A new experiment using an AI-powered browser extension to reorder feeds on X (formerly Twitter), and conducted independently of the X platform’s algorithm, shows that even small changes in exposure to hostile political content can measurably influence feelings toward opposing political parties – within days of X exposure. The findings provide direct causal evidence of the impact of algorithmically controlled post ranking on a user’s social media feed. Social media has become an important source of political information for many people worldwide. However, the platform’s algorithms exert a powerful influence on what we encounter during use, subtly steering thoughts, emotions, and behaviors in poorly understood ways. Although many explanations for how these ranking algorithms affect us have been proposed, testing these theories has proven exceptionally difficult. This is because the platform operators alone control how their proprietary algorithms behave and are the only ones capable of experimenting with different feed designs and evaluating their causal effects. To sidestep these challenges, Tiziano Piccardi and colleagues developed a novel method that lets researchers reorder people’s social media feeds in real time as they browse, without permission from the platforms themselves. Piccardi et al. created a lightweight, non-intrusive browser extension, much like an ad blocker, that intercepts and reshapes X’s web feed in real time, leveraging large language model-based classifiers to evaluate and reorder posts based on their content. This tool allowed the authors to systematically identify and vary how content expressing antidemocratic attitudes and partisan animosity (AAPA) appeared on a user’s feed and observe the effects under controlled experimental conditions.
In a 10-day field experiment on X involving 1,256 participants and conducted during a volatile stretch of the 2024 U.S. presidential campaign, individuals were randomly assigned to feeds with heightened, reduced, or unchanged levels of AAPA content. Piccardi et al. discovered that, relative to the control group, reducing exposure to AAPA content made people feel warmer toward the opposing political party, shifting the baseline by more than 2 points on a 100-point scale. Increasing exposure resulted in a comparable shift toward colder feelings toward the opposing party. According to the authors, the observed effects are substantial, roughly comparable to three years’ worth of change in affective polarization over the duration of the intervention, though it remains unknown if these effects persist over time. What’s more, these shifts did not appear to fall disproportionately on any particular group of users. These shifts also extended to emotional experience; participants reported changes in anger and sadness through brief in-feed surveys, demonstrating that algorithmically mediated exposure to political hostility can shape both affective polarization and moment-to-moment emotional responses during platform use.
“One study – or set of studies – will never be the final word on how social media affects political attitudes. What is true of Facebook might not be true of TikTok, and what was true of Twitter 4 years ago might not be relevant to X today,” write Jennifer Allen and Joshua Tucker in a related Perspective. “The way forward is to embrace creative research and to build methodologies that adapt to the current moment. Piccardi et al. present a viable tool for doing that.”
