Skip to main content
Praxikon

Is emotion recognition allowed in the workplace under the EU AI Act?

··19 min read

No. Emotion recognition in the workplace and in education has been prohibited under the AI Act since 2 February 2025 (Article 5(1)(f)). The only exception is a system intended for medical or safety reasons, and that exception is narrow. The ban applies to the supplier that places such a system on the market and to the employer or education institution that uses it.

Imagine you are the privacy officer of a Dutch municipality. The customer contact centre wants to buy a new telephony module. It analyses the voices of staff during calls and shows the team leader a dashboard with their mood: calm, stressed, irritated. "It is for their wellbeing," says the department manager. Is that allowed? The short answer is no. The reason why determines what is possible.

In short

An AI system that infers emotions or intentions from biometric data, such as voice, face, posture or typing behaviour, may not be used in the workplace or in education (Article 5(1)(f) AI Act). That has applied since 2 February 2025, also to job applicants, trainees, volunteers and students, and regardless of consent or prior information. Systems for medical or safety reasons are excluded, such as a CE-marked medical device; general wellbeing or stress measurement is not. Outside work and education, emotion recognition is not prohibited, but it is in principle a high-risk system, with obligations from 2 December 2027, and since 2 August 2026 people must be informed (Article 50(3)). The GDPR continues to apply alongside. Breaching the ban can be fined up to EUR 35 million or 7 percent of worldwide annual turnover (Article 99(3)).

What the ban says

Article 5(1)(f) prohibits the placing on the market, the putting into service for this specific purpose, or the use of AI systems to infer emotions of a natural person in the areas of workplace and education institutions, "except where the use of the AI system is intended to be put in place or into the market for medical or safety reasons". Article 5(1)(f) has applied since 2 February 2025 (Article 113(a)); only the parts that Regulation (EU) 2026/1744 added to Article 5 apply from 2 December 2026.

The European Commission worked out the ban in its guidelines on prohibited AI practices (C(2025) 5052 final, adopted on 29 July 2025; the Commission approved the content on 4 February 2025). According to those guidelines, four conditions apply at the same time: placing on the market, putting into service or use; the system infers emotions; it does so in the workplace or in an education institution; and there is no medical or safety reason (points 242 and 243). The ban therefore reaches both the provider and the deployer, each within their own role. The guidelines are not binding; authoritative interpretation rests with the Court of Justice.

The ban leaves two ways out closed. Consent from the employee does not make it permitted, and neither does informing them: Article 5 has no exception for a notice or an agreement.

What counts as emotion recognition

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)). Biometric data is defined broadly in the AI Act: personal data resulting from specific technical processing relating to the physical, physiological or behavioural characteristics of a natural person (Article 3(34)). The guidelines read the ban the same way, so limited to inferences from biometric data (point 245).

Recital 18 gives as examples happiness, sadness, anger, surprise, disgust, embarrassment, excitement, shame, contempt, satisfaction and amusement. According to the guidelines that list is not exhaustive and the concept must be read broadly; a system that concludes a person shows an "angry attitude" is also covered (points 247 and 248). Three things fall outside:

  • Physical states, such as pain or fatigue. A system that detects fatigue in a professional driver to prevent accidents is not emotion recognition and therefore does not need the exception (Recital 18; guidelines point 261).
  • The mere detection of readily apparent expressions, such as a smile, a frown or a raised voice. Counting how often someone smiles is not emotion recognition; concluding that someone is happy or angry is (point 249).
  • Text without biometrics. A system that infers sentiment from written text, such as an email or chat message, uses no biometric data and falls outside the ban. A system that infers emotions from the way someone types, their facial expression or their body posture is covered (point 251).

A caution on text analysis: outside the ban is not the same as allowed. If an employer uses sentiment analysis of emails to assess staff, the GDPR applies and possibly the high-risk rules for monitoring and evaluating performance and behaviour (Annex III point 4(b), from 2 December 2027).

What counts as workplace and education

Workplace should be read broadly according to the guidelines: any physical or virtual space where people carry out tasks assigned by their employer or the organisation they work for, from offices, factories and warehouses to shops, cars or temporary work sites. Status does not matter: employee, contractor, trainee or volunteer. The ban also applies to candidates during selection and hiring, because the imbalance of power is already there (point 254). Emotion recognition in a recruitment procedure and during the probationary period is therefore prohibited.

Education covers public and private institutions at every level, including vocational schools and continuous training, without an age limit and both online and in person. Candidates in an admission procedure are covered too (point 255).

There is one nuance. If emotion recognition is used only for an employee's personal training, the guidelines allow it as long as the results are not shared with HR, cannot affect assessment or promotion, and the ban is not circumvented (point 254). The same goes for a role-play in education, such as training actors or teachers, as long as the results cannot affect evaluation or certification (point 255).

Example 1: the municipality as employer

Back to the contact centre. A module that infers from staff voices whether they are stressed or irritated, and shows that to the team leader, is prohibited. The guidelines give almost exactly this example: a call centre using webcams and voice recognition to track its employees' emotions, such as anger (point 254). That it is "for their wellbeing" does not help: measuring wellbeing, motivation or job satisfaction is not a medical reason (point 263), and general monitoring of stress levels at the workplace is not covered by the exception (point 257).

A second situation at the same municipality: HR wants to use a video platform for vacancies that scores applicants on "enthusiasm" and "confidence", inferred from face and voice. That is also prohibited, because the ban already applies at the recruitment stage (point 254). A system that only records and transcribes the video interview, without inferring emotions, is not caught by Article 5(1)(f). If the platform uses the recording to assess or filter candidates, it is in our reading a high-risk system for recruitment and selection (Annex III point 4(a)), from 2 December 2027.

What can the municipality do? A module that analyses the mood of residents on the phone is not caught by this ban, because residents are not at their workplace. The high-risk rules and the information duty apply instead; see below.

Example 2: a bank's call centre

A bank wants to improve its customer service. The same line as at the municipality: inferring the emotions of staff is prohibited. Inferring the emotions of customers, such as anger or impatience, to help staff cope with difficult calls is not prohibited under Article 5(1)(f) according to the guidelines (point 254). Such a system is in principle a high-risk system (Annex III point 1(c)), with obligations from 2 December 2027, and the bank must already inform customers about how it works (Article 50(3), since 2 August 2026). The employee can also be heard in the call; in our reading the bank must therefore ensure, technically and organisationally, that the system only analyses the customer.

The same applies to cameras in a bank branch. Cameras meant to flag suspicious visitors are not caught by this ban according to the guidelines, provided no employees are tracked and there are sufficient safeguards. Cameras that track the emotions of counter staff are prohibited; the guidelines give that example for supermarket staff (point 254).

Example 3: the school and the university of applied sciences

A university of applied sciences is considering software that measures through the webcam how interested and attentive students are during online lectures. That is prohibited; the guidelines name it explicitly (point 255). The same goes for emotion recognition in an admission test and for a language app with emotion recognition that the school requires its pupils to use. A system that tracks the emotions of both teachers and students hits both parts of the ban: workplace and education.

Proctoring in online exams is a borderline case. Software that only tracks gaze to detect unauthorised material does not infer emotions and is not caught by the ban. If the same software also detects anxiety or emotional arousal, it is (point 255). Monitoring prohibited behaviour during tests is also a high-risk use (Annex III point 3(d)), and if the tool recognises students by their face, it also processes biometric data for the purpose of unique identification (Article 9(1) GDPR); more on that in is your AI system on the Dutch DPA's mandatory DPIA list.

HR and recruitment in general

For every employer, public or private, this means in practice: no emotion analysis in video interviews, no "mood meter" in team meetings, no wellbeing dashboard based on voice or face, and no burnout detection. The guidelines explicitly name monitoring the emotional tone in hybrid meetings as prohibited (point 254), and a system intended to detect burnout or depression at the workplace is not covered by the exception either (point 257). Anyone who finds such a feature in an existing HR suite has to switch it off; use is prohibited, not only purchase.

The exception: medical or safety, read narrowly

According to the guidelines, the exception for medical or safety reasons must be interpreted narrowly (point 256):

  • Medical covers at least therapeutic use, which the guidelines read as the use of CE-marked medical devices; measuring general wellbeing does not qualify (points 256 and 257). An aid that helps an employee with autism understand the emotions of colleagues can fall under it (point 263).
  • Safety only concerns the protection of life and health, not protecting property against theft or fraud (point 258). An employer may only measure stress or lack of concentration where that poses a specific danger, such as when operating dangerous machines, and may not then use the data to assess performance (point 263).
  • Strictly necessary and proportionate, limited in time, persons and scale, with safeguards such as a prior written and reasoned expert opinion, and only if no less intrusive alternative exists (point 259). The data may not be used for any other purpose (point 260).

A system that falls under the exception is not done yet. It remains a high-risk system with an information duty (point 263), the GDPR continues to apply (point 262), and the Union and Member States may maintain or introduce rules that are more favourable to workers, and allow collective agreements that are more favourable to workers (Article 2(11); guidelines point 264). In the Netherlands, a works council, where there is one, also has a right of consent over arrangements for observing or monitoring the attendance, behaviour or performance of staff (Dutch Works Councils Act, section 27(1)(l)). That act also covers staff with a public-law appointment, so it applies at a municipality as well. That right of consent only lapses in so far as a collective agreement already regulates the matter in substance (section 27(3) of the Works Councils Act).

Outside work and education: high risk and an information duty

Emotion recognition of customers, residents, patients or visitors is not prohibited under Article 5(1)(f) (point 266). However:

  • It is a high-risk system (Annex III point 1(c)), in so far as its use is permitted under relevant Union or national law. The obligations for providers and deployers apply from 2 December 2027, since the amendment by Regulation (EU) 2026/1744.
  • The deployer must already inform the people exposed to the system about how it works and comply with the GDPR (Article 50(3), since 2 August 2026). More on that in emotion recognition and the Article 50 information duty.
  • In certain cases it may still be prohibited as manipulation or exploitation of vulnerabilities (Article 5(1)(a) and (b)), if all conditions of those bans are met (point 266).

What the GDPR adds

The ban stands apart from the GDPR, and the GDPR continues to apply in full (guidelines point 262). Four points matter here.

Biometric data is defined more narrowly in the GDPR. The GDPR speaks of biometric data where it allows or confirms unique identification (Article 4(14) GDPR), and the prohibition on special categories covers biometric data "for the purpose of uniquely identifying a natural person" (Article 9(1) GDPR). The guidelines point to that difference with the broader AI Act definition (footnote 160). If the system also uses face or voice to uniquely identify the person, that prohibition applies. Article 29 of the Dutch GDPR Implementation Act only makes an exception where this is necessary for authentication or security, which is not the case for emotion recognition, and explicit consent is rarely free in an employment relationship. If an inferred outcome says something about a person's mental health, such as a burnout risk, it is also health data (Article 4(15) GDPR), which is likewise covered by Article 9(1) GDPR.

Employee consent is rarely free. The EDPB considers it problematic for employers to rely on the consent of current or future employees, because the dependency in the employment relationship means it is unlikely to be freely given (Guidelines 05/2020 on consent, paragraph 21). If the employer is a public authority, this weighs even more: where there is a clear imbalance, in particular where the controller is a public authority, consent is not a valid legal basis (Recital 43 GDPR). Member States may lay down more specific rules for the employment context (Article 88 GDPR).

A DPIA is the obvious step. The list of the Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP) names, among others, large scale and/or systematic monitoring of employee activities (category 11), profiling such as the assessment of work performance (category 15), systematic observation of behaviour (category 16) and large scale and/or systematic processing of biometric data to identify a person (category 17) (Stcrt. 2019, 64418). How to test that per application is set out in is your AI system on the Dutch DPA's mandatory DPIA list. A DPIA does not make a prohibited practice permitted; it only matters for what remains, such as emotion recognition of customers.

The AP is critical of the technology itself. In its AI and algorithms report of July 2025 the AP concludes that emotion recognition rests on contested assumptions and must be handled with great caution. A fuller discussion is in our overview of the Dutch DPA report.

Fines and supervision

A prohibited practice carries the highest fine ceiling in the AI Act: up to EUR 35 million or, for an undertaking, up to 7 percent of worldwide annual turnover if that is higher (Article 99(3)). For SMEs and start-ups the lower of the two applies (paragraph 6). Whether and to what extent a public body such as a municipality can be fined is for each Member State to decide (paragraph 8).

The Netherlands has not settled that in law yet, and no supervisory authority has been designated for the AI Act. On 20 April 2026 the government put the Dutch AI Act implementing bill out for public consultation (Rijksoverheid). In that draft, the AP, the AFM and De Nederlandsche Bank, each for its own category of organisations, are the intended supervisors for the prohibited practices (draft bill, Article 2.2(1)(b)). According to the government, the AP and the Dutch Authority for Digital Infrastructure (RDI) will have a coordinating role. The draft does not rule out fines for public bodies: according to the draft explanatory memorandum, a municipality can also be an offender under the Dutch General Administrative Law Act (section 4.5.4.2). Until the act is in force, no Dutch supervisor can impose an AI Act fine. The ban itself applies directly in the meantime, and the AP can already act under the GDPR against unlawful processing of personal data, including against a municipality (Article 18 of the Dutch GDPR Implementation Act). More in AI Act supervision in the Netherlands and AI Act and GDPR fines for municipalities.

The three step test

  1. Does the system infer emotions or intentions from biometric data? Think of voice, face, posture, movements, heart rate or typing behaviour. If it only concerns text, fatigue or pain, or counting visible expressions without a conclusion about an emotion, it falls outside the ban. Continue with the GDPR and Annex III.
  2. Is the system aimed at people at their workplace or in education? That covers employees, contractors, trainees, volunteers, applicants, pupils, students and admission candidates. If so, the system is prohibited unless step 3 applies. A system aimed at visitors that only incidentally affects staff falls outside the ban according to the guidelines, provided the deployer ensures that staff are not screened (point 270). If not, it is a high-risk system with an information duty.
  3. Is the system intended for a medical or safety reason in the narrow sense? That means a CE-marked medical device, or protecting life and health against a specific danger, strictly necessary and without other use of the data. Wellbeing, productivity, customer satisfaction or fraud prevention do not count.

Want to run this for your own application? With the case check you describe the application in a few questions and see which AI Act and GDPR obligations come into view.

Frequently asked questions about emotion recognition in the workplace

Sources

EUR-Lex: Regulation (EU) 2024/1689 (AI Act), Official Journal, Recitals 18 and 44 (13 June 2024, accessed 16 November 2026)
European Commission: Guidelines on prohibited AI practices, C(2025) 5052 final, chapter 7 (points 239 to 270); content approved as C(2025) 884 final (29 July 2025 (content 4 February 2025), accessed 16 November 2026)
EUR-Lex: Regulation (EU) 2016/679 (GDPR), Articles 4, 9 and 88 and Recital 43 (27 April 2016, accessed 16 November 2026)
EDPB: Guidelines 05/2020 on consent under Regulation 2016/679, version 1.1 (4 May 2020, accessed 16 November 2026)
Autoriteit Persoonsgegevens: Decision on the list of processing operations requiring a DPIA, Stcrt. 2019, 64418, categories 11 and 17 (19 November 2019, accessed 16 November 2026)
Autoriteit Persoonsgegevens: AP: Emoties herkennen met AI 'dubieus en risicovol' (15 July 2025, accessed 16 November 2026)
Overheid.nl: Dutch GDPR Implementation Act (UAVG), Articles 18 and 29 (accessed 16 November 2026)
Overheid.nl: Dutch Works Councils Act (WOR), section 27 (accessed 16 November 2026)
Rijksoverheid: Kabinet zet stap met toezicht op Europese AI-regels (20 April 2026, accessed 16 November 2026)
Internetconsultatie: Dutch AI Act implementing bill (Uitvoeringswet verordening artificiële intelligentie), draft, Article 2.2 (consultation 20 April to 1 June 2026, accessed 16 November 2026)

Newsletter

Every Tuesday, the AI Act week ahead in 5 minutes

A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.

Practical and short · No spam · One-click unsubscribe