has gone out. Serge, a server for three years, sits down on a crate of bottles, apron still tied, staring into space. Coline, the floor manager, sees him as she walks past.
That morning, a sales rep had pitched her a "caring listening" tool: an earpiece for every server, continuous tone analysis, and a dashboard showing "stress level", "energy" and "friendliness" for each person. "You'll see who is about to burn out before they quit," he had said.
Coline does not need a chart to know that Serge covered the back section alone tonight. She hands him a bottle of water: "The back section, tonight, was it too much?" The next day she turns the offer down. She does not yet know the tool is also banned.
(Fictional scene and characters.)
What Article 5 actually says
The text and the date
Article 5(1)(f) of Regulation (EU) 2024/1689 prohibits three things: placing on the market, putting into service for this specific purpose, and using AI systems to infer the emotions of a natural person in the workplace and in education institutions. The only exception is medical or safety reasons.
The ban has applied since 2 February 2025, together with the other prohibited practices in Chapter II (Article 113). Regulation (EU) 2026/1744, known as the Digital Omnibus on AI and published in the EU Official Journal on 24 July 2026, added new prohibitions to Article 5 but did not soften this one: the consolidated text of 27 July 2026 keeps it as it was.
What the regulation means by emotion
Article 3 defines an emotion recognition system as one that identifies or infers the emotions or intentions of people on the basis of their biometric data: face, voice, gestures. Recital 18 gives examples: happiness, sadness, anger, surprise, shame, satisfaction. It leaves out two things. Physical states, such as pain or fatigue. And the mere detection of an obvious expression (a smile, a raised voice) as long as no emotion is inferred from it.
What the Commission's guidelines add
On 4 February 2025, the Commission published its guidelines on prohibited AI practices, formally adopted on 29 July 2025 (C(2025) 5052). They are not binding, but they show how the Commission reads the text. Four points matter to an employer:
- the ban covers inferences drawn from biometric data, in line with the definition in Article 3;
- the "workplace" also covers job candidates during recruitment;
- the ban applies to the business using the system, even if the vendor's contract excludes that use;
- the medical or safety exception must be read narrowly.
What is covered in a restaurant, and what is not
The guidelines give examples that translate well. A call centre using webcams and voice recognition to track its employees' anger: prohibited. The same call centre tracking customers' impatience: not covered. A supermarket filming its employees to measure their cheerfulness: prohibited.
| Situation in a restaurant | What the tool does | Banned under Article 5? |
|---|---|---|
| An earpiece analyses servers' tone to score their friendliness | Infers an emotion from the voice | Yes |
| A dashboard shows cooks' "stress level" during the rush | Infers an emotional state from the voice | Yes |
| A dining-room camera reads the team's faces to measure "satisfaction" | Infers an emotion from the face | Yes |
| A video interview tool rates a candidate's enthusiasm | Infers a candidate's emotion | Yes |
| A sensor detects a delivery driver's fatigue and suggests a break | Detects a physical state, for safety | No, but data protection and employment law apply |
| A tool analyses customers' mood in reviews or on the phone | Concerns customers, not employees | No, but other rules apply |
| A voice assistant answers "what's the allergen sheet for today's special?" | Understands a request, infers no emotion | No |
Two rows need a note. Analysing customers' emotions is not a free-for-all: since 2 August 2026, Article 50 requires anyone using an emotion recognition system to inform the people exposed to it, and the GDPR applies. And a tool that infers an employee's mood from written messages alone probably falls outside Article 5, since no biometric data is involved. It is still subject to data protection law and, in France, to the Labour Code's proportionality rule (Article L1121-1, see the guide to AI, voice and privacy at work).
Why lawmakers drew the line
The science is shaky
Recital 44 says it plainly: there are serious concerns about the scientific basis of these systems. Emotions are expressed differently across cultures, situations and individuals. A voice rising mid-service can mean anger, fatigue, the noise of the extractor hood or simply an accent.
The power gap is real
The same recital grounds the ban in the imbalance of power at work and the risk of discriminatory outcomes. A mood score produced by a machine quickly becomes an argument in a review or a turned-down promotion. And it shows fatigue without saying anything about its cause: that night, a section too big for one person.
The penalties
Article 99 sets the highest ceiling in the regulation for prohibited practices: an administrative fine of up to 35 million euros or 7% of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs and start-ups, the cap is whichever of the two is lower. These penalty rules have applied since 2 August 2025, and each Member State sets its own detailed regime within those limits.
The fines target both the vendor and the employer using the tool. They come on top of data protection law and national employment law. The French rules (prior notice to employees, consulting the works council from 50 staff) are summarised in a microphone at work: what the law says in France.
What an employer should check with a vendor
- Ask in writing: does the tool identify or infer emotions, mood, stress or intentions from voice, face or gestures? Insist on a signed answer, not a brochure.
- Watch the labels. "Wellbeing", "engagement", "team climate", "sentiment analysis": ask what is calculated, and from what.
- Ask what the tool does with the voice. Understanding a request to answer it and analysing the pitch or rhythm of a voice are two different kinds of processing.
- If the vendor cites health or safety, ask which precise, documented reason justifies the exception. A promise to "prevent psychosocial risks" does not create one.
- Do not settle for a switched-off option. According to the guidelines, the ban applies to you as the user, whatever the vendor's contract says.
- Check recruitment tools. A video interview that rates a candidate's enthusiasm or confidence is in scope.
- Have the absence of emotion analysis written into the contract, then tell your team before going live.
Common mistakes
- Thinking the ban only covers cameras. The call centre example shows it: a voice analysed by a tool is covered, so earpieces and microphones are too.
- Assuming the employee's consent is enough. The text bans the use itself, and under the GDPR an employee's consent to their employer is rarely considered freely given.
- Confusing fatigue with emotion. A system that spots a physical state for safety is outside Article 5, but it is still data processing that needs a justification.
- Keeping a tool bought before 2025. The purchase date changes nothing: using it has been banned since 2 February 2025.
What to measure
- The share of your tools (scheduling, time clock, recruitment, assistant, cameras) for which you hold a written answer from the vendor on emotion analysis.
- The number of features labelled "mood", "stress" or "engagement" in your software, and what they really calculate.
Where Tsuno fits
Tsuno connects the facts; people decide. Its trust page lists what it refuses to do, including inferring a person's motivation, honesty, loyalty or mood from their messages, and rating a person in place of their manager. It puts dated facts side by side: training completed, checklist done or not, incident, hours.
San, the voice assistant for teams, is available as a demo (San Voice). It acts within the rights of the person using it, evaluates no one, and every action happens in two steps: San prepares and summarises, the person confirms. A good voice tool processes the voice to understand a request, never to read a state of mind. See the Tsuno San page.
Further reading
The full framework is in our pillar guide, AI, voice and privacy at work: the guide. For the rest of the EU timeline, read AI at work: what changes for employers in 2027. The legal framework for a microphone in a French workplace is in a microphone at work: what the law says in France. For our own limits, see our red lines: the ethics of AI at work and what San never does.
Key takeaways
Since 2 February 2025, no AI should infer an employee's emotions at work in the EU, except for a documented medical or safety reason. Ask every vendor in writing, and keep the question that matters for yourself: "Tonight, was it too much?"
Frequently asked questions
Can an employer use AI to analyse employees' emotions?
Not in the European Union. Since 2 February 2025, Article 5(1)(f) of Regulation (EU) 2024/1689 has prohibited placing on the market, putting into service for this purpose and using AI systems that infer the emotions of a person in the workplace. The only exceptions are medical or safety reasons.
Is detecting stress in an employee's voice banned?
A tool that infers an emotional state such as anger, irritation or anxiety from an employee's voice falls within the ban. The European Commission's guidelines give the example of a call centre tracking its employees' anger through voice recognition or webcams: that is prohibited.
Is detecting worker fatigue with AI banned?
Fatigue and pain are physical states, not emotions, according to Recital 18 of the AI Act. A system that spots a driver's fatigue to prevent an accident is therefore outside this ban. It still has to comply with data protection and employment law.
What are the fines for using emotion recognition on employees?
Article 99 of the AI Act sets fines for prohibited practices of up to 35 million euros or 7% of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs and start-ups, the cap is whichever of the two is lower.
Are the Commission's guidelines on prohibited AI practices binding?
No. Published on 4 February 2025 and formally adopted on 29 July 2025, they explain how the Commission reads Article 5, but they are not legally binding. The authoritative interpretation belongs to the Court of Justice of the European Union.