The reply is the ranking signal, and the reply is where disagreement lives

The load-bearing work in this paper is the two checks the authors ran before the headline result. They asked whether users simply follow accounts unlike themselves, and found that most accounts people follow do align with their values ^[5]. They asked whether users only ever engage with material they dislike, and found people engage with plenty of posts that match their values ^[6]. So the mismatch in the feed is not inherited from the follow graph, and it is not a plain readout of what users touch. It is manufactured further down, where the ranker decides which of those touches to believe.

The answer, according to the researchers, is the reply. Likes are how people mostly respond; replies are rarer, and when someone does write one it often goes to a post that clashes with their values ^[7]. X's algorithm treats those rare replies as a much weightier signal than the many likes, and learns most strongly from them ^[8]. The result is a For You page built out of the values of posts the user argued with ^[9]. Put the weighting next to the behaviour and the comment box is functioning as a disagreement detector wired directly to the promote control ^[1].

That is the part with operational teeth for anyone running a ranking or growth surface. Comment rate is treated almost everywhere as the deeper, higher-intent cousin of the like. On the evidence here, the depth it measures is objection density. A team that raises comments per impression and books it as affinity has bought the other thing.

There is also a content-side tilt worth reading literally. The feed most amplifies posts about upholding tradition, following rules, and keeping society safe, and most demotes posts about looking after people, concern for people far away, being dependable, and protecting nature ^[3]. If your output sits in the demoted set, the ranker is not neutral about you, and no amount of comment volume moves you into the amplified set with your values intact.

The asymmetry is the measurement hazard. The authors report the effect is stronger for self-reported Democrats than Republicans, because Democrats object more to the content they reply to ^[10], and that the amplified content is more than four times more misaligned for Democrats ^[11]. Any single platform-level misalignment figure is averaging two populations with very different exposure, and the cohort that argues most is the one supplying the reply volume the ranker learns from.

Two caveats belong on the record. The description of how X weights replies against likes comes from the researchers' own account of their PNAS study rather than from published ranking code ^[15], and the sample is 715 US-based users ^[2], which is thin ground for slicing further than the two political groups they report. The alternative they offer is unglamorous: ask users what they value and sort the feed by the answers ^[12], on the evidence that users can reliably tell an aligned sample feed from a misaligned one ^[13]. That costs a survey, a sort key, and the reply volume you were counting as success.