EU AI Act Article 5(1)(f): Emotion Recognition in the Workplace and Education Is Prohibited
Workplace and education emotion recognition is prohibited under Article 5(1)(f) of the EU AI Act since 2 February 2025. Fine ceiling: €35M or 7%.
Inferring the emotions of employees or students is not a high-risk activity under the EU AI Act — it is a prohibited one. Article 5(1)(f) of Regulation (EU) 2024/1689 bans AI systems that infer the emotional or internal states of natural persons in the workplace and in educational institutions. The prohibition has been in force since 2 February 2025. If your organisation is running — or evaluating — software that analyses facial expressions, voice tone, or body language to assess employee engagement, stress, or mood, you are in Article 5 territory, not Annex III.
The fine ceiling for violating Article 5 is €35 million or 7% of total worldwide annual turnover, whichever is higher (Article 99(3)). For SMEs and start-ups, the cap applies at the lower of those two figures — a proportionality protection, but not a small number.
What Article 5(1)(f) Actually Prohibits
The text bans AI systems "that infer emotions of a natural person in the areas of workplace and education institutions." The key word is infer: the provision targets systems that draw conclusions about internal psychological states — stress, happiness, frustration, alertness, mood — from observable inputs such as facial geometry, vocal acoustics, keystroke dynamics, or movement patterns.
Scope in practice. The prohibition covers:
- Meeting analytics tools that score "engagement" or "sentiment" of individual employees from video feeds
- Call-centre AI that reads agent tone or stress level from voice recordings and feeds the result to supervisors
- HR interview platforms that analyse candidates' facial expressions to generate personality or emotion scores
- Productivity-monitoring software that infers whether employees are "focused" or "distracted" by analysing micro-expressions or gaze patterns
- Student-facing proctoring tools that flag "suspicious emotional states" during exams
The framing does not matter. A vendor calling its product a "wellness insight tool" or an "engagement analytics engine" does not change the classification. If the system derives an inference about a person's emotional or internal state in a workplace or educational setting, Article 5(1)(f) applies.
The safety and medical exception. One carve-out exists: systems intended for medical or safety reasons. The clearest examples are fatigue- and drowsiness-detection tools used in safety-critical environments — an AI that monitors whether a long-haul truck driver or machine operator is becoming dangerously tired, for instance. That use-case targets a physical safety state and operates for the protection of the individual and others. It does not extend to general productivity monitoring framed as safety, and it does not cover routine HR or performance-management applications.
What Article 5(1)(f) does NOT prohibit. Emotion recognition used outside the workplace and education, and not caught by another Article 5 provision, may still be lawful — though it falls into the high-risk tier under Annex III, point 1(c) and/or triggers the deployer disclosure duty in Article 50(3). The prohibition is context-specific: it is the employment and education settings that convert the practice from high-risk to unacceptable.
How This Differs from the Annex III Biometrics Category
There is a meaningful and frequently misunderstood distinction here.
Annex III, point 1(c) lists "AI systems intended to be used for emotion recognition" as high-risk — the full set of technical documentation, risk management, and conformity-assessment obligations from Articles 9–17 apply. But that high-risk classification operates in contexts outside the absolute prohibition. Once the deployment context is the workplace or an educational institution, Article 5(1)(f) overrides Annex III: the activity is no longer merely "high-risk and heavily regulated" but flatly prohibited.
Think of it as a hierarchy. Article 5 sits above the Annex III classification machinery. A system that would otherwise be classified as high-risk under Annex III is instead banned outright when Article 5 applies. No amount of technical documentation, conformity assessment, or human oversight makes a prohibited practice compliant — the only compliant outcome is not deploying it.
A separate but related prohibition. Article 5(1)(g) separately bans AI systems that infer sensitive personal attributes — race, ethnicity, political opinion, trade-union membership, sexual orientation — through biometric categorisation. That provision can also be triggered by emotion-adjacent tools that derive sensitive profile attributes from biometric data, even if the stated output is labelled "engagement" rather than "emotion." Both provisions may apply simultaneously.
The Realistic Trap: HR and Recruitment Video Tools
The most common compliance failure Confir sees in this space is not a malicious employer deliberately surveilling workers. It is a procurement decision made without regulatory scrutiny.
Consider a 60-person logistics company that subscribes to a video-interview platform. The sales deck emphasises "structured scoring rubrics" and "standardised assessments." Buried in the feature list is a capability that assigns a "candidate energy score" or "communication style profile" based on facial expression analysis. The HR manager who signed the contract never connected that feature to Article 5.
Under Article 5(1)(f), the employer who deploys that feature in hiring interviews is the party primarily on the hook. The vendor is a provider and has its own obligations, but the prohibition on use falls on the deployer operating the system in an employment context. The fine is calculated on the deployer's worldwide turnover, not the SaaS vendor's subscription fee.
The same pattern applies to engagement tools that analyse Slack tone, monitor email sentiment, or score meeting participation by inferring emotional states from video. Each of these — if deployed in the workplace and producing an inference about an employee's internal state — falls within Article 5(1)(f).
What is not prohibited. Tools that aggregate behavioural signals without making any individual-level emotional-state inference sit outside the prohibition. A system that measures response latency in ticketing tools or counts words per sentence in performance reviews is not inferring emotions; it is counting. The line is the inference of an internal psychological state from observable data, attributed to an identifiable individual.
Article 50(3): The Separate Transparency Duty
Even where emotion recognition is not prohibited — i.e., in contexts outside workplace and education — a deployer using an emotion-recognition system must inform persons that they are being subjected to it, under Article 50(3). This limited-risk transparency obligation applies from 2 August 2026 as part of the general application of the Act.
In a workplace setting, this duty is academic: you cannot satisfy a transparency obligation for a practice that is itself prohibited. But the provision is relevant for deployments in retail (customer emotion analysis), consumer products, or public-service contexts that fall outside the employment relationship.
What Happens to Existing Deployments
Article 5 has applied since 2 February 2025. There is no transition period for prohibited practices. If a workplace emotion-recognition tool was running on 2 February 2025, it became non-compliant that day.
Organisations that deployed such tools before that date should:
- Determine whether the system falls within the Article 5(1)(f) scope — specifically, whether it infers emotional or internal states of employees or students, outside the medical/safety exception.
- If it does: cease the prohibited function immediately. Retention of historical outputs processed after 2 February 2025 may itself present an ongoing compliance issue under GDPR Article 9 (special-category data, to the extent emotional inference data constitutes it).
- Document the cessation decision, including a record of what the system did and the legal basis for ceasing the practice. This supports your Article 12 logging posture and evidences good-faith compliance in any supervisory inquiry.
- Evaluate whether a modified use of the tool — stripped of the emotion-inference capability — can continue. Many video-interview platforms, for instance, can operate without their affect-scoring features. The structural scoring rubrics (standardised questions, timed sections, human rater panels) are not caught by Article 5.
For vendors: if your product includes a workplace emotion-inference feature, that feature must be removed from EU-market deployments. Placing a prohibited AI system on the market or putting it into service under your own name makes you a provider subject to Article 5 liability. You cannot contractually transfer this risk to your deployer clients.
How Confir Helps
Confir's Article 5 prohibited-practice checklist is the first gate every AI system passes through when registered. The intake questions ask, in plain language, whether the system infers emotional states in an employment or education context. If the answer triggers the Article 5(1)(f) rule, Confir returns an Unacceptable Risk finding — no further compliance workflow opens, because there is no compliant path.
The logic is deterministic and rule-based: same intake, same finding, no ambiguity. The rule that fired — Article 5(1)(f) — is displayed alongside the output, so you can read the statutory text and confirm the reasoning yourself. That explainability matters when you are documenting a decision to cease a deployment or explaining the outcome to a vendor.
For systems that fall just outside the prohibition — emotion recognition in contexts other than workplace or education — Confir continues into the Annex III high-risk classification workflow, with the full technical documentation, AIRC, AITR, AITO, and AIGM controls mapped to the applicable articles.
Frequently Asked Questions
Is all emotion recognition software prohibited under the EU AI Act?
No — the prohibition in Article 5(1)(f) is context-specific. It bans emotion inference in the workplace and in educational institutions (with a medical/safety exception). Emotion recognition deployed outside those contexts — in retail, consumer products, or public-space applications — is not caught by Article 5(1)(f), though it may be high-risk under Annex III, point 1(c), and is subject to the Article 50(3) transparency requirement from 2 August 2026.
When did the Article 5(1)(f) prohibition come into force?
2 February 2025. Article 5 (the full set of prohibited practices) and Article 4 (AI literacy) were the first obligations to apply under the Act's phased timeline. There is no grace period for prohibited practices — any deployment of a workplace emotion-recognition system after that date is non-compliant.
What is the fine for deploying a prohibited emotion-recognition system?
Article 99(3) sets the ceiling at €35 million or 7% of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs and start-ups, the fine is capped at the lower of those two figures (Article 99(6)) — but 7% of turnover is still material for a fast-growing company.
Does the prohibition cover AI that detects fatigue or drowsiness for safety?
No. Article 5(1)(f) explicitly exempts systems intended for medical or safety reasons. A fatigue-detection system used in transport, manufacturing, or other safety-critical industries — where the goal is preventing accidents — falls within the exception. The exemption does not extend to routine productivity or performance monitoring reframed as safety.
Our video-interview platform vendor says emotion scoring is just one optional module. Are we still at risk?
Yes, if you have enabled or could enable that module. The prohibition falls on the deployer operating the system in an employment context. "We didn't enable that feature" is a defensible position only if the feature is genuinely inaccessible, not merely switched off by default in your settings. Review your contract for data-flow provisions — if the vendor's model processes facial data and generates an inference, even as a backend signal, you need to confirm whether that inference is surfaced anywhere in your workflow.
Can a candidate or employee consent to emotion recognition in a hiring or employment context?
Consent in an employment context is widely considered not freely given under GDPR, given the power asymmetry between employer and employee. But the question is moot for Article 5(1)(f): the provision does not include a consent exception. Unlike GDPR special-category data (which can be processed with explicit consent in certain circumstances), the AI Act's prohibited practices are absolute bans. No consent mechanism unlocks a prohibited AI practice.
The original article on this page described workplace emotion recognition as Annex III high-risk. Is that wrong?
It was. Emotion recognition in the workplace is prohibited under Article 5(1)(f), not merely high-risk under Annex III. The slug and URL are unchanged; the legal analysis has been corrected to reflect the true classification. High-risk Annex III classification applies to emotion recognition in other contexts (outside workplace and education) where the prohibition does not operate.
Related guides
- Article 6 high-risk classification
- fundamental rights impact assessment
- risk classification framework
- Annex III high-risk use cases
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 7-day free trial.
Start free trial →