Key takeaways
- Inferring emotions from voice, face or other biometric data is prohibited in workplaces and education institutions (Article 5(1)(f)), unless the use is for medical or safety reasons.
- Outside those settings, assess high-risk status under Article 6 and Annex III, including any applicable exception, and provide Article 50(3) notices.
- The prohibitions have applied since 2 February 2025, and fines for breaching them reach €35 million or 7% of worldwide annual turnover (Article 99(3)).
When is voice emotion recognition prohibited?
If a system infers emotions from voice or other biometric data in a workplace or education institution, Article 5(1)(f) generally prohibits it. The exception covers medical or safety reasons. For other settings, assess Article 6 and Annex III high-risk classification and Article 50 notices. A vendor’s “sentiment” label does not establish what its system actually does.
Banned, high-risk or allowed: where your use case falls
| Use | Status | Why |
|---|---|---|
| Inferring employees' emotions from their voice, for example anger in call-centre agents | Prohibited | Article 5(1)(f), workplace. The Commission's guidelines use this very example. |
| Inferring students' emotions in a school or training institution | Prohibited | Article 5(1)(f), education institution. |
| Inferring customers' emotions from their voice on a support line | Assess high-risk classification | Annex III, point 1(c), is subject to Article 6 classification rules. Assess the actual use and provide Article 50(3) notice where applicable. |
| Detecting a driver's or pilot's fatigue | Not emotion recognition | Recital 18 excludes physical states such as pain or fatigue. |
| Scoring the sentiment of the words in a transcript | Outside the definition on its face | No biometric data is processed. The Commission's guidelines are non-binding, so check borderline tools. |
| Inferring emotions at work for a medical or safety reason | Exception, read narrowly | Article 5(1)(f) exception. The Commission says it should be interpreted narrowly, so it does not cover general wellbeing monitoring. |
An emotion recognition system (Article 3(39)) identifies or infers emotions or intentions from biometric data. Voice counts when it is processed to measure characteristics of the speaker. Simply detecting an obvious expression, such as a raised voice, is not enough unless it is used to infer an emotion.
How the rules are enforced
Breaching Article 5 can cost up to €35 million or 7% of worldwide annual turnover, whichever is higher (Article 99(3)). Failing to tell people about an emotion recognition system, an Article 50 duty, falls in the next tier: up to €15 million or 3% (Article 99(4)). Each Member State designates the national authorities that enforce the Act, and the Commission's guidelines are a non-binding aid for reading it.
What to check this week
- List every tool that scores mood, stress, sentiment or intent from audio, video or typing, including features inside software you already license.
- For each one, note who is exposed: employees, students or customers.
- Switch off anything that infers the emotions of employees or students, unless a medical or safety reason is documented.
- For customer-facing use, assess Article 6 and Annex III classification, provide required notices and decide whether a DPIA is required.
- Ask each vendor in writing what the system infers and from which data.
Send us the voice features you are considering and who they analyse. We can map the technical behaviour so your legal team can assess the intended use.
Review your voice AI featuresFAQ
Is emotion recognition from voice banned in the EU?
Article 5(1)(f) generally prohibits biometric emotion inference in workplaces and education, with medical or safety exceptions. Other uses still need assessment under Article 6, Annex III, Article 50 and data-protection rules.
Can a call centre analyse its customers' emotions?
The customer’s role does not by itself establish that the use is lawful. Check whether the tool infers emotions from biometric data, also analyses employees, falls within Annex III and meets GDPR requirements. Provide the applicable notice and assess DPIA requirements before use.
Is voice always biometric data?
Not by itself. A recording is audio until it is processed to measure characteristics of the speaker, which the AI Act's definition of biometric data (Article 3(34)) covers. The GDPR treats biometric data as special category only when it is used to identify a person uniquely (Article 9).
Sources
- Regulation (EU) 2024/1689 (AI Act), Articles 3, 5, 50 and 99, Annex III and Recital 18, Official Journal of the EU
- European Commission, Guidelines on prohibited artificial intelligence practices (C(2025) 5052, 29 July 2025, non-binding)
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), application date for Annex III systems, Official Journal of the EU
- Regulation (EU) 2016/679 (GDPR), Article 9, Official Journal of the EU
About the author
Erik Salo Akimov · Co-founder · Engineering
Erik runs engineering at ESADEX. He designs the platforms, infrastructure and security controls that sit behind client systems.
