AI Streamer Faces: Too Perfect Loses Engagement

What the study measured across 347 AI streamers

An AI streamer is an AI character that appears in a livestream. Liao et al. asked how the face of such a character relates to the way viewers respond.

Stream data from a large livestreaming platform

The authors analyzed 347 AI streamers from a large livestreaming platform. In the introduction they describe the data as 347 livestreaming sessions from a leading e-commerce platform, so the number is counted in streamers in one place and in sessions in the other. Engagement here means the size of the viewer response.

The study combined three methods. Geometric morphometrics, which captures facial shape as numbers, and machine learning turned facial features into measurements. Econometric modeling, a statistical approach that separates the effect of each factor, then related those measurements to engagement.

Four facial measures instead of one attractiveness score

The study does not rate a face with a single attractiveness score. It splits attractiveness into two physical measures, asymmetry and typicality, and two emotional measures, richness and expression intensity. Typicality is how close a face is to an average one.

The paper sets this against earlier work on human facial aesthetics, which highlights symmetry and averageness as critical to perceived attractiveness. With four separate measures, each of those two cues, and two emotional cues besides, can be set against engagement.

Asymmetric faces drew more engagement, typical faces less

Three of the four measures were linked to engagement: asymmetry positively, typicality and expression intensity negatively. Emotional richness was the exception.

Asymmetry tracked higher engagement, typicality lower

Faces with more physical asymmetry were linked to more consumer engagement. Faces closer to the average, the typical face, were linked to less. A viewer who expects a polished, symmetrical host to do best would find the pattern reversed: in this data set the less perfect face went with the stronger response.

The abstract words the asymmetry result as an association, and the other two results are likewise links, not causes. None of the three shows what would happen if a particular face were changed.

Intensity tracked lower engagement; richness showed no link

Expression intensity was negatively associated with engagement. Emotional richness, the other emotional measure, had no significant effect.

So faces that showed stronger expressions did not go with a bigger response. For readers searching for a "perfect AI face", the reported links run the other way: more typical faces and more intense expressions went with less engagement.

Avatar realism and product type change the effects

The effects are not the same for every AI streamer or every product. The authors report two conditions that change them.

Human-like form amplifies asymmetry and typicality effects

Form anthropomorphism is how human-like the AI streamer's form is. When it is high, the effects of asymmetry and typicality are amplified, and the impact of expression intensity is weakened.

In practice the asymmetry and typicality measures count for more on streamers whose form is more human-like. The paper's summary does not say more than that about what counts as a high or low level, so none is assumed here.

The effects differ for hedonic and utilitarian products

A hedonic product is one people buy mainly for pleasure. A utilitarian product is one they buy mainly for function.

For hedonic products the positive effect of asymmetry was stronger. For utilitarian products the negative effects of typicality and expression intensity on engagement were stronger. The abstract names no particular products for either type.

How the authors explain the reversal from human faces

The introduction gives the authors' reading of why the human-face rule flips. It is their interpretation of the pattern.

Symmetry and typicality read as signals of artificiality

In humans, symmetry and typicality signal genetic fitness. The authors argue that in synthetic agents the same cues work as signals of artificiality. On this reading, the cues that signal fitness in a person mark an AI character as artificial, which fits the negative link with typicality.

Asymmetry may make an AI face look less machine-made

The authors suggest that facial asymmetry may disrupt the impression of mechanical perfection. They offer this as one way to read its positive link with engagement. They also tie it to the uncanny valley effect, the unease people feel toward faces that look almost but not quite human, and say the cue may ease that feeling when the form is highly human-like.

The paper asks whether AI emotion reads as genuine warmth

The introduction says that facial expression intensity triggers avoidance, while emotional richness yields a null effect. It also leaves a question open: it is unknown whether high-intensity emotional displays by AI agents are perceived as genuine warmth or as inauthentic, performative output.

The study measured the link between intensity and engagement. It did not settle how viewers interpret strong expressions.

The abstract and introduction give no recipe for building an AI face. The evidence here supports the associations above, not a design rule.