From Simulated Online Communities to Hybrid Publics with YSocial: Operational Validation and Experimental Evidence

Day - Time: 01 July 2026, h.14:00

Place: Area della Ricerca CNR di Pisa - Room: C-29

Speakers

• Aleksandar Tomasevic (Institute of Physics Belgrade, University of Belgrade, Serbia)

Referent

Giulio Rossetti

Abstract

Large language model agents are increasingly used to simulate social systems, but their validity depends on more than plausible text. A credible social simulation must generate observable interaction processes: activity rhythms, reply networks, topic structure, toxicity, repeated encounters, and the failures that reveal missing mechanisms. I will present YSocial, a platform for simulating Reddit-like communities, through two connected studies.The first study operationally validates YSocial against 30 matched 30-day windows from Voat’s v/technology. Across 30 independent simulations, YSocial reproduces several aggregate regularities, including root-post volume, unique users, daily active users, large connected component size, and broad topic coverage. Its failures are equally diagnostic: comments and thread size are too high, toxicity is misallocated across posts and replies, and the interaction core is too large and diffuse. These gaps point to specific engineering targets, especially durable memory, activity concentration, reply ecology, and content-layer calibration.The second study turns YSocial into a 21-day human-AI field experiment: human participants and autonomous memory-enabled LLM agents shared a Reddit-like popular-culture community. In one condition, AI accounts were labeled as AI; in the other, the labels were hidden. The setting generated a detailed social trace: participants discussed live pop-cultural topics, formed recurring interaction patterns, argued, exchanged toxic or confrontational replies, and encountered agents that remembered prior exchanges.The main result is straightforward: when labels were hidden, participants interacted with AI accounts more often and more repeatedly.In same-thread situations where people could reply to human or AI content, AI-authored comments were more likely to be chosen as reply targets. This pattern remained after using browser logs to account for whether comments were actually visible, and for how long. At the user level, human-to-AI contact became more concentrated around a few AI accounts and more likely to repeat. At the network level, the hidden-label condition had more human-AI ties, while human-human interaction still remained the core of the public. Together, these results show that disclosure labels shape whether AI accounts become socially available interaction partners in hybrid online communities.The strength of the evidence comes from the setting itself: the effects appeared in a live platform with persistent identities, public threads, reputation, memory, topical drift, conflict, and repeated encounters.The experiment therefore captures human-AI interaction as an evolving public setting with enough social texture to study disclosure, uptake, memory, and conflict as connected processes.