Emotional and behavioral asymmetries in user responses to virtual versus human influencers

As AI virtual influencers (AIVIs) proliferate, understanding human-AI emotional engagement is vital for advancing affective computing. To examine whether AIVIs evoke comparable affective responses to human influencers (HIs), we analyzed 156,918 Instagram comments that received 12+ million likes, integratinglarge language model (LLM)-derived knowledge graphs, emotion analysis, and machine learning. HI-directed discourse shows more differentiated emotional expression and a pattern consistent with more cohesive semantic networks than AIVI-directed discourse. AIVI-directed comments are less likely to express positive affect (24.8% vs. 41.5%) and also less likely to express negative affect (5.8% vs. 9.3%), indicating reduced affective expression of either valence rather than heightened negativity. This group-level tendency is not uniform: affective engagement varies substantially across AIVIs, and that variation is more closely associated with positive than with negative expression. Our Multi-Signal Embedding Network (MSEN), incorporating behavioral metrics and OpenAI semantic embeddings, classified influencer types from behavioral, affective and semantic features (F1 = 0.89), outperforming baselines including a state-of-the-art zero-shot LLM. These findings identify a boundary condition for the Computers Are Social Actors (CASA) paradigm: human-like appearance can elicit social responses without producing affective and semantic response patterns as rich as those elicited by HIs.