A PNAS study says X's feed learns from your arguments, not your likes

A study of X published in the Proceedings of the National Academy of Sciences reports that the platform's feed algorithm is more likely to promote posts that conflict with a user's values than posts that align with them ^[1][5]. The mechanism the authors identify is not ideological: it is that replies are weighted far more heavily than likes, and replies are where disagreement lives ^[8][9].

The team, writing up its own work in The Conversation, built a measurement tool based on standard psychological classifications of human values and applied it to the feeds of 715 US-based users on X ^[2][3]. At the aggregate level, the feed leaned in a consistent direction: posts about upholding tradition, following rules and keeping society safe were the most likely to be amplified, while posts about looking after people, concern for people far away, being dependable and protecting nature were the most likely to be demoted ^[4]. Compared against the values users expressed in their own posts, promotion skewed toward conflict ^[5].

The authors say they checked two duller explanations first. The accounts people follow mostly do align with their values, so the follow graph is not the source of the mismatch ^[6]. And users engage with plenty of content that reflects values they share, so it is not simply a case of people asking for fights and getting them ^[7]. What differs is the form of the interaction: liking is the common response, replying is rarer, and replies disproportionately land on posts that clash with the user's values ^[8]. Because the system learns most strongly from replies, it serves more of what the user argued with ^[9]. The practical consequence is that the ranking signal driving the feed is drawn from the smaller and more adversarial slice of a user's behaviour ^[10].

That has an uneven distribution. The effect was stronger for users who reported being Democrats than for those who reported being Republicans, and the authors attribute this to Democrats objecting more than Republicans to the content they reply to, producing a tighter feedback loop ^[11][12]. The size they report is not marginal: the amplified content was more than four times more misaligned for Democrats than for Republicans ^[13].

For anyone running a ranking system, the useful reading is that "engagement" is not a single quantity. A comment costs more than a like and therefore looks like a stronger preference signal, but it can encode the opposite of preference, and the exchange rate between the two is a tuning decision rather than a fact of nature ^[8][9]. It also means an audit that reports one average alignment number across a user base can hide a fourfold difference between groups ^[13].

The authors propose an alternative in separate work: ask users directly what they value and sort feeds accordingly, a method they say works because users can reliably tell whether a sample feed matches their values ^[14][15]. They argue this may be a better route out of echo chambers than unmediated exposure to opposing views ^[16]. That claim is theirs, from the same group that produced the audit, and the article summarising the findings is written by the researchers rather than by an independent reviewer ^[2].