What kind of study it is
Study at a glance
An AI streamer is an AI that takes the human role of selling products in a livestream.
An analysis of 347 sessions
The study covers 347 AI streamers on a large livestreaming platform. It combines geometric morphometrics, which captures facial shape numerically, with machine learning and econometric modeling that statistically separates the effect of each factor, estimating the relationship between facial features and consumer engagement (the size of viewer response). What distinguishes it is that it uses data from streams actually running, not laboratory subject tests.
"This research combines geometric morphometrics, machine learning, and econometric modeling to analyze 347 AI streamers from a large livestreaming platform"(Abstract)／"We empirically test this framework using a dataset of 347 livestreaming sessions from a leading e-commerce platform."(Introduction) — from the Journal of Business Research
Breaking facial attractiveness into four measures
The crux of the study is that it did not treat facial attractiveness as one thing. Physical attractiveness was split into asymmetry and typicality, emotional attractiveness into richness and expression intensity, and all four were measured separately. Typicality is how close a face is to an average, familiar one. In human interaction, symmetry and averageness have been treated as conditions of attractiveness.
"exploring how physical (asymmetry and typicality) and emotional (richness and expression intensity) dimensions of facial attractiveness affect consumer engagement"(Abstract) — from the Journal of Business Research
The human rule of thumb reverses
The four measures against engagement
The results run opposite to what research on human faces has said.
Asymmetry helps; a typical face hurts
Facial asymmetry was positively associated with engagement, while a typical face had a negative effect. In humans, symmetry and averageness have been treated as conditions of attractiveness, so the direction is simply reversed. The authors explain this by noting that in synthetic agents, symmetry and typicality function as signals of artificiality. Asymmetry, conversely, works as a cue that disrupts mechanical perfection.
"Contrary to previous studies on human facial aesthetics, which highlight symmetry and averageness as critical to perceived attractiveness, this study finds that facial physical asymmetry is positively associated with consumer engagement, while facial physical typicality and facial expression intensity negatively influence engagement."(Abstract)／"While symmetry and typicality signal genetic fitness in humans, we show that in synthetic agents, these cues function as signals of artificiality."(Introduction)／"facial physical asymmetry may serve as an important cue that disrupts perceptions of mechanical perfection"(Introduction) — from the Journal of Business Research
Turning up the expression does not work
The emotional side also cuts against intuition. Expression intensity had a negative effect on engagement, and emotional richness showed no significant effect. With a human salesperson, it is easy to assume that more expressive means better selling. The results suggest that when an AI produces intense expressions, they may be received not as warmth but as something performative.
"Facial emotional richness has no significant effect."(Abstract)／"we find that facial expression intensity triggers avoidance, while emotional richness yields a null effect."(Introduction)／"it is unknown whether high-intensity emotional displays by AI agents are perceived as genuine warmth or as inauthentic, performative algorithmic outputs"(Introduction) — from the Journal of Business Research
The conditions that matter in practice
Two conditions that shift the effects
The effects are not uniform; two conditions change them.
More human-looking amplifies the effects
The closer the avatar's appearance is to a human, the stronger the asymmetry and typicality effects become. The influence of expression intensity, conversely, weakens. With a stylized character, how the face is built matters less; the closer you push toward photorealism, the more an over-symmetric face drags results down.
"Furthermore, high form anthropomorphism amplifies the effects of asymmetry and typicality, while weakening the impact of facial expression intensity."(Abstract) — from the Journal of Business Research
The best answer depends on the product
The other condition is what is being sold. For hedonic products the positive effect of asymmetry was stronger, while for utilitarian products the penalties attached to a typical face and intense expressions stood out more. In other words, distinctive features work for apparel and indulgences, whereas for everyday and functional goods, a safe, symmetric face plus strong expressions becomes a handicap. When designing an AI face, starting from the character of the product is the sound order.
"Notably, the positive impact of asymmetry is stronger for hedonic products, whereas utilitarian products accentuate the negative effects of typicality and facial expression intensity on consumer engagement."(Abstract) — from the Journal of Business Research
When you want to turn an English web page into markdown with its headings and tables intact so it reads more easily on your own machine, the following tool can help.
