If your For You feed on X feels like a machine engineered to irritate you, new research suggests that is not paranoia. A study published in the Proceedings of the National Academy of Sciences collected what X genuinely served to hundreds of real American users, compared it against what those same people said they valued, and found a consistent gap between the two.
The headline finding is not that the algorithm amplifies outrage. Most of us worked that out years ago. The interesting part is the mechanism. According to the researchers, the strongest single driver of a worse timeline is the one action you probably think of as fighting back: hitting reply.
The short version
- The paper: “Value misalignments in X’s feed algorithm is a reflection of value tensions in engagement,” published in PNAS
- Lead author: Ziv Epstein, a postdoctoral researcher at Stanford
- Sample: 715 Americans, quota matched on ethnicity, gender and partisanship
- Method: a browser extension that captured each participant’s real For You page
- Key number: replies were just 6.8 percent of all interactions but carried outsized weight with the algorithm
- Uneven effect: users who identified as Democrats were served more value misaligned content than Republicans
What the researchers actually measured
Most studies of social media feeds run into the same wall. Platforms do not hand over their ranking systems, and asking people to describe their own timeline from memory produces mush. The Stanford team went around the problem by installing a browser extension on participants’ machines that quietly downloaded the posts X chose to show them during September and October 2024.
That gave the researchers two datasets sitting side by side. On one side, the content X selected. On the other, a values profile for every participant, built using the Schwartz Theory of Basic Values, a framework psychologists have used for decades. It maps human motivations onto 19 points arranged in a wheel, covering things like tolerance, humility, dominance, hedonism, security and openness to change.
Then they asked a simple question. Does the stuff the algorithm pushes hardest line up with what these people say matters to them?
It does not. The paper reports “an overall negative correlation (misalignment) between users’ explicit values and the values in content that the algorithm is more likely to amplify.” In plain terms, the more strongly X promoted a post, the less likely it was to reflect the viewer’s stated values.
| Study element | Detail |
|---|---|
| Participants | 715 US adults, quota matched on ethnicity, gender and partisanship |
| Data collected | Real For You page content, captured by a browser extension |
| Collection window | September and October 2024 |
| Values framework | Schwartz Theory of Basic Values, 19 value dimensions |
| Headline result | Negative correlation between user values and amplified content values |
| Strongest signal | Replies, despite being only 6.8 percent of interactions |
The 6.8 percent problem
Here is the part worth sitting with. Almost everything people do on X is agreeable behavior. You like a post because you enjoyed it. You repost something because you want others to see it. Those actions broadly track what you actually care about.
Replying is different. People reply to things they like, sure, but they also reply to argue, to correct, to dunk, to tell a stranger they are wrong on the internet. It is the one interaction type that fires just as readily on content you hate as on content you love.
And it is rare. Across the study, replies accounted for a small slice of all recorded interactions.
Share of all recorded interactions
Source: Epstein et al., PNAS, 2026. Replies were the rarest interaction type in the study yet carried disproportionate weight in what the feed served next.
A ranking system optimized for engagement sees a rare signal and treats it as valuable. Rarity implies information. If almost nobody replies and you did, the model reasonably concludes that this post moved you in a way a scroll past did not. It cannot tell the difference between moved and enraged.
Epstein put it directly: “It turns out that X’s feed algorithm, like a lot of these social media algorithms, is optimized for engagement, but it turns out that not all types of engagement are considered equally.”
How the loop closes on itself
The result is a feedback cycle that is genuinely difficult to escape once it starts, because every attempt to push back on the content feeds the thing producing it.
As Epstein described it, “the algorithm learns that you get outraged and then continues to serve more content in that direction.” Nobody at the company has to want this. It falls out of optimizing for a number.
Why Democrats got hit harder
The finding that will generate the most heat is that self identified Democrats in the sample were served more value misaligned content than Republicans. The researchers are careful here, and so should everyone else be.
The paper’s explanation is behavioral rather than political. Democratic users in the study confronted content they disagreed with by replying to it, and the algorithm learned preferentially from exactly that behavior. If one group argues more and another group scrolls past, the group that argues gets more of what it argues with. That is a mechanical outcome, not evidence of a thumb on the scale.
It also leaves the obvious question unanswered. Why does one group reply more in the first place? A likely factor is the overall composition of content on the platform since 2022, which would put left leaning users in contact with more disagreeable posts to begin with. The study does not settle that, and the authors say more research is needed. Anyone telling you the paper proves deliberate bias in either direction is reading something into it that is not there.
X says it already fixed part of this
X did not respond to press requests for comment, but Nikita Bier, the platform’s former head of product, pushed back on the idea that the findings still describe the live system. The study captured data from late 2024, and ranking systems get rewritten constantly.
What the former head of product 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.”
Take that as a partial admission rather than a rebuttal. The fix he describes does not remove the reply signal. It reweights it toward people you actually follow, which should mean fewer strangers’ hot takes and more of your own circle. Whether that holds up at scale is not something anyone outside the company can verify, which is its own problem. X open sourced parts of its recommendation code, but the version that runs in production and the version on GitHub are not the same thing.
What you can do about your own feed
None of this requires deleting your account. If the reply signal is the lever, then the practical countermeasures follow directly from that.
| Move | Why it works | Effort |
|---|---|---|
| Stop replying to bad posts | Removes the highest weighted signal you can send | Hard, honestly |
| Quote post instead of replying | Still engagement, but a weaker direct interest signal on that account | Low |
| Live in the Following tab | Chronological feed, no ranking model deciding for you | One tap |
| Build topic Lists | Curated timelines that bypass For You entirely | Medium, once |
| Mute keywords, not just accounts | Cuts off the topic before the reply reflex kicks in | Low |
| Use “Not interested in this post” | Sends an explicit negative signal instead of an ambiguous one | Low |
The awkward truth in that list is that the most effective option is also the least satisfying. Silence reads as disinterest to a ranking model. Anger reads as interest. There is no button that means “show me less of this” as loudly as a reply means “show me more.”
This is bigger than one platform
It would be convenient to file this under things Elon Musk broke, but the mechanism described in the paper is not unique to X. Any system that ranks content by predicted engagement, and that treats rare interactions as strong signals, will drift the same direction. That is the same underlying argument at the center of the case where a jury found Meta and Google negligent over social media addiction, and it is why regulators keep circling feed design rather than individual posts.
Europe has been the most aggressive on this front, and X has already drawn scrutiny there for how its systems behave, including the EU investigation into Grok generated content. Meanwhile the definitional fights continue elsewhere in the industry, from YouTube quietly changing what counts as a view to platforms rewriting what engagement even means when the numbers stop looking good.
What to watch next
- Replication on current data. The study window closed in October 2024. Someone needs to run the same extension in 2026 to test whether Bier’s fix actually held.
- Whether other platforms get the same treatment. The method works anywhere a browser extension can capture a feed, which means TikTok, Instagram and Threads are all fair game.
- Regulatory interest in signal weighting. If replies are the lever, disclosure rules about how platforms weight interaction types become far more meaningful than generic transparency reports.
- Whether X publishes updated ranking code. Open sourcing the model was the promise. Keeping it current is the test.
For now the takeaway is uncomfortably simple. The timeline is not reading your mind. It is reading your fingers, and your fingers move fastest when you are annoyed.
