From Science Fiction to Reality: The Beginning of Emotionally Intelligent AI
The idea of machines that can recognize, interpret, and respond to human emotions has long been a staple of science fiction — from HAL 9000's unsettling calm to Samantha's evolving relationship with Theodore in Her. Over the past several years, a combination of advances in multimodal AI, facial recognition, voice analysis, and natural language processing has brought emotion-sensing AI considerably closer to commercial reality.
The consequences — legal, ethical, and psychological — are only beginning to be worked out.
What Affective Computing Can Do Now
Modern affective computing systems combine signals from multiple modalities: facial expression analysis, voice tone and prosody, word choice, physiological signals (heart rate, skin conductance, eye movement) where available, and behavioral patterns. By training on large datasets of human emotional expression and response, these systems can classify emotional states with accuracy that, in controlled conditions, rivals human perception.
Commercial applications have proliferated: hiring assessment platforms that analyze video interview affect; call center software that monitors customer emotional states in real time; educational technology that adjusts content delivery based on detected engagement or frustration; in-vehicle systems that monitor driver alertness. Research applications extend to mental health monitoring, pain assessment, and autism support.
The EU AI Act's Explicit Restrictions
The EU AI Act takes a notably strong position on emotion recognition AI. The Act prohibits the use of AI systems to infer emotions in workplace and educational settings — a category that encompasses the hiring assessment and educational technology applications that have attracted the most investment. The prohibition is grounded in concern that emotion inference is often inaccurate (particularly across demographic groups), that individuals cannot meaningfully consent when emotion recognition is a condition of employment or education, and that the potential for misuse and discrimination is severe.
Outside these prohibited categories, high-risk emotion recognition systems face substantial compliance obligations: documentation, transparency requirements, human oversight, and conformity assessment before deployment.
The Accuracy Problem
A substantial body of research challenges the foundational assumption of much affective computing — that emotions are reliably expressed through facial expressions and other physiological signals in ways that generalize across individuals and cultures. Work by psychologist Lisa Feldman Barrett and others in the "theory of constructed emotion" school argues that emotional expression is highly variable and culturally mediated, making universal emotion recognition models scientifically suspect. These critiques have influenced regulators and generated significant controversy within the affective computing research community itself.