New simulator shows social media moderation fails in real-time

A team of computer scientists led by Enrico Verdolotti and colleagues has developed an empirically grounded extension of the SimSoM (Social Media Simulation) model that reveals critical flaws in how platforms evaluate content moderation policies.

By calibrating their simulator using real-world COVID-19 vaccine discourse data and optimizing parameters via CMA-ES (Covariance Matrix Adaptation Evolution Strategy), the researchers created a tool that accurately reproduces key statistical signatures of online behavior, including activity distributions and temporal patterns. When they tested established misinformation detection methods—both retroactively (static) and in real-time (dynamic)—they found static evaluations consistently overestimated the effectiveness of user bans. Specifically, dynamic moderation showed that compensatory resharing by remaining users reduced the expected decline in low-quality content by over 40% compared to static estimates, highlighting the limitations of current evaluation practices.