Reranking partisan animosity in algorithmic social media feeds alters affective polarization

TL;DR

The paper introduces a platform-independent, real-time feed reranking approach using a browser extension and LLM-based scoring to causally study how algorithmic curation affects affective polarization. In a preregistered field experiment on X with 1,256 US participants around the 2024 campaign, exposure to Antidemocratic Attitudes and Partisan Animosity content (AAPA) was manipulated to induce higher or lower polarization. Results show that reduced exposure to AAPA post content increases warmth toward the out-party by about 2 degrees, while increased exposure decreases warmth by about 2.5 degrees; in-feed surveys captured corresponding shifts in negative emotions, with effects largely short-lived and bipartisan. The study demonstrates a scalable, independent method for auditing feed-ranking interventions in naturalistic settings, with implications for platform accountability and democratic discourse, while highlighting trade-offs with engagement for downranking strategies.

Abstract

Today, social media platforms hold sole power to study the effects of feed ranking algorithms. We developed a platform-independent method that reranks participants' feeds in real-time and used this method to conduct a preregistered 10-day field experiment with 1,256 participants on X during the 2024 U.S. presidential campaign. Our experiment used a large language model to rerank posts that expressed antidemocratic attitudes and partisan animosity (AAPA). Decreasing or increasing AAPA exposure shifted out-party partisan animosity by two points on a 100-point feeling thermometer, with no detectable differences across party lines, providing causal evidence that exposure to AAPA content alters affective polarization. This work establishes a method to study feed algorithms without requiring platform cooperation, enabling independent evaluation of ranking interventions in naturalistic settings.

Figures (27)

• Figure 1: Overview of the field experiment on X. (A) Participants completed a pre-experiment survey to rate their feelings toward the opposing party and report their emotions. Then, they were randomly assigned to one of the two parallel experiments, reduced or increased AAPA content exposure, and further randomized into treatment and control groups. After ten days, they completed a post-experiment survey. (B) Timeline of the experiment. During the first three days, no intervention was applied to measure the participants' baseline responses. For the next seven days, participants in the treatment groups received increased or decreased AAPA exposure. In-feed surveys were periodically shown throughout the baseline and the intervention period. (C) The interventions downranked or upranked AAPA posts in the participants' feeds, depending on their treatment assignment.
• Figure 2: Effects of reducing and increasing exposure to AAPA content in participants' feeds on their feeling towards the out-party (feeling thermometer scale between 0: cold, and 100: warm) relative to the corresponding control group. Participants were surveyed within the feed during the intervention and off-platform after the experiment. The error bars represent 95% confidence intervals.
• Figure 3: Effects of reducing and increasing exposure to AAPA content in participants' feeds on their experiences of emotion, relative to the corresponding control group. Participants were surveyed within the feed during the intervention (scale ranged from 0: "none at all" to 100: "extremely") and off-platform after the experiment (scale ranged from 1: "never" to 5: "all the time"). The filled circles represent statistical significance ($P<0.05$, adjusted for multiple hypothesis testing), and the error bars represent 95% confidence intervals.
• Figure S1: Summary of the participants' recruitment funnel.
• Figure S2: Distribution of the AAPA scores assigned to political posts in the "For you" feed on X (pre-intervention). The green bars represent the fraction of non-AAPA, while the orange bars describe the AAPA posts recommended by X (not necessarily seen by the users).