An innovative scientific study has revealed the mechanism by which social media platforms can rapidly generate political hostility among their users. Researchers demonstrated that imperceptible changes to the content displayed on X can create levels of political division in one week that would normally take three years to develop naturally in society.
The experiment represents a breakthrough in understanding how algorithmic content curation affects democratic discourse. More than 1,000 X users participated during the 2024 presidential election without realizing their feeds were being manipulated for research purposes. Scientists used artificial intelligence to classify posts based on their divisiveness, then adjusted participants’ feeds to show either more or fewer posts containing antidemocratic attitudes, partisan aggression, opposition to bipartisan cooperation, and biased presentations of political facts.
Assistant professor Martin Saveski from the University of Washington emphasized that the algorithm’s power lies in its subtlety. Despite barely perceptible changes to their feeds, users reported significant differences in their feelings toward political opponents. His colleague Tiziano Piccardi noted that the observed shift corresponds to roughly three years of polarization based on established trends in American political attitudes.
The measurement methodology involved asking participants to assess their feelings toward the opposing political party using a “feeling thermometer” ranging from warm and favorable to cold and unfavorable. Participants exposed to more divisive content showed increased negativity of more than two degrees on the 100-point scale—matching the polarization increase that occurred throughout the four decades leading to 2020. Conversely, participants who saw fewer divisive posts exhibited decreased polarization of a similar magnitude.
These findings emerge at a time when political division has reached alarming levels in many democracies. Polling indicates that overwhelming majorities believe political opponents cannot agree even on basic facts, let alone policies. The research demonstrates that platforms possess the technical capability to reduce this division through algorithmic design choices. However, the study also acknowledges a business model challenge: interventions that reduce divisive content may slightly decrease engagement volume, creating tension between profit maximization and social responsibility. The fact that users in less-divisive conditions showed higher rates of meaningful interaction suggests this trade-off may be less severe than platforms fear.