People expressed less emotion to virtual influencers—but Instagram comments cannot explain why
Across 156,918 public comments, both positive and negative language appeared less often around AI-created personas than human influencers. The observational, model-labelled data do not measure individual feelings or causal effects.
The short version
- What happened
- Researchers analysed 156,918 public Instagram comments that collectively received more than 12 million likes, using emotion analysis, knowledge graphs and machine learning.
- Who it affects
- People increasingly encounter synthetic personalities in entertainment and advertising. The study suggests their public conversations differ from those around human creators, but it does not show that individual users are less attached, more manipulated or less satisfied.
- What remains unproven
- Peer reviewed
Keep in mind
Peer reviewed
156,918 public comments
More than 12 million comment likes
Observational comparison
Automated emotion labels
No random assignment
No private content
No funding or conflicts reported
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 4 min The evidence, methods, implications and limitations
The difference was lower expression, not a wave of hostility
A peer-reviewed study from Bar-Ilan University compared how people responded to AI virtual influencers and human influencers on Instagram. The researchers analysed 156,918 public comments that together received more than 12 million likes, combining automated emotion analysis, language-model-derived knowledge graphs and behavioural features. Positive affect appeared in 24.8% of comments directed at virtual influencers, compared with 41.5% around human influencers. Negative affect was also less common: 5.8% versus 9.3%.
That pattern matters because it is easy to misread lower positivity as rejection. Both emotional directions declined, suggesting flatter or less expressive engagement on average rather than simply more negative reactions. The authors also report wider variation between individual virtual influencers, especially in positive expression. Their Multi-Signal Embedding Network classified whether engagement concerned a virtual or human influencer with an F1 score of 0.89, outperforming the comparison methods reported in the paper.[1]
Comments are traces of behaviour, not a controlled test of emotion
The denominator is large, but the design is observational. People chose which accounts to follow and whether to comment; the study did not randomly expose comparable users to otherwise identical virtual and human creators. Differences in topics, audiences, posting frequency, language, moderation, fame or commercial strategy could contribute to the measured gap. A comment also cannot reveal how strongly its author felt, whether they understood the persona as synthetic or whether they behaved differently offline.
The emotional and semantic labels were generated through computational methods, including OpenAI embeddings and language-model analysis. Those tools can detect recurring patterns at scale, but their categories are not direct psychological measurements. The public early-access article reports the overall comment and like counts, not a causal experiment or an independent clinical-style validation of every inferred emotion. The classifier's F1 score shows separation between the two groups; it does not establish why the groups differ or whether users formed weaker relationships.[1]
The distinction that matters
What the Instagram comparison can support
Large observational data reveal a pattern but cannot assign its cause.
Observed
Compare 156,918 public comments
Less positive and less negative expression around virtual influencers
Modelled
Combine emotion, semantic and behavioural features
Influencer type classified at F1 0.89
Not tested
Randomly assign matched users and posts
Causal effect on feelings remains unknown
Brands should not turn averages into assumptions about people
For media teams and advertisers, the result suggests that a human-looking synthetic persona does not automatically generate the same texture of conversation as a human creator. Raw likes or comment counts may hide a difference in what people express. Yet the group average should not become a claim that virtual influencers cannot build communities: some virtual accounts produced substantially different patterns from others, and the study did not measure long-term loyalty, purchasing or wellbeing.
The data came from public content accessed through the Meta Content Library; accounts were anonymised and no private comments or identifiers were reported. The institutional review board treated the work as exempt, and the authors declared no funding or competing interests. The assessment would change with preregistered comparisons of matched posts, human-validated emotion labels, disclosure experiments and repeated measures of the same users. For now, the study maps a difference in public language—not the inner emotional effect of synthetic personalities.[1]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 44 min 28 sec · Background explainer
Is this the AI future we want? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Published 2026-07-01. Source checked 14 October 2026. Check captions on the publisher’s video page.
Check the healthcare evidencePhoto: Daniel Sone / National Cancer Institute / Unsplash · Licence
Continue the story
Read next
Continue exploring affective computing: compare the finding and its evidence limits.
Can AI carry emotion across languages?
Partly, in one English-to-Korean experiment. GPT-4o-generated text and images produced responses aligned with intended emotions above chance, but same-language reference ratings were closer for 11 of 18 constructs.
Read the next reportExplore the full source and evidence record
What this means for people
- People increasingly encounter synthetic personalities in entertainment and advertising. The study suggests their public conversations differ from those around human creators, but it does not show that individual users are less attached, more manipulated or less satisfied.
Global context
Instagram audiences and influencer cultures vary across languages, markets and account types. The article describes a platform-level sample, but the public record does not establish that its averages transfer to every country or synthetic persona.
What the evidence does not yet show
- Peer reviewed
- 156,918 public comments
- More than 12 million comment likes
- Observational comparison
- Automated emotion labels
- No random assignment
- No private content
- No funding or conflicts reported
What to watch next
- Matched-post experiments
- Synthetic-identity disclosure
- Human label validation
- Language and region results
- Longitudinal communities
- Purchasing and wellbeing outcomes
- Platform replication
Living evidence record
Impact record IAI-0LLYX37
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
29 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Scientific Reports published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 29 October 2026
Research question, topics and regions
Differences in affective and semantic patterns between comments directed at virtual and human influencers
Topics: Virtual influencers · Instagram · Affective computing · Social media · Online behaviour
Regions: Israel · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.