Kirak Kim, Hyojin Kim, Yejin Son, Sungyoung Kim, Kyung Myun Lee
6 min
This study investigates whether assigning distinct personas to LLM agents can recreate the sense of collective presence typically missing from asynchronous online concert viewing. The authors developed a multi-agent system where ten LLM agents generate real-time chat reactions to a K-pop performance video. They compared a persona-conditioned condition—where each agent was assigned a unique fan identity, bias, and chat style—against a no-persona baseline where agents lacked individual traits. The study used a within-subjects pilot with 11 K-pop fans to evaluate both model-level output quality and user-level subjective experience.
Persona conditioning successfully achieved its technical goal: it produced highly differentiated agent behaviors, reduced repetitive chat patterns, and increased overall output diversity. Participants explicitly recognized the persona-conditioned audience as more natural than the baseline. However, this increased naturalness failed to improve the users' sense of social connectedness, engagement, or affective response. The quantitative results showed no significant differences in these metrics between the two conditions, suggesting that while the agents appeared more realistic, they did not foster a more meaningful collective experience.
This research highlights a critical gap between the technical performance of AI agents and their ability to facilitate meaningful human social experiences. The authors suggest that online concert chat functions more as a collective monologue than an interpersonal dialogue. Consequently, simply making agents appear more human-like through persona prompting is insufficient; true collective experience may require deeper alignment between AI behavior, specific fandom identities, and the cultural expectations of the audience. The study underscores the need for domain-specific, culturally situated evaluation when designing AI agents for social contexts.
A concert is a collective experience, but recorded performance videos are typically watched alone, stripping away the shared audience presence that makes concerts feel eventful. We investigate whether persona-based LLM audience agents can recreate aspects of this collective experience by generating real-time fan chat alongside a K-pop performance video. We present a multi-agent system in which ten LLM agents react through live-chat messages, comparing a persona-conditioned audience (each agent assigned a distinct fan identity, bias, and chat style) with a no-persona baseline. In a within-subjects pilot with K-pop fans (N=11), persona conditioning substantially improved model-level chat quality and perceived naturalness, but did not translate into differences in social connectedness, engagement, or affective response. Interviews suggest that online K-pop concert chat may operate as collective monologue rather than interpersonal dialogue, and that meaningful participation depends on shared identification with the specific artist and fandom. Persona conditioning can make LLM audiences appear more natural, but culturally meaningful collective experience may require deeper alignment between persona, crowd behavior, fandom identity, and user expectations.
Alex: But if ten agents are all trying to talk at once, doesn't it just collapse into noise?
Sam: That's actually a known failure mode they had to design around. It's called "output collapse"—when all the agents start saying the exact same thing, like a room full of people simultaneously shouting "Wow!" To prevent it, each agent gets a personal memory of what it said recently, so it doesn't just repeat itself. They also share a common chat log, so agents can see what the others—and the human viewer—are saying.
Alex: And how did they test whether any of this actually worked?
Sam: They ran a pilot study with eleven K-pop fans. Each participant watched the same four-minute performance twice: once with the persona-conditioned agents, and once with a baseline version where all the agents were identical. Afterward, they asked participants how connected they felt to the audience and how natural the experience seemed.
Alex: So did the personalized agents make fans feel more connected?
Sam: This is where the results become genuinely interesting. The persona-conditioned agents were rated as significantly more natural-feeling than the identical ones. Participants noticed they seemed more like real people. But—and this is the key finding—that increased naturalness didn't translate into stronger social connection or higher engagement. People could tell the agents felt more human, but they didn't feel more like part of a real community.
Alex: So the naturalness didn't actually produce a better social experience.
Sam: That's exactly it. The interviews pointed to something the researchers describe as a "collective monologue." In online K-pop concert chats, people aren't really having conversations with each other—they're expressing excitement simultaneously, like a stadium crowd cheering in unison. Everyone is broadcasting their own reaction, not responding to anyone else's.
Alex: Oh—so that's why the individual personas didn't matter as much? The fans weren't trying to make friends with specific AI agents. They were looking for the overall vibe of a crowd.
Sam: Exactly. What participants cared about was the collective atmosphere—the trends, the repetitive cheers, the shared energy of the fandom as a whole. Making an individual AI agent seem more human doesn't automatically create that sense of shared identity. The "soul" of the fandom, as one might put it, is a group property, not an individual one.
Alex: It's like being at a concert. You don't need to know the life story of the person standing next to you to feel the energy of the crowd. You just need to know you're both there for the same reason.
Sam: That's a useful way to frame it. The researchers conclude that meaningful participation in these spaces depends on shared cultural context. If you don't know the fandom's specific slang, or the history of the group, the AI crowd just feels like noise—no matter how realistic the individual agents seem.
Alex: So the vibe of the collective matters more than the realism of any single agent.
Sam: That's the takeaway. If we want AI audiences that genuinely foster a sense of community, the research suggests we need to focus on the collective dynamics of the fandom—not just on making individual agents seem more lifelike. It's a meaningful step toward understanding how technology might bridge the gap between watching something alone and feeling like you're experiencing it together.
Alex: That's a thought-provoking result—the thing that makes a crowd feel real isn't any one person in it, but the shared energy between all of them. Thanks for listening to ResearchPod.