Scroll through a livestreaming shopping app in China these days and you may find yourself watching a host who never blinks in the wrong direction, never takes a break, and never has a bad day. That is because the host is not a person at all, but an artificial intelligence (AI) streamer, a computer-generated figure built to pitch products around the clock. As these synthetic salespeople spread across platforms like Taobao and TikTok, retailers face a practical question: what should their digital host actually look like?
A team of researchers set out to investigate one slice of that question, focusing on the face. Their work, published in the Journal of Business Research, offers evidence that some of the facial traits humans find appealing in other humans can backfire when they appear on a machine.
The puzzle of the synthetic face
Decades of research on human attractiveness point in a consistent direction. People tend to rate faces as more appealing when they are symmetrical and when they look “typical,” meaning close to the average of many faces blended together. Psychologists have long linked these preferences to instinctive readings of health, genetic fitness, and trustworthiness.
But an AI streamer is not a person. It is a synthetic figure generated by algorithms, and the researchers wondered whether the usual rules still apply. Their concern draws on an idea called the uncanny valley, first described by roboticist Masahiro Mori in 1970. The theory holds that as an artificial figure gets closer and closer to looking human, there is a point where near-perfection starts to feel eerie or unsettling rather than reassuring.
Miyan Liao of Chongqing Technology and Business University and her colleagues at the University of Electronic Science and Technology of China and Henan University suspected that a flawless synthetic face might trip that alarm. A perfectly symmetrical, perfectly average AI face, they reasoned, could read as “engineered” and fake, while small imperfections might make it feel more genuinely alive.
Measuring faces with math and machine learning
To test these ideas, the team gathered data from 347 livestreaming sessions on Taobao, each featuring a different AI streamer, recorded between February and July of 2024 across all twelve of the platform’s product categories. They then broke facial appeal into two categories.
The first was physical attractiveness, which they split into two measurable traits. Facial asymmetry captured how much the two halves of a face differed, calculated by mapping 68 points on each face and measuring the distances between matching landmarks on either side of the midline. Facial typicality captured how closely a face resembled the average of all the faces in the sample.
The second category was emotional attractiveness, also split in two. Emotional richness counted how many distinct expressions, such as happiness or surprise, a streamer displayed during a session. Expression intensity measured how strong those expressions were, gauged by how far the face moved from a neutral resting state. To capture this, the researchers pulled a frame every ten seconds from each stream, producing more than 22,000 images that they analyzed using facial-recognition software.
For their outcome, the team built a combined measure of consumer engagement, pooling likes, shares, new followers, new members, and the scrolling “bullet screen” comments that fly across the screen during Chinese livestreams. They then ran a statistical model suited to this kind of count data, controlling for factors like the number of viewers, video length, product prices, and the streamer’s gender and posture.
What the analysis revealed
The findings ran counter to the human beauty playbook. Facial typicality was linked to lower engagement, and it had the strongest effect of any trait the team measured. In other words, the more average and generic an AI face looked, the less audiences responded to it. Strong facial expressions also dragged engagement down. Asymmetry, meanwhile, showed a small positive link to engagement, the opposite of what the human literature would predict.
Emotional richness, the sheer variety of expressions, had no measurable effect at all. The researchers interpret this null result through what they call a “zone of indifference.” In the fast-moving, sales-driven setting of livestreaming, they suggest, viewers may simply tune out an AI’s emotional range, treating it as background noise rather than a meaningful social signal. They are careful to note this interpretation is speculative and would need experiments to confirm.
The authors read the broader pattern as a reversal of established beauty cues. In human faces, symmetry and typicality signal fitness and trust. In synthetic faces, they argue, those same traits may signal artificiality, while a touch of asymmetry works as an authenticity cue that nudges the AI back out of the uncanny valley.
When context changes the message
The study also examined two conditions that shaped how these facial cues played out. The first was form anthropomorphism, meaning whether the AI looked human-like or cartoon-like. The second was product type, splitting goods into utilitarian items like appliances and electronics versus hedonic items like food, cosmetics, and jewelry.
Both mattered. When the AI streamer looked strongly human-like, the benefit of asymmetry grew stronger and the penalty for a typical face grew harsher. The researchers interpret this as the uncanny valley at work: the closer a face gets to looking human, the more a subtle imperfection helps and a generic appearance hurts. High human-likeness also softened the penalty for intense expressions, suggesting viewers grant more emotional latitude to figures that already look like people.
Product type shifted things too. The positive effect of asymmetry was stronger when the streamer was selling hedonic products, the kind tied to pleasure and self-expression, where a distinctive face seems to align with a shopper’s desire for something unique. For utilitarian products, the picture flipped in a way the researchers found striking: asymmetry actually reduced engagement. They interpret this as a change in what shoppers want. When buying practical goods, people prize reliability and competence, and a lopsided face may read as sloppy or imprecise rather than charming.
What it might mean for firms
For companies designing AI hosts, the study points toward a few practical ideas. The authors suggest that building in moderate asymmetry and distinctive features, through small tweaks to eye alignment, brow shape, or mouth curve, may help a synthetic streamer stand out and feel more trustworthy, especially for pleasure-oriented products. They also caution against over-the-top expressions, recommending that emotional intensity be dialed to match audience expectations, which may vary across cultures.
The researchers frame their overall takeaway as a warning against chasing maximum realism. As they put it, “a simple strategy of maximizing human-likeness is not only suboptimal but may be detrimental,” arguing that commercial success rests “not on its perfection, but on its plausibility.”
Several caveats deserve attention. The data came from a single platform in one cultural setting, and the authors note that norms around emotional display differ across cultures, so an “imperfection” that reads as authentic in one place might not travel. The study is also observational, meaning it can show that certain faces were associated with more engagement but cannot prove the faces caused that response. The team ran instrumental-variable tests to address this and reported that their main results held, but they call for controlled experiments to open the “black box” of what viewers are actually feeling. For now, the work suggests that when it comes to a machine’s face, a little imperfection may go a long way.
