Overview
- A peer-reviewed study published in PNAS on Tuesday, August 18, 2026, analyzed feeds from 715 active U.S. X users who installed a browser extension to collect their For You and Following pages.
- The team measured users’ self-stated values with a 19-point Schwartz inventory and found the algorithm amplified content that was overall negatively correlated with those values, a pattern the authors call value misalignment.
- The misalignment was much larger for self-identified Democrats than for Republicans, with reported correlations of about −0.229 for Democrats and −0.041 for Republicans, showing a clear partisan gap in exposure to value-clashing posts.
- Researchers identified replying as a low-frequency interaction (about 6.8% of engagements) that the recommendation model weights heavily, which can create outrage-driven feedback loops when users reply to content they dislike.
- The paper is observational and does not settle why the partisan gap exists or what role bots play, and the findings sit against X’s recent January 2026 move to open-source its Grok-powered recommendation system and public notes from company leadership that the algorithm needs major fixes.