Emotion AI: The Politics of Feeling

Photo by Hirzul Maulana (Unsplash)

When I first wrote about emotion AI, the question of whether AI could understand human emotion was more theoretical than a material reality. Fast forward to now, and what was once a dinner party debate has become a high-stakes question of safety and civil rights. Systems that detect, interpret, and respond to human emotion are moving from research labs into hiring offices, retail stores, automobiles, and police departments. Are we ready?

The Unsettled Science of Affective Computing

Affective computing, defined by MIT Media Lab professor Rosalind Picard in 1997 as “computing that relates to, arises from, or deliberately influences emotions”, stands on an incomplete foundation. The most widely used framework is basic emotion theory, which proposes that a small set of emotions is biologically universal and expressed through consistent facial configurations across all cultures. Its appeal is its simplicity: if emotions are universal and tied to specific facial muscle movements, they can be mapped, measured, and automated with AI.

But neuroscientist Lisa Feldman Barrett’s psychological constructivist view argues that emotions are not hardwired but constructed by the mind. They are shaped by culture, language, and personal experience. The same sensation might be experienced as excitement in one context and anxiety in another. Emotion is not simply a facial expression; it is a state shaped by context.

Algorithmic blind spots carry real consequences. A system designed on basic emotion theory may confidently report that a job applicant seems “disengaged” when that person is managing their expression differently due to cultural background, disability, or individual personality. Microsoft retired its general-purpose emotion detection API in 2023, citing the lack of scientific consensus and the inability to generalize the linkage between facial expression and emotional state across regions and demographics.

The Cost Is Not Distributed Equally

One of the most consequential deployments of emotion AI has been in employment screening. Research from the Berkeley Haas Center for Equity, Gender, and Leadership AI showed that 44% of AI video interview systems demonstrate gender bias, while 26% show both gender and race bias. An algorithm trained on data from white, male, native-English-speaking candidates will penalize everyone who deviates from that profile through the compounding effect of underrepresentation. These systems do the most harm to the candidates who already face the most barriers.

This pattern extends to neurodivergent people. Research from the University of Cambridge found that autistic participants displayed different facial expressions for the same internal emotional states compared to neurotypical individuals, yet emotion AI operates on assumed universal signals. These frameworks encode barriers into software.

Emotion AI deployment in retail raises the same structural problem. Walmart has patented emotion-sensing technology and has a documented history of deploying computer vision in its stores. When the ACLU asked 20 major retailers whether they scan customer faces, 17 refused to answer. For wealthier consumers, opting out of a surveillance-heavy retailer is a lifestyle choice. For low-income families who depend on stores like Walmart as the only affordable grocery option within a reasonable distance, there is no opt-out. Working-class people become the primary subjects of emotion AI experimentation. That is not a neutral outcome. It is a structural inequity embedded in deployment decisions.

Law enforcement has been among the most aggressive adopters of this technology, with documented consequences disproportionately affecting women and racialized minorities. The ACLU has confirmed that at least 14 people in the United States have been wrongfully arrested due to facial recognition technology. Several cities have moved to set terms for how AI can be used in policing, establishing precedent for further regulation.

Where the Benefits Are Real

AI systems trained on speech patterns and vocal tone can identify signs of depression or anxiety with reported accuracy as high as 90%, though these figures vary by dataset and methodology. A 2024 study in Nature Medicine suggested that AI could reduce the accessibility gap in mental health screening for people of minority ethnicity by providing more consistent and culturally responsive evaluation. For people in rural areas, low-income communities, or regions with acute shortages of mental health professionals, emotion AI tools could offer a meaningful additional layer of support.

Researchers and humanitarians are experimenting with how AI tools can empower oppressed populations. For example, under the oppressive rule of the Taliban, Afghan women have increasingly turned to AI chatbots to access companionship, education, and mental health support. Afghan women, forbidden from seeing male doctors or leaving the house without a male escort, have found in AI chatbots a source of medical advice and education. The distinction between these applications and the problematic ones matters.

Successfully deploying emotion AI in healthcare requires further development and strict human oversight. AI tools can help supplement existing healthcare systems, but are not a replacement for in-person care. MIT researchers have uncovered evidence that increased use of AI tools by medical professionals risks leading to worse health outcomes for women and ethnic minorities, pointing to the need for further research and development, and raising questions about how we quantify the potential benefits or consequences of AI deployment.

What Needs to Change

Legislation like Illinois’ Biometric Information Privacy Act and the EU AI Act’s restrictions on biometric surveillance in public spaces point toward what meaningful accountability can look like. The covert deployment of emotion-reading systems in spaces where attendance is functionally mandatory should require explicit disclosure and meaningful opt-out options.

Beyond regulation, another shift is needed. The people most likely to be harmed by biased emotion AI systems are the people whose knowledge, experience, and emotional expression are most absent from the training data. Including marginalized communities in the design and governance of these systems is how you build something that actually reflects the full range of human experience and human potential.