Can We Fix Social Media? Testing Using Generative Social Simulation (arxiv.org)

Researchers introduce "generative social simulation" — embedding large language models as agents inside an agent-based social platform — to test whether prosocial interventions can fix social media harms. In a minimal simulated platform where agents can post, repost and follow, the LLM-driven population self-organizes into three well-known dysfunctions: partisan echo chambers, concentrated influence among a small elite, and amplification of polarized voices (which the authors call a "social media prism" that distorts political discourse). The model reproduces these macro patterns from micro-level agent behavior and network growth, showing that realistic social dynamics emerge without hand-coded biases. The team experiments with six interventions (including switching to chronological feeds and deploying bridging/recommendation algorithms) and finds only modest improvements — sometimes backfiring. Their analysis points to a core mechanism: feedback between reactive engagement (what users respond to) and network growth (who gains followers) tends to amplify polarization and influence concentration. For AI/ML researchers and platform designers this method offers a powerful sandbox to evaluate policies and algorithmic designs before deployment, but the results caution that surface-level algorithm tweaks may be insufficient. Meaningful reform may require rethinking foundational platform dynamics and incentive structures rather than incremental recommendation changes.