Comments directed at virtual influencers contained less positive and less negative emotional expression than replies to human creators in a Scientific Reports study published on 29 October. The observational finding concerns the sampled Instagram discussions: it does not show that using an avatar caused a change in the same audience’s response.[1]
Kathleen Jagodnik and Alon Bartal analyzed 156,918 public Instagram comments that collectively received more than 12 million likes. They combined emotion classification, large-language-model-derived semantic representations and network analysis. Within the sampled comments, positive affect appeared in 24.8% of responses to virtual influencers, compared with 41.5% for human influencers. Negative affect appeared in 5.8% versus 9.3%. Both emotional polarities were less frequent in the synthetic-influencer group.[1]
The most important result is not a like count
The 12-million-plus likes belong to the collected comments. They do not establish that virtual influencers were more or less popular overall, because the study did not randomly assign audiences to follow one type of creator. The reported percentages describe the emotion labels of comments in the selected dataset, not the proportion of all social-media users who feel positively about AI personalities. Inferring a population preference from those figures would require a different study design.[1]
Emotion labels in sampled Instagram comments
| Expression | Virtual influencers | Human influencers |
|---|---|---|
| Positive affect | 24.8% | 41.5%Source 1 |
| Negative affect | 5.8% | 9.3%Source 1 |
The authors also report richer differentiation of emotional language and more cohesive-looking semantic networks in human-influencer discussions. Their Multi-Signal Embedding Network combined behavioral, affective and semantic information to distinguish the two categories of influencer, reaching an F1 score of 0.89 in the study’s classification task. F1 measures a balance of precision and recall for the classification, not how accurately a machine has detected an individual commenter’s true inner state.[1]
A smaller emotional signal is not universal indifference
The study finds substantial variation between individual virtual influencers. Some generated personalities elicited more positive expression than others, and that variation was more closely associated with positive language than with negative language. That heterogeneity matters to brands and platforms: the synthetic label alone is too crude to predict how an audience will respond. Appearance, subject matter, posting frequency and the community already following an account may all be associated with the observed pattern.[1]
Because the analysis is observational, it cannot isolate why the comments differ. Human creators may disclose personal experiences, have offline histories or attract audiences with different expectations. Virtual accounts may be promoted in different formats, on different schedules and to different followers. The language-model emotion labels themselves introduce measurement uncertainty, particularly with irony, slang, multilingual replies and inside jokes.[1]
The paper’s conclusion is therefore more precise than a claim that people cannot form relationships with digital figures. Users clearly respond socially to computer-generated personalities. But human-like appearance by itself does not guarantee the same distribution or richness of the responses observed around human creators. The finding places a useful boundary around the idea that people always treat computers as social actors in exactly the same way as people.[1]
Sources & further reading
- Emotional and behavioral asymmetries in user responses to virtual versus human influencers Scientific Reports · 29 Oct 2026 · Accessed 30 Oct 2026