Emotion AI and Measurement in 2026
Originally published Feb 2011. Updated June 14, 2026.
Emotion AI — also called affective computing — is the discipline of using machine learning to recognize, interpret, and respond to human emotional states from facial expressions, voice tone, and biometric signals. The category was pioneered by MIT spinout Affectiva in 2009, consolidated under Swedish eye-tracking company Smart Eye in 2021 ($73.5 million acquisition), and is now subject to direct regulation under the EU AI Act, which prohibits emotion recognition in workplace and educational settings from February 2, 2025.
What emotion AI is and where it now operates
Emotion AI systems take inputs (facial video, audio, physiological signals from wearables) and output probability scores across emotional or attentional states — joy, surprise, anger, sadness, contempt, fear, engagement, distraction, drowsiness. The technical foundation rests on facial-action-coding-system (FACS) work pioneered by Paul Ekman in the 1970s, adapted for machine vision in the 2000s, and combined with deep-learning architectures since the 2015 ImageNet inflection point.
The commercial applications cluster in five domains. Market research and ad-effectiveness measurement (the original Affectiva business). Automotive driver-monitoring systems (the application that drove Smart Eye's acquisition strategy). Customer-experience analytics in call centers. Healthcare diagnostics, particularly mental-health screening and autism-spectrum research. And — controversially — workplace productivity monitoring and educational attention tracking, the use cases now restricted under EU law.
For adjacent reading, see EPR's health tech and AdTech coverage.
Affectiva, Smart Eye, and the consolidation of the category
Affectiva, spun out of the MIT Media Lab in 2009 by Rana el Kaliouby and Rosalind Picard, became the most-cited emotion-AI vendor of the 2010s. The company built one of the largest annotated facial-expression datasets in the world — more than 12 million face videos from 90+ countries by 2020 — and licensed its Affdex emotion-analytics platform to market-research firms including Kantar, Mars, Unilever, and CBS for ad testing.
Affectiva was acquired by Smart Eye, a Gothenburg-based eye-tracking specialist, in June 2021 for $73.5 million. The acquisition rationale was automotive: Smart Eye supplies driver-monitoring systems to OEMs including BMW, Polestar, and GM, and Affectiva's facial-expression and cognitive-state recognition complemented eye-tracking for driver-attention systems. The combined company now competes for the automotive interior-sensing tier with Seeing Machines (ASX-listed), Tobii (Swedish, listed on Nasdaq Stockholm), and the OEM-internal solutions emerging from Tesla, Mercedes-Benz, and Mobileye (NASDAQ: MBLY).
Realeyes, Hume AI, and the current vendor field
The current emotion-AI vendor landscape includes several distinct players. Realeyes, London-based and founded in 2007, focuses on attention and emotion analytics for advertising effectiveness measurement. The company's clients include Mars, Hershey, Coca-Cola, and AT&T. Realeyes raised $12.4 million in Series B funding in March 2024.
Hume AI, founded in 2021 by former Google DeepMind researcher Alan Cowen and headquartered in New York, raised $50 million in Series B in March 2024, reportedly at a valuation above $200 million. Hume's Empathic Voice Interface (EVI), launched in March 2024, integrates voice-based emotion measurement with large language model responses — a category sometimes labeled emotionally intelligent AI rather than emotion measurement, with applications in customer support, mental-health screening, and conversational interfaces.
Other notable vendors include Audeering (Munich-based, voice-emotion specialist), MorphCast (in-browser emotion analytics), and Cogito (real-time call-center emotion coaching, raised $25 million in 2022). The category remains less consolidated than market-research peers; many vendors operate at $5 million to $50 million in annual revenue.
The Microsoft Azure Face API exit and the technology-ethics inflection
The clearest signal that the broader market had reservations about emotion AI came from Microsoft. In June 2022, Microsoft announced that Azure Face API would retire its emotion-detection capabilities and limit access to facial recognition more broadly. The Responsible AI Standard release, authored under Natasha Crampton's office, cited concerns about scientific consensus on the validity of emotion inference from facial expression alone — a critique advanced influentially by psychologist Lisa Feldman Barrett and colleagues in their 2019 review paper for Psychological Science in the Public Interest.
The Microsoft decision shaped the regulatory framing that followed. The EU AI Act's emotion-recognition provisions, finalized in 2024, draw heavily on the same scientific-validity critique. The position now broadly held in the academic literature is that facial expressions carry useful information about affective states but cannot reliably support the kind of fine-grained emotion classification (six or seven discrete emotions inferred from a single frame) that the first generation of commercial emotion AI claimed.
The EU AI Act emotion-recognition restrictions
The EU AI Act, in force from August 1, 2024, places emotion recognition under tiered restrictions. The Article 5 prohibitions, effective February 2, 2025, ban emotion recognition systems "in the areas of workplace and education institutions" — with limited safety and medical exceptions. Other emotion-recognition applications are classified as high-risk under Annex III, requiring conformity assessment, data governance, transparency obligations, and human-oversight provisions before commercial deployment.
The penalty ceiling for non-compliant deployment of prohibited systems reaches €35 million or 7% of global annual turnover, whichever is higher. The practical implication for emotion AI vendors is significant: workplace productivity monitoring (Vibe, Teramind, and similar) and education-platform attention tracking are now legally constrained within the EU's 27 member states.
Healthcare and mental-health applications — the durable use case
Healthcare remains the most credible emotion-AI deployment domain. Cognoa, the Palo Alto-based pediatric behavioral-health company, received FDA De Novo authorization in June 2021 for an AI-based autism diagnostic aid for children aged 18 to 72 months — the first such authorization. The system uses caregiver questionnaires and short video assessments analyzed by machine-learning models trained on more than 2,800 clinical labeled videos.
Voice-based emotion analysis has clinical applications in depression screening (Ellipsis Health, Kintsugi, both venture-funded) and Parkinson's disease early detection (research-stage, with multiple academic groups publishing). Wearable-based affective state inference is integrated into devices from Apple Watch, Oura Ring, and Whoop, though the marketing claims here remain looser than the medical-device-grade vendors.
For broader frames, see EPR's wellness coverage.
Where emotion measurement actually creates value in 2026
The honest answer in 2026 is narrower than the 2011 framing suggested. Emotion AI creates measurable value in three settings. First, where the use case is aggregated rather than individual — ad-testing panels of 500 viewers produce statistically valid attention and engagement signals even if no individual frame is reliably classified. Second, where the system supplements rather than replaces a human judgment — clinician-supported autism diagnostic aids, call-center coaching that surfaces patterns for human supervisor review. Third, where the measurement is functional rather than emotional — driver drowsiness detection from blink rate and head pose is a well-validated safety system that does not depend on inferring discrete emotions.
The use cases where emotion AI does not credibly create value in 2026 are also clearer. Hiring decisions based on candidate-video emotion scoring (HireVue retired its facial-analysis component in January 2021 under sustained criticism). Workplace productivity monitoring based on individual employee affect (now restricted in the EU). Educational attention scoring for individual students (now restricted in the EU and subject to growing US state-level legislation).
