Social media algorithms are designed to hold attention: they show what triggers a reaction. But, as scientists have found, in the X feed (formerly Twitter) automatic sorting often promotes posts that diverge from the user's own beliefs. The work was published in the Proceedings of the National Academy of Sciences.
The authors are Ziv Epstein of the Massachusetts Institute of Technology and Michael Bernstein of Stanford University, with colleagues. They relied on Shalom Schwartz's theory of basic human values. This model describes universal motivations, including societal security, tradition, openness to change and care for others. The researchers matched these values against the texts of posts.
The study involved 715 American adults who actively use X. They installed a browser extension that recorded two feeds: "Following" (subscriptions without sorting) and "For You" (the algorithm's personalized selection). A special artificial intelligence model determined which values each post expressed.
It turned out that in the "Following" feed, posts generally corresponded to what participants reported about their values in surveys. However, in "For You" the picture changed: a negative correlation was found - a statistical association that in itself does not prove causation - between declared values and those the algorithm conveyed. In other words, the system more actively showed material contradicting the person's views.
At the most general level, the algorithm slightly favored individualistic values over collectivist ones. It consistently amplified posts about order and preserving the familiar way of life - for example, about obeying rules, public safety and traditions. At the same time it suppressed material about care for others, reliability and protecting nature.
The imbalance affected users with different political views, but it was more pronounced among Democrats.
To understand the mechanism, the scientists separately studied how people interact with posts over 24 hours. Likes and reposts usually matched personal values - this is conscious support. Replies, however, more often diverged from beliefs, which looks like confrontation. Among Democrats, such replies on average slightly failed to match their values.
Because the X algorithm gives great weight to replies, the researchers hypothesized a feedback loop: a person enters into an argument with an unpleasant post, the system sees high engagement and shows even more irritating content. As a result, the feed fills with what provokes a reaction rather than what is close to the person.
In addition, the overall analysis showed that globally popular content more often matched Republicans' values. This suggests that right-leaning users do not need to get into conflictual discussions to receive material close to them.
The study is based on observations: the scientists did not interfere with the algorithm's operation. Therefore it cannot definitively prove that confrontational replies are the only cause of the mismatch. The authors note the need for controlled experiments and repeated checks on larger samples.
Nevertheless, the results raise a question: what patterns of behavior should social media algorithms encourage? If the system learns primarily from arguments and friction, it will continue to create an environment that contradicts users' own preferences.