Artificial Intelligence Is Misreading Human Emotion

Abstraction: Flawed universal emotion theory underpins commercial AI affect recognition

Key points:

• Paul Ekman's theory of 6 universal facial emotions is the scientific foundation of a projected $56B emotion-AI market, but a 2019 systematic review (Barrett) found no reliable evidence faces reliably reveal emotional states
• FACS (Facial Action Coding System, 1978) requires 75–100 hours of training per human coder; computer vision automated this at scale without resolving the underlying scientific validity problems
• Amazon Rekognition, Microsoft Face API, IBM Tone Analyzer, Affectiva, and HireVue all deployed affect recognition commercially; HireVue dropped facial analysis in 2021 after criticism
• Studies show affect recognition software rates Black faces as angrier and more contemptuous than white faces, even controlling for degree of smiling
• TSA's SPOT program ($900M cost) applied Ekman's microexpression theory at airports; GAO found no evidence of effectiveness and documented racial profiling
• Article adapted from Kate Crawford's Atlas of AI (2021); traces affect recognition roots to physiognomy and phrenology

Connections: Affectiva · Kate Crawford · Hirevue · Affect Recognition · Algorithmic Bias · Facial Recognition

Source: https://www.theatlantic.com/technology/archive/2021/04/artificial-intelligence-misreading-human-emotion/618696/