Point a camera at a face and emotion AI returns a number: how happy, angry or engaged the person is. The face below is smiling, and the software is sure it means happiness. Tap through four real situations that all produce this exact smile. The reading never changes. What the person actually feels does.
Illustrative demo. The label is a stand-in score, not a vendor figureEmotion AI reads
Happiness
91% confident
Tap a situation. The face and the score above will stay exactly as they are. Watch what happens underneath.
Four situations, one unchanging readout. The movement of the face was the same every time, so the model reported the same confident number every time. The feeling underneath was different in each case, and the model could not see it.
That is the whole trap. A confident emotion score describes how a face moved, not what a person felt (Krumhuber et al., 2023). Treat the score as a hypothesis about a feeling, never as a reading of one.
The face and the score are a fixed illustration, not a live model call, which a self-contained embed cannot make. The claim they carry is the article's: a facial movement does not map one-to-one onto a felt emotion (Barrett et al., 2019; Krumhuber et al., 2023). The 91% is the kind of figure vendors quote, and it measures agreement with an expression label, not proof of what anyone felt (Kopalidis et al., 2024).