X's ranker pays more for replies than likes, and that makes reach a bad proxy for quality

A study published in PNAS that instrumented the feeds of 715 American X users found the platform's ranking was more likely to amplify content that did not align with the values those users said they held ^{[1] [2] [5]}. For anyone who treats organic reach on X as a rough proxy for whether a post was good, that is a modeling error rather than a moral complaint.

The method was direct. Participants installed a browser extension that collected both their For You and Following feeds ^[2]. They filled out a values inventory built on the Schwartz Theory of Basic Values, which breaks a belief system into 19 points covering qualities such as "tolerance" and "dominance" ^[3], and they reported their political alignment ^[4]. Researchers then tracked how users engaged with posts and how far those posts reflected the stated values ^[5].

The mechanism the authors point to sits in how engagement types are priced. Co-author Ziv Epstein, a Stanford researcher, told 404 Media that X's feed algorithm is optimized for engagement but that "not all types of engagement are considered equally" ^[6]. Replying is weighted much more heavily than liking, even though replies account for under seven percent of interactions on the platform ^[7], which is fewer than one interaction in fourteen ^[8]. Epstein described the result as a feedback loop of outrage baiting: the algorithm learns that you get outraged, then serves more content in that direction ^[9].

Read that as a pricing sheet rather than a scandal. If the scarce signal carries the heavy coefficient, the cheapest way to win distribution is to provoke a reply, which is not the same thing as being worth replying to. Reach measured under those weights is a measure of provocation efficiency, and a growth forecast that assumes otherwise is fitting the wrong variable.

The study also found the pattern was stronger for users who identified as Democrats ^[10]. The researchers say they do not yet have a reason, and offer two possibilities: there is simply more right-wing content on X, or Democrats engage more with posts they disagree with ^[11]. That is an open question in the paper, not a finding.

X has not formally responded to the research ^[12]. Nikita Bier, the company's former head of product, said in a post dated 18 August 2026 that the reply predictor was the largest contributor to seeing ragebait and that the team was aware angry replies were causing people to see more of it, and that the previous month they gave the reply predictor a 15x boost when the post came from a friend ^{[13] [14]}. Bier says ragebait has since been reduced by an order of magnitude ^[15]. Engadget's writer says their own feed suggests otherwise ^[16].

Worth noting what that fix is: another re-weighting of the same reply predictor, gated on the friend graph ^[14], not a move to a quality measure. The coefficient changed; the thing being counted did not.

Watch whether anyone outside X can measure the order-of-magnitude claim, since the 715-user extension method is repeatable by academics but the weights are not disclosed ^{[2] [12]}. Watch whether the friend-graph gate survives, because a multiplier that large is a tempting knob to turn back ^[14]. And if you are planning distribution on X, price in that the ranker's objective is revisable without notice and that your reach series may contain a step change you did not cause.