Beyond the Social Graph: Simulating Algorithmic Content Curation Effects on Opinion Dynamics with LLM Agents

ACL ARR 2026 January Submission3364 Authors

Keywords: LLM-based Agents, Opinion Dynamics, Social Polarization

Abstract: Algorithmic content curation shapes public opinion, yet traditional opinion dynamics models largely overlook the interplay between algorithms, rich content, and user cognition. We propose a platform-centric simulation framework integrating Large Language Model (LLM) agents into an opinion dynamics setting mediated by content-based interactions. Users are modeled as heterogeneous agents with dynamic stance and sentiment, exposed to content curated via random, popularity-based, and steering strategies. By simulating these dynamics on two real-world datasets---a polarized election and a negative news event---we demonstrate that while steering strategies can effectively shift aggregate opinion, they exacerbate polarization. Conversely, popularity-based algorithms lead to severe traffic concentration and unequal exposure. Furthermore, we analyze how user traits like stubbornness and activity level, along with the presence of social comments, modulate these effects. Our work provides a data-driven approach to understanding platform governance and its impact on the information ecosystem.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: Computational Social Science and Cultural Analytics

Languages Studied: English

Submission Number: 3364