Auditing the impact of social media’s policy shift on anti-vaccine discourse: A large language model-driven empirical study

Abstract

The sudden termination of X’s (formerly Twitter) misinformation policy on November 23, 2022, provides an opportunity to assess the effects of lifting content moderation restrictions on vaccine-related discourse. This study examines changes in the prevalence, thematic composition, and engagement of anti-vaccine discourse following X’s policy shift, analyzing tweets from a seven-day period before and after the policy termination (November 16–30, 2022), excluding the announcement date itself from regression analyses. Using GPT-4o for stance classification, thematic categorization, and stance consistency assessment, with validation through external benchmarks and cross-annotator agreement, we find that anti-vaccine tweets increased significantly post-policy (OR = 1.60, 95% CI: 1.50–1.72), particularly via retweets, suggesting content amplification. Sensitivity analyses excluding highly retweeted content revealed that the policy change was also associated with increased creation of new anti-vaccine content. Thematically, health concerns over vaccination became more prominent, while conspiracy-related and anti-mandate narratives declined in relative prevalence. Stance consistency in quote tweets increased, indicating reinforced ideological alignment in anti-vaccine discourse. These results suggest that content moderation policies may constrain both the volume and amplification of anti-vaccine content, with policy removal associated with rapid shifts in discourse patterns.