A groundbreaking experiment has demonstrated that social media platforms can precisely control the rate at which political polarization develops among their users. Research on X showed that minimal adjustments to content feeds produced as much political division in one week as would have taken three years to emerge naturally, while also proving that similar adjustments can significantly reduce political animosity between partisan groups.
The study was published in the journal Science by researchers from Stanford, Johns Hopkins, Northeastern, and the University of Washington. They conducted a carefully controlled experiment involving more than 1,000 users during the 2024 presidential election campaign. Using artificial intelligence to classify posts based on divisiveness, they manipulated what appeared in participants’ “for you” feeds. Some users received slightly more posts containing antidemocratic sentiments, partisan hostility, opposition to bipartisan cooperation, and biased political information, while others saw fewer such posts, all while keeping modifications imperceptible to users.
The election campaign provided a charged context for the research, marked by viral spread of manipulated images and AI-generated propaganda on X. Since the platform’s acquisition and transformation, its algorithmic “for you” feed has prioritized content calculated to maximize engagement rather than simply displaying posts from followed accounts. This approach has generated ongoing debate about the platform’s impact on political discourse and democratic culture.
Researchers measured polarization effects using a “feeling thermometer” ranging from 0 to 100 degrees. After one week, participants exposed to more divisive content showed increased negative feelings toward political opponents of more than two degrees on this scale—matching the polarization increase that occurred across four decades of American history from 1978 to 2020. Assistant professor Martin Saveski noted that barely perceptible changes to feeds resulted in significant differences in how users felt about others. Co-author Tiziano Piccardi emphasized that the shift corresponds to approximately three years of polarization based on established trends.
The research offers both warnings and opportunities. While it confirms platforms’ power to rapidly intensify political division, it also proves they could choose to reduce polarization through algorithmic redesign. Users who saw fewer posts with antidemocratic attitudes and partisan animosity exhibited decreased political division by a similar amount. Although platforms have been accused of amplifying divisive content to maximize engagement and advertising revenue, the study found that down-ranking such content produced only slight reductions in overall engagement volume, while users actually engaged more meaningfully through likes and reposts. This suggests platforms could integrate methods to mitigate harmful societal consequences while maintaining viable business operations, though success would require prioritizing social well-being alongside profit maximization.