A examine revealed within the Proceedings of the Nationwide Academy of Scientists (PNAS) has discovered what all of us suspected, that X’s algorithm prioritizes ragebait with a purpose to juice engagement. The examine additionally discovered that this content material was served up extra to individuals who establish as Democrats.
This isn’t a shock to anybody who has checked out X within the final couple of years. Presently, my feed is an countless cavalcade of right-leaning troll accounts popping off about how evil it’s {that a} store on the Mall of America sells hijabs (or one thing.)
One of many paper’s authors, Ziv Epstein, informed 404 Media that they know that is “not essentially the most shocking headline ever” however that they needed to get to the why of all of it. “X’s feed algorithm, like loads of these social media algorithms, is optimized for engagement [but] it seems that not all kinds of engagement are thought-about equally,” the Stanford researcher mentioned.
The examine concerned following 715 American X customers who put in a browser extension that collected their For You and Following feeds. Individuals have been requested to fill out a “values stock” utilizing one thing known as the Schwartz Principle of Fundamental Values, which breaks down somebody’s perception system into 19 factors that correspond to qualities like “tolerance” and “dominance.” Volunteers additionally reported their political alignments.
Subsequent, researchers tracked how these customers engaged with posts and the way these posts mirrored their self-reported values. The examine discovered that the algorithm was “extra prone to amplify” content material that did not align with these self-reported values. In different phrases, it did not take lengthy till they began seeing stuff that angered them.
If a publish infuriates somebody, they’re extra prone to have interaction with it which, in flip, sends extra outlandish posts their means. Epstein notes that replying (to a baseless or offensive publish, as an example) is weighted rather more closely than liking, regardless of replies accounting for underneath seven % of interactions on the platform.
“Replying is barely a fraction of engagement, however there does appear to be some proof that these algorithms are prioritizing and studying extra from this sort of rarer type of engagement,” he mentioned. “So it is this suggestions loop of concern baiting. The algorithm learns that you simply get outraged after which continues to serve extra content material in that course.”
Nevertheless, the examine additionally discovered that this ragebaiting occurred extra for the customers who recognized themselves as Democrats. The researchers do not but have a motive for this, suggesting that it could possibly be accounted for by the truth that there’s merely extra proper wing content material on X or that Democrats have a tendency to interact extra with posts they disagree with.
Epstein mentioned the examine was designed to get individuals speaking about social media apps and the way they function. “There are these social media algorithms which have huge quantities of energy in our lives, they form the data that we devour and now we have little or no transparency into how they function and what their implications are,” he mentioned. “I feel that has sort of necessary implications for civil society and simply preventing among the techno-feudalistic tendencies of platforms to manage these algorithms.”
That is not true:
The most important contributor of seeing ragebait was the reply predictor and we have been conscious that offended replies have been inflicting individuals to see extra of that content material.
So final month, we gave the reply predictor a 15x increase if it is a good friend’s publish—and it decreased… https://t.co/dl8qVld4K2
— Nikita Bier (@nikitabier) August 18, 2026
X hasn’t formally responded to this analysis. Nevertheless, the corporate’s former head of product Nikita Bier says the algorithm has been adjusted because the examine came about and that ragebait content material has been decreased “by order [sic] of magnitude.” My feed suggests in any other case, however not less than the platform is conscious of the problem and dealing on it.
