LLM audience agents with personas feel more natural but don't boost social connection
Persona-based LLM chat agents made K-pop concert viewing feel more real, but fans still felt alone.
A new study from researchers at multiple Korean universities explores whether persona-based LLM agents can recreate the collective experience of watching a concert with an audience. The team built a multi-agent system with 10 LLM agents that generate real-time fan chat alongside a K-pop performance video. In a within-subjects pilot study with 11 K-pop fans, they compared a persona-conditioned audience (each agent given a distinct fan identity, bias, and chat style) against a no-persona baseline.
The results showed that persona conditioning substantially improved model-level chat quality and perceived naturalness. However, this did not translate into measurable differences in social connectedness, engagement, or affective response. Follow-up interviews suggested that online K-pop concert chat often operates as a collective monologue rather than interpersonal dialogue. Meaningful participation may depend on shared identification with the specific artist and fandom, not just realistic-looking agents. The work was accepted at the ICML 2026 Workshop on Culture x AI.
- 10 LLM agents with distinct fan personas generated real-time chat during a K-pop concert video
- Persona conditioning improved chat quality and naturalness but didn't boost social connectedness or engagement
- Interviews suggested online concert chat is collective monologue, requiring deeper cultural alignment with fandom identity
Why It Matters
Shows that making AI chat agents more human-like doesn't automatically create meaningful social experiences—context and culture matter.