X’s Algorithm Feeds Off Ragebait and Impacts Democrats More, Study Finds

X’s algorithm actively learns what provokes its users and then feeds them more of that content, with a particular skew toward users who identify as Democrats, according to a study published in the Proceedings of the National Academy of Sciences. The paper, titled “Value misalignment of X’s feed algorithm is a reflection of value tensions in engagement,” confirms what many users have long suspected: the platform’s For You Page is optimized for engagement above all else, and not all forms of engagement are weighted equally. The researchers found that replies, which accounted for only 6.8% of all interactions, carried disproportionate influence over what content the algorithm surfaced, effectively creating a feedback loop of outrage.

The study’s lead author, Ziv Epstein, a postdoctoral researcher at Stanford University, told 404 Media that the algorithm learns that a user gets outraged and then continues to serve more content in that direction. This dynamic is not surprising given the platform’s business model, but the research exposed a more subtle and concerning asymmetry: X served more rage-inducing content to users who said they were Democrats than to those who said they were Republicans. The study did not determine the cause of this disparity with certainty, but two plausible explanations emerge from the data. One is that the platform’s overall content pool tilts rightward, meaning Democratic users are more likely to encounter material that contradicts their views and provokes a reply. The other is behavioral: Democratic users may be more inclined to engage with content they disagree with, thereby training the algorithm to serve them more of it.

X’s former product chief, Nikita Bier, acknowledged the mechanism on the platform itself. He confirmed that the site had been set to favor replies and that the reply predictor was the largest contributor to users seeing ragebait. Bier stated that the company was aware angry replies caused people to see more of that content and had implemented a change the previous month: giving the reply predictor a 15-times boost when the reply was on a friend’s post, a move he said reduced ragebait by an order of magnitude. Whether that adjustment addresses the deeper structural incentives of the platform remains an open question. The algorithm still prioritizes engagement, and outrage remains one of the most reliable forms of it.

The implications extend beyond user experience. For professional readers monitoring platform governance and information integrity, the study offers concrete evidence of a value misalignment. X’s stated values around free expression and user satisfaction are at odds with a system optimized to maximize time on site through emotional manipulation. The partisan asymmetry is especially troubling for political discourse. If one side of the political spectrum is systematically fed more inflammatory content, the platform is not merely a neutral conduit but an active participant in warping that group’s information diet. Markets and advertisers should take note: an environment optimized for rage is not a stable foundation for sustained user trust or advertising effectiveness.

The core takeaway is straightforward but often ignored in the noise of platform politics. X’s algorithm is not broken; it is working exactly as designed. The problem is that the design incentives produce a product that is toxic for users and corrosive for democratic debate. The fix, as Bier’s admission suggests, requires deliberately suppressing the most reliable drivers of engagement rather than optimizing for them. Until platform incentives are realigned with user values, the ragebait will keep flowing, and the asymmetry in who sees it will persist.