X Algorithm May Amplify Content Users Hate, Study Suggests

Arguing with users on X may be inadvertently training the platform's algorithm to show more content that users dislike, according to a new study. Researchers from MIT, the University of Michigan, Johns Hopkins, Stanford, and Yale analyzed the behavior of 715 active U.S. X users in September and October 2024.

The study, published in the Proceedings of the National Academy of Sciences, found that X's 'For You' algorithm amplified content that was less aligned with users' explicitly stated values. This effect was particularly noticeable among Democrats, with a correlation of -0.229 between amplified content and stated values, compared to -0.041 for Republicans. Across all users, amplified content showed a misalignment score of -0.150.

Researchers suggested that user engagement, such as replying to posts that clash with their values, could be misinterpreted by the algorithm as a signal to recommend more of the same. The study did not investigate whether the accounts generating these posts were automated or human.

This finding follows X's decision in January to open-source its Grok-powered 'For You' algorithm, which ranks posts based on predicted user actions like likes and replies. Elon Musk acknowledged at the time that the algorithm required significant improvements.