Emotion Recognition at Work Is Now Illegal in the EU. The Rest of the World Is Still Deciding

---

Four of the world’s biggest technology companies have been steadily filing patents for reading an employee’s mood from their face or their voice. The Financial Times reported on 12 August that Microsoft, Amazon, Nvidia and Meta all hold filings that, in different ways, try to work out how a worker feels without asking them.

Here’s the part that matters for HR: the EU has already made the decision for you. Emotion recognition at work is banned there, by law, full stop. No exemption for good intentions.

The rule sits in the EU AI Act, and it’s blunt. Employers cannot use AI to guess how a worker feels from their face, their voice or their body. That’s been the law since February 2025, and the version of the Act currently in force, dated 27 July 2026, hasn’t softened it.

There’s one exception: genuine medical or safety use, the kind that needs a certified medical device behind it. A stress dashboard for line managers doesn’t qualify just because stress sounds like a health issue. Most of what vendors currently pitch as “wellbeing tech” would fail that test, and picture the pitch: a demo dashboard scoring each employee’s voice on a call, green for calm and red for stressed, sold as care rather than surveillance. That is exactly the kind of emotion recognition at work the ban was written to catch.

What counts as emotion recognition at work, and what doesn’t

The EU’s rule on emotion recognition at work is specific about what it catches. It covers AI that infers emotion from facial expressions, tone of voice, or physical signals like heart rate. It does not cover a survey asking how someone feels, and it does not cover a system that simply flags tiredness rather than a feeling.

EU regulators admit there’s a grey area between spotting a mood and inferring an emotion, and between a tool that’s “about work” and one that isn’t. Lawyers will be arguing over individual cases for a while yet.

Here’s a simpler test than the statute itself: was the data given, or was it taken? An engagement survey, a pulse check, a one-to-one where someone tells their manager they’re struggling: that’s data an employee chose to hand over. A system that scores a call centre agent’s voice for stress in real time, without the agent doing anything to trigger it, is data taken from them.

That’s also why the EU skipped consent as a fix. Most privacy law, GDPR included, manages this kind of risk by asking people to agree to it. The AI Act doesn’t bother, because an employee can’t meaningfully say no to the person who pays them and can end their contract. A consent box signed under that pressure isn’t really a choice, so the law removed the choice from the equation entirely.

It’s worth being fair to the technology too. There are real, well-meaning uses on the table: spotting burnout before someone resigns, catching fatigue in safety-critical jobs like transport, flagging a contact-centre agent before a bad call tips them over. Vendors lead with exactly these examples when they pitch it to HR and safety teams.

The EU accepts the intentions are genuine. The objection sits with the technology instead.

Reading emotion from a face or a voice is unreliable, and it performs worse across different cultures, ages and people. Put shaky science next to consent nobody can truly refuse, and prohibition starts to look like the safer call than light-touch regulation.

Outside the EU, there’s no equivalent rule yet. Vendors already sell sentiment-analysis and biometric wellbeing tools to American employers that would fall foul of the EU’s law if used on staff there. Until local regulators catch up, US and global HR leaders are on their own to draw that line, which means the decision about emotion recognition at work sits with whoever signs off the procurement, not with a regulator.

It’s a familiar pattern. GDPR ended up setting the practical standard for multinationals long before other countries passed anything similar, and this ban could easily follow the same path. That’s a reasonable bet, not a certainty this early.

Staff already notice when they’re watched without being asked, and they don’t tend to shrug it off. Covert AI use by staff tends to rise wherever people suspect they’re monitored without consent, and trust in how AI shapes decisions about someone’s job is fragile even before an emotion-reading tool enters the picture. Loyalty research keeps landing on the same point: staff commitment stays conditional on being treated as a person in the decision, not a data source for it.

Given or taken is worth applying to any monitoring tool a vendor pitches, EU rule or not, before it ever reaches procurement. Ask it whether the vendor’s own slide deck calls it emotion recognition at work, employee sentiment, or something softer. It’s one more item on a list that was already stacking up, alongside pay transparency deadlines and the employment law changes piling on over the past year and a half.

Does emotion recognition at work belong on a compliance checklist or a procurement one?

Both, and increasingly at the same meeting. The next time a wellbeing platform, a call-quality tool or a wearable pitch lands on a desk promising to read how staff feel, the sharper question is simple: was that data given freely, or taken without asking.