MOSAIC: Modeling Social AI for Content Dissemination and Regulation in Multi-Agent Simulations

TL;DR

MOSAIC tackles modeling content diffusion and misinformation regulation in large-scale social networks by integrating memory-enabled LLM-driven agents with a directed follower graph. The framework supports four moderation regimes (No FC, Community FC, Third-Party FC, Hybrid FC) and validates agent realism against human engagement patterns, showing scalability and reproducibility. Key findings indicate that hybrid/third-party moderation can reduce misinformation while boosting engagement, and that misinformation does not universally spread faster in the agent-based setting, contrasting with some human studies; engagement follows a power-law with a small set of actors driving most activity. The work contributes a flexible, open-source platform for evaluating governance interventions, informing policy, and advancing AI safety research in online information ecosystems. Overall, MOSAIC offers a practical testbed for simulating complex social dynamics, testing regulatory strategies, and understanding the interplay between content, moderation, and user engagement in digital ecosystems.

Abstract

We present a novel, open-source social network simulation framework, MOSAIC, where generative language agents predict user behaviors such as liking, sharing, and flagging content. This simulation combines LLM agents with a directed social graph to analyze emergent deception behaviors and gain a better understanding of how users determine the veracity of online social content. By constructing user representations from diverse fine-grained personas, our system enables multi-agent simulations that model content dissemination and engagement dynamics at scale. Within this framework, we evaluate three different content moderation strategies with simulated misinformation dissemination, and we find that they not only mitigate the spread of non-factual content but also increase user engagement. In addition, we analyze the trajectories of popular content in our simulations, and explore whether simulation agents' articulated reasoning for their social interactions truly aligns with their collective engagement patterns. We open-source our simulation software to encourage further research within AI and social sciences.

Figures (14)

• Figure 1: Overview of the MOSAIC, a multi-agent social simulation framework where agents interact in an environment mimicking a social network, form dynamic memory-based behaviors, and respond to misinformation using community-based, third-party, or hybrid fact-checking mechanisms. Personas are replicated from human surveys or generated using synthetic distributions. Memories are retrieved before an agent takes certain actions, and are updated after certain events.
• Figure 2: Average engagement received per post: Human vs. Agents. Our t-test validates that the difference in reaction patterns across the three engagement types are not statistically significant, suggesting that agents can simulate individual human reactions to social media feed realistically.
• Figure 3: Effectiveness of content moderation approaches in promoting factual content, across models. Positive values: factual content receives more engagement. Negative values: misinformation receives more engagement.
• Figure 4: A consolidated view of content engagement and fact-checking metrics, averaged across all models used in our experiments. The first four panels display engagement metrics—specifically, the sum of likes, comments, and shares—under each fact-checking condition. The last two panels show fact-checking metrics, which combine both the number of community notes and note ratings, for the two methods where these are applicable (Hybrid and Community-Based). All values represent the average behavior across models, providing a holistic summary of the system’s dynamics under each experimental setting.
• Figure 5: User engagement's best power-law fit. Engagement is defined as the sum of reposts, likes, and comments received by the user.