Emotions, Group Cues, and Algorithmic Interventions: An Agent-Based Model of Elite Retweeting on Twitter

<p>Misinformation interventions are commonly targeted at ordinary users, despite political elites having an out-sized role in misinformation amplification, and in shaping public opinion. Individual-level interventions for ordinary users may be ineffective for political elites, whose interests and motives for sharing differ substantially. In this study, we develop an empirically grounded agent-based model of intra-elite retweeting behavior to evaluate content-reranking interventions aimed at reducing the spread of tweets linking to untrustworthy news websites. Using observational data from 2021 to early 2023, we estimate a retweeting submodel that captures how emotions, group references, partisan alignment, and domain quality influence amplification decisions among U.S. Members of Congress. We embed this decision model within a directed network of elite agents and their partisan audiences to model information diffusion. We then test seven interventions using two main criteria: multiplicative discernment (the relative amplification of trustworthy vs. untrustworthy information) and relative change in total retweet volume (changes in overall retweeting volume compared to the baseline condition). We find that only directly targeting tweets with links to low-quality news websites meaningfully reduced the spread of untrustworthy information. No indirect intervention, like down-ranking angry tweets or down-ranking politicians with an extreme ideology, was effective at improving multiplicative discernment. Moreover, the intervention that improved discernment also reduced the total retweet volume, revealing an inherent structural tension between platform incentives and public interest goals. This trade-off can help explain why effective interventions may remain unimplemented under current incentive structures for social media platforms.</p>