Arguing with a post on X, formerly known as Twitter, may be doing more than keeping the debate alive. A new study suggests the X algorithm can learn from those replies and keep serving users content that clashes with what they actually believe.
According to a newly published Proceedings of the National Academy of Sciences study, researchers examined whether the personalized For You feed on X reflects the values users say they hold. The researchers found an overall mismatch between participants’ stated values and some of the content the algorithm was more likely to amplify.
The study followed 715 active American X users during September and October 2024. According to the researchers, the group was quota-matched by ethnicity, gender, and political affiliation. Participants installed a browser extension that allowed researchers to collect posts appearing in both their personalized For You feeds and the feeds made up of accounts they intentionally followed. The study also used the Schwartz Theory of Basic Values to examine priorities including tolerance, dominance, hedonism, and openness to change.
What researchers found becomes especially interesting once replies enter the picture.
“When we look at commenting, the act of replying to posts, that’s where we actually see some kind of meaningful misalignment between people’s values and the values of the content they’re replying to,” study coauthor Ziv Epstein told 404 Media.
Replies represented only 6.8 percent of the engagement researchers observed, according to the PNAS study, but researchers found evidence that this relatively uncommon interaction carried considerable weight in what users were shown afterward. In other words, liking something can tell a platform you enjoy it, but arguing underneath something you hate can still tell the system that the post successfully captured your attention.
“Replying is only a fraction of engagement, but there does seem to be some evidence that these algorithms are prioritizing and learning more from this kind of rarer form of engagement,” Epstein told 404 Media. “So it’s this feedback loop of outrage baiting. The algorithm learns that you get outraged and then continues to serve more content in that direction, and that seems to be particularly true of the Democratic users of our study.”
That partisan difference became one of the study’s biggest findings. According to the researchers, users who identified as Democrats encountered more content that conflicted with their values and were more likely to respond to that content through replies. The researchers did not establish exactly why that difference occurred, and Epstein cautioned against assuming it reflects either the psychology of Democrats or the overall ideological makeup of X without additional research.
“Democrat users confront the abundant value-misaligned content by replying to it, which the algorithm in turn preferentially learns from and continues to feed them,” the study said.
The X algorithm finding fits into a much larger body of research showing that conflict can be unusually valuable currency online. A 2021 study published in Science Advances found that social rewards such as likes and shares can reinforce moral outrage expression, making people more likely to express outrage again later. Researchers concluded that both reinforcement learning and perceptions of social norms can shape how outrage spreads across social networks.
Another 2021 PNAS study examining political posts on Facebook and Twitter found that language referring to the opposing political group was a particularly strong predictor of engagement. Researchers reported that each additional term referring to a political out-group increased the odds that a post would be shared by 67 percent.
The pattern extends beyond political arguments. A 2024 study published in Science examined Facebook and Twitter data alongside behavioral experiments and found that misinformation sources generated more outrage than trustworthy sources. The researchers also found that outrage increased sharing and that people were more willing to share misinformation that provoked outrage without reading the linked material first.
Separate research has also raised questions about what X chooses to place in front of users. A 2026 Nature study based on a randomized seven-week experiment conducted in 2023 found that X’s algorithmic feed promoted more conservative political content than the chronological feed for both Democrats and Republicans or independents. Researchers reported that algorithmic exposure shifted some policy and current events opinions in a more conservative direction, while finding no significant effect on participants’ political party identification or affective polarization.
There is also an important timing issue. The new PNAS research reflects how X behaved during its 2024 study period, not necessarily how the platform operates today. Former X product head Nikita Bier said the company later changed how its system treats replies.
“This is no longer true,” Bier said. “The largest contributor of seeing ragebait was the reply predictor, and we were aware that angry replies were causing people to see more of that content. So last month, we gave the reply predictor a 15x boost if it’s a friend’s post — and it reduced ragebait by [an] order of magnitude.”
That response does not erase what researchers observed in 2024, but it underlines one of the hardest parts of studying social media algorithms. Platforms can change ranking systems quickly, while independent researchers are often studying behavior that occurred months or years earlier. The new findings nevertheless add another piece to a growing research record suggesting that online outrage is not simply something users bring to social media. The way platforms interpret attention can help determine how much of that outrage comes back around.
