Emotion Recognition System (EU AI Act Definition)

An emotion recognition system is an AI system used to identify or infer the emotions or intentions of natural persons from their biometric data. Under Regulation (EU) 2024/1689 — the EU AI Act — this category carries one of the Act's sharpest compliance lines: use in workplace and education settings is outright prohibited, while use outside those contexts remains possible but triggers high-risk or transparency obligations depending on deployment.

The EU AI Act definition

Article 3 of Regulation (EU) 2024/1689 provides the statutory definitions. Point 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."

Three elements of that definition matter in practice.

Biometric data as the input. The system must draw on biometric data — physiological, behavioural, or physiognomical characteristics. Facial expressions captured on video, vocal inflection patterns, heart-rate variability from a wearable, micro-gestures tracked by a camera: all qualify. A purely text-sentiment classifier that never processes a person's physical signals sits outside this definition (though it may face other transparency duties under Article 50).

Inference, not just detection. The definition captures both identification (labelling an observed state, such as "this person is frowning") and inference ("therefore this person is unhappy or uncooperative"). Systems that go further and infer intentions — assessing whether a job candidate seems motivated, whether a student appears engaged — are squarely covered.

The physical-state boundary. The Act draws a practical distinction between detecting readily apparent physical states and inferring emotional or intentional states. Detecting fatigue in a long-haul driver from eye-tracking data, or recognising acute pain in a patient through facial expression analysis, sits closer to physiological measurement than emotional inference. The prohibition in Article 5(1)(f) targets emotion and intention inference, not every use of biometric data related to a person's state. This line is not always bright, but it is the boundary regulators will apply.

The Article 5(1)(f) workplace and education prohibition

Article 5(1)(f) bans the placing on the market, putting into service, or use of emotion recognition systems in two specific settings: the workplace and educational institutions.

This prohibition has applied since 2 February 2025. It is not a future obligation — it is live law. Any organisation currently running such a system in either context is already in breach.

The medical and safety exception is narrow. Article 5(1)(f) carves out use "for medical or safety reasons." The canonical examples are clinical: a hospital system monitoring post-operative patients for signs of distress, or a driver-monitoring system in a heavy-goods vehicle that flags acute fatigue before a microsleep event. The exception does not extend to general employee wellness monitoring, productivity tracking, or any use that mixes safety detection with behavioural or performance inference. If a system does both — flags fatigue and scores employee engagement — the prohibited function taints the whole deployment unless the functions are genuinely separable systems.

What "workplace" covers. The prohibition is not limited to traditional employment relationships. Any professional setting in which natural persons are monitored during the performance of work activities falls within scope. This includes remote workers monitored via webcam, gig-economy workers tracked through app interactions, and contractors on a client's premises.

What "educational institutions" covers. Schools, universities, vocational training centres, and online learning platforms that monitor students for engagement, attention, or emotional response during instruction are all within scope. Exam-proctoring software that infers dishonest intent from facial expressions is a straightforward example of a prohibited use.

The penalty ceiling for an Article 5 breach is €35,000,000 or 7% of total worldwide annual turnover, whichever is higher (Article 99(3)). For growing technology companies, the turnover percentage will often be the binding figure. There is a proportionality protection: Article 99(6) caps fines on SMEs and start-ups at the lower of the percentage or the fixed amount.

When emotion recognition is high-risk

Outside the prohibited workplace and education contexts, emotion recognition systems fall under Annex III, area 1 (Biometrics) of the Act. Specifically, Annex III lists emotion recognition as a high-risk AI use case — alongside remote biometric identification and biometric categorisation — where the system is deployed in circumstances not covered by the Article 5(1)(f) prohibition.

Practical examples of potentially high-risk, non-prohibited deployments include:

• A law-enforcement tool assessing behavioural cues during witness interviews (Annex III area 6, law enforcement, also in scope).
• A border-control system inferring the emotional state of travellers during document screening (Annex III area 7, migration and border control).
• A clinical-research platform inferring patient emotional responses to treatment — distinct from the Article 5 medical exception in that the purpose is research, not direct clinical care.

For high-risk emotion recognition systems, the full Article 6 and Annex III compliance stack applies: a risk management system under Article 9, technical documentation under Article 11 meeting the requirements of Annex IV, data governance under Article 10, transparency to deployers under Article 13, human oversight under Article 14, accuracy and robustness requirements under Article 15, and conformity assessment under Article 43.

Because emotion recognition sits in Annex III area 1 (biometrics), the applicable conformity-assessment route is the Annex VII notified-body route (where the provider has not applied harmonised standards), rather than the internal self-assessment route under Annex VI that applies to most other Annex III categories.

Under the Digital Omnibus agreed in May 2026, obligations for stand-alone high-risk AI systems under Annex III apply from 2 December 2027, pushing back the original August 2026 date. For emotion recognition systems that are safety components of products covered by EU product law listed in Annex I, the date is 2 August 2028.

Providers claiming the Article 6(3) filter — that their Annex III system does not pose a significant risk to health, safety, or fundamental rights — must document that assessment. The filter is available only where the system performs a narrow procedural task, improves a previously completed human activity, detects decision patterns without replacing or influencing human assessment, or carries out preparatory work. Any system that profiles natural persons remains high-risk regardless of the filter.

The Article 50 transparency duty

Article 50(3) imposes a disclosure obligation on providers and deployers of emotion recognition systems operating outside the Article 5 prohibition (where use is still lawful). Natural persons exposed to an emotion recognition system must be informed that they are subject to such a system.

This duty applies regardless of whether the system is classified as high-risk. A limited-risk emotion recognition system — say, a customer-service tool that adjusts communication style based on inferred emotional tone — still triggers the Article 50(3) notification requirement.

The disclosure obligation applies from 2 August 2026. That date was not deferred by the Digital Omnibus; it applies as originally scheduled.

Practically, Article 50(3) means organisations must:

1. Identify every context in which individuals encounter an emotion recognition system.
2. Design and deliver a clear notification before or at the point of exposure — not buried in a privacy policy.
3. Document the disclosure mechanism as part of their broader transparency record.

A breach of Article 50 falls in the €15,000,000 or 3% tier under Article 99(4) — not the Article 5 prohibition ceiling, but still a material risk.

Frequently Asked Questions

Is detecting stress or fatigue the same as emotion recognition under the Act?

Not automatically. The EU AI Act targets systems that infer emotions or intentions from biometric data. Detecting physiological indicators of fatigue — yawning frequency, eyelid drooping, reduced reaction time — in a driver-assistance or occupational safety context sits closer to physical-state measurement than emotional inference. Providers of such systems should document why their system detects states rather than infers emotions or intentions, because the distinction is not always obvious and regulators will scrutinise the purpose and outputs of the system, not just its technical architecture.

Can an employer use emotion recognition for safety monitoring under the Article 5(1)(f) exception?

Yes, but the exception is narrow. The use must be genuinely for medical or safety reasons — detecting acute fatigue or distress to prevent a workplace accident, for example. It does not cover employee engagement scoring, productivity analysis, or any inference about an employee's attitude, honesty, or performance. If a system serves both a safety function and a broader behavioural-assessment function, the prohibited purpose contaminates the whole deployment unless the functions are genuinely implemented as separate systems.

Does Article 5(1)(f) cover online education and remote work?

Yes. The prohibition is not tied to physical premises. An e-learning platform that uses webcam analysis to infer whether students are paying attention, or a remote-work monitoring tool that analyses facial expressions during video calls, falls within scope. "Workplace" and "educational institution" describe the purpose and relationship of the monitoring, not the physical location.

What conformity-assessment route applies to a lawful, high-risk emotion recognition system?

Emotion recognition sits in Annex III area 1 (biometrics). Under Article 43, biometric AI systems generally require the Annex VII notified-body conformity-assessment route, unless the provider has applied harmonised standards that cover all relevant requirements, in which case Annex VI internal self-assessment is available. Most providers should assume a notified body will be involved unless harmonised standards are confirmed to apply.

When does the Article 50(3) disclosure duty apply?

Article 50(3) applies from 2 August 2026 — the general application date of the Act, which was not deferred. Any operator of a lawful emotion recognition system (outside the Article 5 prohibition) must inform affected persons from that date. The high-risk compliance stack (Articles 9–15, conformity assessment) applies from 2 December 2027 for stand-alone systems, but the transparency disclosure duty is earlier and independent.

Does the prohibition cover systems that infer intentions, not just emotions?

Yes. Article 3, point 39 expressly includes systems used for "identifying or inferring emotions or intentions" of natural persons from biometric data. A recruitment system that analyses video interview footage to score a candidate's apparent enthusiasm or sincerity is inferring intentions, not just observing physical states. In a workplace or education context, such a system is prohibited from 2 February 2025.

Related terms

• Emotion recognition in the workplace — Article 5(1)(f) prohibition
• Article 5 — Prohibited AI practices under the EU AI Act
• Biometric categorisation system — Annex III definition and rules
• Biometric data — definition and scope in the EU AI Act
• Article 50 — Transparency duties for limited-risk AI systems
• Annex III — The EU AI Act high-risk use-case list