The emotion recognition prohibition is narrower than its reputation in three separate ways, and the third — that inferring a physical state is not inferring an emotion — is the one that decides most engineering questions.
Two settings, and nowhere else
Article 5(1)(f) of Regulation (EU) 2024/1689 prohibits the placing on the market, putting into service for this specific purpose, or use of AI systems to infer emotions of a natural person in the areas of workplace and education institutions, except where the use is intended to be put in place or into the market for medical or safety reasons.
Two settings, exhaustively. Emotion inference in a retail environment, in a call centre serving customers rather than monitoring staff, in a consumer product, or in a public space is not prohibited by this provision. Recital 44 explains the choice: the concern is the imbalance of power in employment and education, where a person cannot reasonably decline.
A call centre is the instructive case because it is often both. A model inferring the caller’s emotion is not operating on a person in a workplace in the relevant sense — the caller is a customer at home. A model inferring the agent’s emotion, or scoring the agent on inferred warmth, is squarely within the prohibition. The same audio stream, the same technique, two different answers, and the difference is whose emotional state is being inferred.
Not legal advice. Workplace emotion inference in the EU engages the AI Act, the GDPR (biometric data under Article 9), national employment law, and in several member states a works council consultation requirement that can bind before any of the above. A system outside Article 5(1)(f) can be unlawful under any of them. Take advice on the specific deployment.
What counts as inferring an emotion
Article 3(39) defines an emotion recognition system as an AI system for the purpose of identifying or inferring emotions or intentions of natural persons on the basis of their biometric data. Two elements matter.
- Emotions or intentions. Intention inference is included, which is broader than an emotion label and catches predictive systems that never name a feeling.
- On the basis of biometric data. Article 3(34) defines biometric data as personal data resulting from specific technical processing relating to the physical, physiological or behavioural characteristics of a natural person. Face, voice, posture, gaze are all in.
Text is the interesting boundary. Sentiment analysis of what an employee wrote in a message is not obviously operating on biometric data — writing is a behavioural product rather than a physiological or behavioural characteristic in the Article 3(34) sense. On the plain words, text sentiment analysis sits outside the definition and therefore outside the prohibition. That reading is not settled, and keystroke dynamics or typing rhythm would be a different matter entirely, since those are behavioural characteristics of the person rather than the content they produced.
The physical-state boundary
Recital 44 draws a line that a great many summaries omit and that resolves most practical questions. The notion of emotion recognition does not include physical states such as pain or fatigue, and the recital gives the example of systems used to detect the state of fatigue of professional pilots or drivers for the purpose of preventing accidents. Nor does it include the mere detection of readily apparent expressions, gestures or movements, unless those are used to identify or infer emotions.
So a driver monitoring system that detects eyelid closure and head nodding to raise a drowsiness alert is not an emotion recognition system at all — it never reaches the exception, because it never engages the definition. A system that detects a smile and logs “smiling” is detecting a readily apparent expression. The same system that maps the smile to “happy” and scores the employee on it has crossed the line.
The distinction is between observing what a body is doing and asserting what a mind is feeling. It is a real distinction and it is also easy to cross by accident: a fatigue detector whose output is labelled “engagement” and used in a performance review has become an emotion inference in substance, whatever the model architecture.
The medical and safety exception
The exception is for use “intended to be put in place or into the market for medical or safety reasons”. Three features are worth noting.
- It turns on intended purpose, not on effect. The wording is about what the use is intended for. Intended purpose is a defined concept in the Act — Article 3(12) ties it to the use for which the system is intended by the provider, including in the instructions for use and the promotional materials — so the documentation is evidence of the claim.
- Recital 44 gives therapeutic use as the medical example. Genuine clinical application, not wellbeing branding attached to a monitoring product.
- Safety is not productivity. A system that detects distress in a lone worker to trigger a welfare response is a safety case. A system that infers frustration to route work differently is an operations case wearing safety language, and the intended purpose documentation would not support it.
The exception is the part of the provision most likely to be tested, because it is the part with commercial pressure on it. There is no case law and the Commission’s February 2025 guidelines do not supply a test for distinguishing a genuine safety purpose from a pretextual one. A deployment resting on it should expect to have to evidence the purpose rather than assert it.
What happens outside the ban
Falling outside Article 5(1)(f) moves a system down a tier, not out of the Act.
Annex III point 1(c) classifies emotion recognition systems as high-risk where they are not prohibited, so a permitted deployment carries the whole Chapter III apparatus — risk management under Article 9, data governance under Article 10, technical documentation under Article 11, logging, human oversight, conformity assessment and registration. And Article 50(3) imposes a transparency duty on deployers of emotion recognition systems: inform the natural persons exposed to the system of its operation, and process any personal data in accordance with the applicable data protection law. See the Article 50 notice duty for emotion and biometric systems.
There is also a scientific validity objection recorded in recital 44 itself, which notes serious concerns about the scientific basis of emotion recognition systems, particularly their limited reliability, lack of specificity and limited generalisability. That is unusual language for a recital and it is worth reading as a signal about how a decision-maker is likely to receive accuracy claims made under Article 15. The adjacent prohibition on inferring sensitive traits from biometrics is at Article 5(1)(g), and employment-context high-risk classification at Annex III point 4.
Top comments (0)