Emotion recognition systems assign an emotional label, such as angry, anxious or happy, to a person based on signals they give off: a facial movement, a tone of voice, word choice or a heart rate. Contact-centre analytics, interview screening tools, driver monitoring and classroom engagement software have all used it.
The scientific base is weaker than the product claims. A 2019 review led by Lisa Feldman Barrett concluded that facial movements are not reliable or specific enough, across contexts, individuals and cultures, to diagnose an emotional state. People scowl when concentrating as well as when angry. The EU AI Act cites this limited reliability in Recital 44, has banned emotion inference in workplaces and education since February 2025 except for medical or safety reasons, and from August 2026 requires deployers elsewhere to tell people they are exposed to it.
Treat every reading as a hypothesis to check with the person. A system that asks “you sound tired, is that right?” uses emotion recognition safely. One that logs “agent showed low engagement” in a performance file does not, and in the EU it is unlawful.