What the design can carry
The researchers recruited 715 American X users under quotas for ethnicity, gender and partisanship, and had each of them install a browser extension that captured both the For You and Following feeds; collection ran through September and October 2024. Participants filled in the Schwartz value inventory across 19 dimensions, and the posts that landed in their feeds were annotated on those same dimensions. The measurement at the centre of the study, published in PNAS, follows from that pairing: how much of what a person was shown matched what that person said they valued?[1]
The finding is that posts out of step with a participant's stated values turned up in the For You feed more often than aligned posts, and drew more engagement. That is a relationship between two annotated quantities, and it carries only as much as the annotation captures. A value inventory records what someone endorses in the abstract; a post annotation records what a coder judged the post to express. Two people can score alike on the tolerance dimension and disagree entirely about which posts display it, so part of what reads as misalignment may be a disagreement about the label rather than about the content.[1]
The weight replies carry
Replying accounted for 6.8 per cent of the observed interactions, yet its influence on ranking ran well past that share. Likes and reposts are common and replies are rare, and still much of the signal that decides what appears in the For You feed comes from the rare one. The study also reports that users who identify as Democrats received more of this content.[1]
Read as a mechanism, this describes a loop: content provokes a reply, the reply teaches the ranker, the ranker serves more of the same. An observation cannot separate the loop's directions. Everything was measured on feeds as they arrived, and no group was assigned a different feed, so a ranker that already promotes provocative posts for reasons unrelated to values would produce the same pattern: the author's reach, how fresh the post is, how contested the topic is. The partisan gap has no established cause either, and co-author Ziv Epstein says answering it needs more research.[1]
The evidence that would settle it
The design that answers the causal question is an intervention: give one group a feed with the reply signal removed or down-weighted, hold the rest fixed, and see whether misaligned content still rises in the For You feed. Short of that, an independent replication on the current ranker would show whether the 2024 pattern still holds; X has changed its ranking since, and a measurement taken in the autumn of 2024 is not evidence about today's feed. If such a replication or a platform-side experiment is published by April 30, 2027, the loop moves from a plausible account to a tested one; if the new work is another observational snapshot, the claim stays where it is.[1]