Is emotion recognition AI banned at work under the EU AI Act?

In the workplace, yes, with narrow exceptions. Article 5 of the EU AI Act prohibits AI systems that infer the emotions of people in workplaces and education institutions, except where the system is used for medical or safety reasons, and the prohibition has applied since 2 February 2025. It was not deferred by the Digital Omnibus.

The question matters in 2026 because monitoring and productivity suites are quietly growing emotion features, and because the deferral of the high-risk regime has left many buyers assuming the whole Act is parked until December 2027. The prohibitions are not parked. They have been live for well over a year.

What exactly does Article 5 prohibit at work?

Using AI systems to infer the emotions of a natural person in the areas of workplace and education institutions. The carve-out is narrow: systems put in place for medical or safety reasons. An AI that flags fatigue in a safety-critical driver can argue the safety exception; an AI that scores the enthusiasm of a sales team from video calls is squarely what the prohibition was written for. The test is not what the vendor calls the feature but what the system infers about the person.

Was the workplace emotion ban deferred to 2027?

No. The Digital Omnibus moved the high-risk Annex III obligations, once due on 2 August 2026, to 2 December 2027, but the Article 5 prohibitions were untouched and have applied since 2 February 2025. Outside workplaces and education, emotion recognition is not banned but sits in Annex III as high-risk, and that regime is on the deferred timeline. Inside the workplace, the ban stands today.

Where is the line between sentiment analysis and emotion inference?

Identifiability and target. Running sentiment across aggregated, anonymised survey feedback to learn whether a policy landed badly is analysis of text the organisation was voluntarily given, at group level. Inferring the emotional state of an identified worker from their video, their voice or their keystrokes is inference about a person, and in the workplace that is prohibition territory. A workable test: if the output can be traced to a named individual and describes their emotional state, treat it as inside the ban unless a genuine medical or safety purpose is documented.

Is your monitoring software already doing this?

Possibly, and the deployer carries exposure alongside the vendor. Monitoring suites have been adding engagement scoring, tone analysis and attention tracking, sometimes enabled by default, sometimes arriving silently in a feature update. An organisation can drift into Article 5 territory through a feature it never consciously adopted. The control is a feature audit: list every analytic the monitoring stack runs on staff, trace each to a documented purpose, and disable anything that infers emotional state without a medical or safety justification. Repeat the audit after every vendor update, because the feature set is not static.

Does the ban reach UK workplaces?

Not directly, because the EU AI Act is not UK law. UK employers are not unconstrained, though. UK GDPR requires a lawful basis, necessity and proportionality for workplace monitoring, Article 8 of the European Convention on Human Rights protects private life in ways that reach the workplace, and the ICO's guidance on monitoring at work sets clear expectations on transparency and necessity. A UK employer with EU staff or EU operations can also find the EU ban applying to part of its workforce. The direction of travel on both sides of the Channel points the same way.

What does a sovereign deployment change here?

It makes the boundary provable. Whatever analytics an organisation legitimately runs on staff data should run inside its own perimeter, not in a vendor cloud where features change without notice. On Mickai, our Sovereign Intelligence Operating System, which runs offline on operator-owned hardware, every analytic is an explicit, sealed action: a post-quantum signed audit ledger records what was inferred, from what data and under whose authority, and the record verifies offline. That cuts both ways, usefully. It evidences the analytics the organisation does run, and it evidences that prohibited inference was never configured at all, which is the record an employer wants when a works council, a regulator or an employment tribunal asks the question.

“A prohibition is easiest to defend when the sealed record shows exactly what the system inferred and what it never did.”

The full architecture behind that sealed record is set out at /sovereign-ai, and the film at /film shows the system in operation.

Frequently asked questions

Can I use AI to analyse sentiment in our staff engagement survey?

Aggregated, anonymised sentiment analysis of voluntary written feedback is generally understood to sit outside the Article 5 prohibition, which targets inference about the emotions of identifiable people. Keep the output at group level, prevent re-identification and document the design. The closer the analysis moves to a named individual's emotional state, the closer it moves to the ban.

Is emotion recognition AI banned outside the workplace in the EU?

Not banned, but regulated. Outside workplaces and education institutions, emotion recognition is treated as high-risk under Annex III, which brings conformity duties rather than prohibition, and those duties sit on the timeline the Digital Omnibus deferred to 2 December 2027. The prohibitions and the high-risk regime are different instruments with different dates.

What counts as a medical or safety reason under the exception?

The exception should be read narrowly. Fatigue or distress detection genuinely deployed for safety in high-hazard settings is the kind of use the carve-out contemplates; productivity, engagement or performance purposes are not. An organisation relying on the exception should document the medical or safety purpose precisely and be prepared to defend it as the real purpose, not a label.

My company is UK-only. Can we ignore the EU ban?

Ignoring it is unwise even where it does not apply. UK GDPR, Article 8 ECHR and the ICO's monitoring-at-work guidance all constrain emotion-level monitoring of staff, and any EU workforce brings the prohibition straight back into scope. Treating the EU ban as the design baseline is the lower-risk position for a UK employer.

How do I find out whether our monitoring tools infer emotions?

Ask the vendor in writing for a complete list of analytics the suite runs on staff data, including features added by update, then map each analytic to a documented purpose. Anything that scores the emotional or engagement state of identifiable workers needs a medical or safety justification in the EU, and a strong necessity case in the UK, or it should be switched off.