Can an employer read workers’ emotions by calling it wellness?

On 12 July 2024, the EU published its AI Act in the Official Journal. Article 5 prohibits AI systems used to infer people’s emotions in workplaces, except where the system is intended for medical or safety reasons. The question is whether an employer can put an AI wellness wristband inside that exception simply by describing emotional monitoring as employee wellbeing.

The text provides no wellness exception. It distinguishes prohibited emotion inference from particular medical and safety purposes. A reassuring product name does not settle which side of that boundary a system occupies.

What does Article 5 prohibit?

Regulation (EU) 2024/1689, Article 5(1)(f), prohibits placing these systems on the market, putting them into service for this specific purpose, or using them to infer emotions in workplaces and educational institutions. Its exception is for systems intended for medical or safety reasons.

The prohibition is therefore not confined to the vendor. It reaches use. An employer cannot answer the legal question merely by producing a supplier’s description of its product.

The European Commission’s implementation timetable places the prohibitions in effect from 2 February 2025. The later application dates for other AI Act obligations do not postpone this ban.

Recital 44 explains the distinction. It describes the scientific shortcomings of emotion inference and the imbalance of power at work. It then says the prohibition should not cover systems placed on the market strictly for medical or safety reasons, giving therapeutic use as an example. That is narrower than any use an employer regards as beneficial.

What counts as reading emotions?

Article 3(39) defines an emotion recognition system as an AI system that identifies or infers emotions or intentions from biometric data. Article 3(34) defines biometric data through technical processing of physical, physiological or behavioural characteristics.

The relevant question for a wristband is therefore not whether it measures something about a worker. It is whether the system uses qualifying data to infer an emotional state. Measurement and inference are different operations.

Recital 18 makes that distinction explicit. It excludes physical states such as pain or fatigue, including fatigue detection for professional pilots or drivers to prevent accidents. It also excludes merely detecting readily apparent expressions, gestures or movements, unless those observations become inputs for identifying or inferring emotions.

A system that detects a smile is not necessarily an emotion recognition system. A system that treats the smile as evidence of happiness crosses into a different function. Likewise, the fatigue exclusion does not establish a general exemption for anything attached to the word health.

The Commission published guidelines on prohibited AI practices on 4 February 2025. Those guidelines assist interpretation but are nonbinding. The controlling distinction remains the regulation’s wording, not a vendor’s preferred category.

Why does the purpose matter?

The medical exception is real. Recital 44 expressly contemplates therapeutic use. It would be wrong to read Article 5 as banning every clinical application involving a person who also happens to be an employee.

But workplace wellbeing is not itself a defined substitute for medical purpose. Calling a dashboard preventive medicine does not establish what condition it addresses, what its clinical function is, or how its outputs are used.

The Medical Devices Regulation, Regulation (EU) 2017/745, supplies a useful comparison. Article 2(1) defines medical devices by specified medical purposes, including diagnosis, prevention, monitoring and treatment of disease. Recital 19 distinguishes software intended for medical purposes from software intended for lifestyle and wellbeing purposes.

The Commission’s published medical-device software guidance, MDCG 2019-11, develops that qualification framework. It asks about intended purpose and the software’s function, rather than treating all health-adjacent software as medical software.

The AI Act does not simply incorporate medical-device status as its exception test. The comparison nevertheless exposes the weakness in the branding argument: EU medical-device law itself does not treat wellbeing and medicine as interchangeable. An employer claiming a medical exception needs a medical purpose, not just medical vocabulary.

Can an employer call it safety?

Council Directive 89/391/EEC requires employers to ensure workers’ safety and health in every aspect related to work. Article 6 requires risk prevention and assessment, alongside measures addressing working conditions and organisation.

That duty is broad. Article 5’s exception cannot sensibly mean that citing the duty makes every emotional surveillance system permissible. Otherwise, the workplace prohibition would depend on whether an employer remembered to mention safety.

The two texts require different questions. The directive asks employers to address occupational risks. The AI Act asks whether this particular prohibited technology is intended for a permitted medical or safety reason.

A claimed safety purpose must therefore be examined as a purpose of the system, not accepted as a general description of the employer’s responsibilities. The legal distinction matters most where the same emotional output could serve different ends. An inference used to assess an occupational hazard is not the same stated purpose as an inference used to rank employee enthusiasm.

And fatigue monitoring does not prove the broader case. Recital 18 places qualifying physical-state detection outside emotion recognition. An exclusion from the definition is not permission to infer any emotion that might interest a safety department.

Does employee consent fix it?

Article 5(1)(f) contains no exception for workers who agree to emotion recognition. A consent form cannot supply an exception that the prohibition does not contain.

Consent also raises a separate data-protection question. Article 7 of the General Data Protection Regulation sets conditions for valid consent. The European Data Protection Board’s Guidelines 05/2020 explain why dependence within employment usually makes freely given consent difficult to establish. They do not say employee consent is impossible in every circumstance.

If a system genuinely qualifies for the medical or safety exception, the remaining data processing still needs its own legal basis. GDPR Article 5 requires purpose limitation and data minimisation. Article 9 imposes additional conditions where processing concerns health data or other covered special categories.

Article 35 requires a data protection impact assessment for processing likely to create a high risk to people’s rights and freedoms. Passing the AI Act’s exception test does not perform that assessment or establish a GDPR legal basis. These are separate obligations, not alternative routes to permission.

Why is a benevolent purpose insufficient?

The prohibition addresses more than malicious employers. Recital 44 identifies limited reliability, specificity and generalisability as shortcomings of emotion inference. A benevolent objective does not establish that the inference is sound.

The 2019 review “Emotional Expressions Reconsidered,” by Lisa Feldman Barrett and colleagues, examined the evidence for inferring emotion from facial movements. It found substantial variation across people, situations and cultures. Its scope is facial movement, not every physiological signal a wearable might collect.

That distinction still matters for the wellness argument. Naming a desired benefit does not validate the measurement used to pursue it. The EU Charter separately protects private life and personal data in Articles 7 and 8. An employer’s preferred account of its intentions does not displace those protections.

Does the label change the answer?

No. An employer cannot make prohibited workplace emotion inference lawful by calling it wellness. The system must either fall outside the relevant definition or qualify for the actual medical or safety exception, with separate data-protection obligations still applying. Wellness describes an aspiration. Article 5 requires a permitted purpose.