“Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements,” by Lisa Feldman Barrett, Ralph Adolphs, Stacy Marsella, Aleix Martinez, and Seth Pollak, appeared in Psychological Science in the Public Interest in 2019 (volume 20, issue 1, pages 1 to 68). It is the most cited scientific challenge to a premise that underlies a whole class of commercial AI: that you can read a person’s emotional state from their face.
The authors note that this assumption is not academic - it “influences legal judgments, policy decisions, national security protocols, educational practices, psychiatric diagnosis and treatment,” and many commercial products. Vendors sell systems that claim to detect anger, fear, or deception from facial expressions in job interviews, classrooms, and border crossings. The review examined the actual scientific evidence behind the common view, focusing on six popular emotion categories: anger, disgust, fear, happiness, sadness, and surprise.
Their conclusion was that the evidence does not support reliable inference. People do not reliably scowl when angry or frown when sad, and the same facial movement can express different things in different contexts and cultures. The review framed this as a problem of limited reliability and specificity: a given configuration of the face is neither a consistent signal of a given emotion nor specific to it. The authors called for new research on how people actually move their faces in everyday life, rather than assuming a fixed dictionary of expressions.
Why business readers should care: emotion-recognition or “affect detection” AI is built on the very assumption this review found unsupported. The paper became a key reference for companies and regulators pulling back from the technology - it was cited in the wave of facial-analysis restrictions, and the EU AI Act later moved to ban emotion inference in workplaces and schools. A product can be technically accurate at measuring face muscles and still be measuring something that does not mean what it claims.