EU AI Act Emotion Recognition Ban: Workplace and Education

Article 5(1)(f) of the EU AI Act stops a specific chain of activity: an AI system takes biometric data, identifies or infers a natural person's emotion or intention, and operates in a workplace or education institution. The rule covers placing that system on the market or putting it into service for that purpose, including using it. The exception for medical or safety reasons is narrow.

I would block this use case at intake rather than ask HR or a school administrator to remember a legal exception during deployment. Put the system purpose and input modality on one review sheet. Add the affected people and deployment setting, plus the exception evidence. An empty exception field should stop the route.

TL;DR

• Article 5(1)(f) prohibits biometric emotion inference in workplaces and education institutions.
• Recruitment and probation can fall inside the ban. Employee monitoring, admissions and student-attention scoring can too.
• The medical and safety exception is narrow. General wellbeing and stress monitoring are prohibited, as are burnout, boredom or satisfaction monitoring.
• DeepInspect can deny authenticated calls to identified emotion-inference model routes. Cameras and sensor collection require separate controls, as does model validity.

Article 5(1)(f) sets a four-part legal test

The official text of Regulation (EU) 2024/1689 prohibits placing on the market or putting into service AI systems for the specific purpose of inferring emotions of a natural person in workplaces and education institutions, including using those systems. The exception applies when the intended use has medical or safety reasons. Article 5 has applied since 2 February 2025.

The Commission's guidelines on prohibited AI practices break the test into cumulative conditions. There must be an AI system placed on the market or put into service, including one already in use. The system must identify or infer emotion or intention from biometric data. Its setting must be a workplace or education institution. A qualifying medical or safety purpose takes the use outside the prohibition.

The Commission published those guidelines as non-binding interpretation. The Court of Justice of the European Union retains authority over the law's meaning. Teams should keep the legal text beside the classification record and obtain jurisdiction-specific advice.

Emotion inference starts with biometric data

Article 3(39) defines an emotion recognition system by purpose: identifying or inferring emotions or intentions of natural persons on the basis of biometric data. The Commission reads Article 5(1)(f) consistently with that definition. Face images and voice characteristics can qualify. So can gait and keystroke patterns. Eye tracking, body movement, EEG or heart signals can also qualify when the system processes those physical or physiological characteristics, or behavioural characteristics, as biometric data.

The output label matters less than the inference. Happiness, sadness, anger and embarrassment are listed examples, along with excitement, satisfaction and amusement, rather than a closed vocabulary. Calling the output an "attitude" or "engagement signal", including a "customer-facing readiness score", leaves the underlying inference unchanged.

A camera that counts visible smiles is different if it stops at the observable expression. The legal boundary changes when the system concludes that the person is happy or angry, anxious or bored. Text sentiment analysis also is outside Article 5(1)(f) when it infers tone from written content without biometric data. Other EU rules may still govern those systems.

Workplace coverage includes recruitment and remote work

The Commission interprets workplace broadly. It covers physical and virtual places where people perform assigned tasks, including offices, factories and warehouses, as well as vehicles, shops, mobile sites and video calls. Employment status is not an escape. Employees and contractors can sit within the protected relationship, as can trainees, volunteers and self-employed people.

Recruitment is part of the boundary. A provider that sells interview software to infer enthusiasm from a candidate's face or voice can engage the placing-on-the-market limb. An employer that activates the feature engages the use limb. The same analysis reaches probation reviews and call-centre anger monitoring. It also reaches supermarket cameras that score employee happiness and software that infers emotional dynamics during hybrid meetings.

I would reject "wellbeing" as a sufficient purpose statement. The Commission says general monitoring of workplace stress and systems intended to detect burnout or depression are prohibited under this exception analysis. A dashboard tile marked "team mood" may look harmless. Its data path can still contain a biometric inference about each named worker.

Education coverage reaches admissions and online classrooms

Public and private education institutions fall within the Commission's interpretation, across age groups and delivery modes. Schools and universities can qualify, as can vocational institutions and continuous-training bodies. The covered setting includes in-person classes and online instruction, plus blended learning and the admissions process.

A school that requires webcam software to infer student attention or anxiety uses an emotion recognition system in education. Admissions testing with the same function is also prohibited. Eye tracking used only to locate a student's gaze during an online exam has a different purpose. Adding an arousal or anxiety inference crosses into emotion recognition.

Institutional control changes the classification of a language-learning application. Private use outside an education institution is outside this particular workplace-and-education ban. A required biometric emotion feature is a covered deployment when the institution assigns it to students. Role-play training may also be outside when the result serves the learner and cannot affect evaluation or certification. That limit belongs in the system configuration and evidence file, rather than in a sales assurance.

The medical and safety exception is narrow

The Commission says the medical and safety exception should receive a narrow interpretation. Therapeutic use points to a CE-marked medical device. Safety concerns must protect life or health. Protecting property against theft or fraud is outside that safety concept. General wellbeing and motivation fail the stated exception. Satisfaction, burnout and routine stress monitoring also fail it.

Necessity and proportionality need evidence. The guidelines call for limits on time and scale, with the affected people defined and safeguards in place. They suggest a prior written and reasoned expert opinion as one possible safeguard. The assessment should test less intrusive alternatives against the medical or safety purpose. It should avoid treating an employer's preference as objective necessity.

Data collected for a valid safety case cannot migrate into performance scoring. A dangerous-machinery deployment may justify a narrowly configured warning when a specific state creates danger. HR cannot reuse that output for promotion or discipline. Separate route permissions and retention rules around the exception before activation. Define access roles and downstream exports separately.

Physical states and emotion inferences require separate records

Recital 18 excludes physical states such as pain or fatigue from the emotion recognition definition. A system that detects driver or pilot fatigue to prevent accidents therefore may be outside emotion recognition, rather than relying on the Article 5(1)(f) exception. The classification record should state which path applies because the controls and legal reasoning differ.

Detecting a raised voice or smile is outside when the system reports only that apparent signal. The same applies to a frown or head movement. Using the same signal to infer anger or happiness, anxiety or intention brings emotion recognition back into view. Procurement language such as "behaviour analytics" rarely answers this question.

Emotion recognition outside workplaces and education institutions is outside this particular prohibition. It can still be high-risk under Article 6(2) and Annex III point 1(c). Article 50(3) can also require notice to exposed people. The Annex III classification guide covers the high-risk path. The Article 50 transparency guide owns the separate disclosure decision.

Build a prohibition record before model access

Use one record for each proposed system purpose and deployment setting. Start with the provider and model or service version. Then state the intended function. Name the natural persons affected and the institution. Record the business owner separately. Record every input modality and state which inputs count as biometric data. Preserve sample outputs, including labels that claim to measure attention or engagement, stress or attitude, or intent.

Then make the Article 5(1)(f) decision explicit:

• Activity: Record who places the system on the market or puts it into service for the stated purpose, including who uses it.
• Inference: Identify the emotion or intention inferred and the biometric input used to produce it.
• Setting: Name the workplace or education context, including recruitment and admissions, remote work or online instruction.
• Exception: State the medical or life-and-health safety reason. Attach necessity and proportionality evidence. Include expert and purpose-limit evidence, plus the alternative analysis.
• Decision: Record whether the outcome is denial or bounded permission. Route uncertain cases for legal review. Name the approver and policy version.
• Technical enforcement: Map the decision to model routes and application identities. Record permitted roles and blocked fallback paths.
• Review trigger: Reopen the record after a model or input change. Do the same after an output, purpose or population change, or after an institution or downstream-use change.

Keep runtime evidence for systems that proceed. The Article 12 logging guide explains the adjacent request record. A log supports traceability. It never converts a prohibited use into a permitted one.

DeepInspect

DeepInspect can enforce the route decision after the organization has classified the use. For authenticated HTTP calls deliberately routed through its stateless proxy, a policy can deny access to model endpoints or features designated for prohibited emotion inference. The decision record can bind the caller identity and application. It can also bind the destination route, policy version and permit or deny outcome.

That control covers the governed model call. Camera capture and microphone collection sit outside the proxy boundary. Biometric sensor processing and model validity do too, along with legal interpretation and medical-device status. Direct provider calls and local models also need separate controls. Keep those owners on the prohibition record so a blocked gateway route is never presented as coverage of the entire system.

Book a technical deep dive at deepinspect.ai.