Algorithmic Intervention Decreases Polarization in Content Moderation

Abstract

Crowd-based moderation relies on people's collective judgments to combat misinformation, but raters bring their own ideological biases to the task, shaping how they moderate each other's contributions. We ask whether an algorithmic redesign can counteract this bias, using X's Community Notes, where contributors write context for potentially misleading posts and rate each other's notes. Community Notes replaced simple majority voting with an algorithm that instead rewards agreement across people with different ideological viewpoints. Comparing 8,805 notes and 4,367 contributors before and after this change, we find that ideological polarization decreased in the algorithm's estimates of contributor and note positions. Although negative interactions increased, contributors became less structurally aligned, indicating that negative interactions were less organized along partisan lines. These findings show that algorithmic incentives can reshape both the social structure and behavioral dynamics of crowds, extending beyond their intended effects on content moderation.