Study links ChatGPT era to diverging political speech homogenization on Reddit

Researchers at arXiv have released a new paper titled "Asymmetric Discourse Homogenization and Shared Language Technology: Evidence from Reddit," authored by Fengming Liu. The study examines 6 million Reddit comments from two cross-partisan forums spanning 2019 to 2025, tracking how political discourse diversified or converged over time. The headline finding: conservative users experienced an interruption of their prior diversification trajectory starting around late 2022, while progressive users showed no comparable change in their speech patterns. This asymmetry held across multiple estimation strategies, including interrupted time series, difference-in-differences, regression discontinuity, and propensity-score matching.

Why does this matter for AI adoption? The paper tested whether the ChatGPT launch specifically caused the shift. A daily-frequency permutation test over 2,377 candidate cutoff dates found the ChatGPT threshold produced an unremarkable estimate at the 49.8th percentile—meaning the homogenization grew gradually rather than breaking sharply at a single date. The authors also built a cumulative LLM exposure index tracking seven AI model releases, which stayed significant even after accounting for time trends. Importantly, a 'stayer' analysis that restricted the sample to authors active throughout the entire study period made the homogenization effect disappear, indicating the convergence is ecological (community-level discursive drift) rather than a story of individuals adopting AI to write their posts. The paper includes a full replication package with code and instructions.