The suspicion is close to universal now, and it does not feel like a theory. It feels like something you can verify from the couch. Spend an hour in a feed, notice what it hands you, notice how you feel afterward, and conclude that the thing was built to set people against each other.
Part of that is better evidenced than most people realize. And part of it, the part almost everyone adds without noticing they have added it, has now been tested properly twice and has not come out the way the theory predicts.
What actually spreads
Steve Rathje, Jay Van Bavel and Sander van der Linden collected 2,730,215 posts from news media accounts and members of the US Congress on Facebook and Twitter, and published the analysis in the Proceedings of the National Academy of Sciences in 2021.
The Artful Age
A weekly letter on aging well, family across generations, and the creative life after the kids leave home.
Posts about the political out-group were shared or retweeted about twice as often as posts about the in-group. Each individual term referring to the out-group raised the odds of a post being shared by 67 percent. Out-group language was about 4.8 times as strong a predictor of sharing as negative emotional language, and about 6.7 times as strong as moral-emotional language, both of which were already known to travel well.
The pattern of reactions matched. Out-group language strongly predicted angry reactions. In-group language predicted love. None of it depended on which side the account was on, or which platform it was posted to, though it ran stronger for politicians than for news outlets.
If you wanted a portrait of a system that pays out for hostility, that is it.
The sentence that does not get quoted
The authors are careful about what they have, which is a very large correlation. They say plainly that the data cannot establish causation and that experiments are needed.
They say something else that matters more here, and it is almost never repeated when this study is cited. They cannot tell how much of the amplification comes from the algorithms, because the companies do not disclose how the algorithms work.
So the finding is that hostile posts travel. Who or what is carrying them, the ranking system or the people doing the sharing, the study is not in a position to say.
Switching it off
Somebody did run the experiment. Around the 2020 US election, a large team led by Andrew Guess moved consenting Facebook and Instagram users off the default algorithm and onto reverse-chronological feeds for three months, and reported the results in Science.
A great deal changed. Time on the platforms dropped substantially, and so did activity. Political content went up, and so did untrustworthy content. On Facebook, people saw more from moderate friends and from sources with ideologically mixed audiences. And the amount of content classified as uncivil or containing slur words went down, which is precisely what the critics would predict, and which means the algorithm had been supplying more of it.
Then the outcomes. Issue polarization, affective polarization, political knowledge and other key attitudes were measured, and across the three-month study period none of them significantly changed.
And then again, somewhere else
This year, Germain Gauthier, Roland Hodler, Philine Widmer and Ekaterina Zhuravskaya published a randomized experiment in Nature that did the same thing on X, over seven weeks in 2023, with 4,965 people completing both surveys.
The algorithmic feed was dramatically more engaging, and it changed the diet. Posts from traditional news organizations appeared 58 percent less often. Posts from political activists appeared 27 percent more often. Content annotated as conservative was about 20 percent more likely to show up.
And it moved opinions. Policy priorities shifted roughly 0.11 standard deviations in a conservative direction, with movement of a similar size on the Trump investigations and on Ukraine. The effect was strangely one-way. Switching the algorithm on did this; switching it off did not reverse it, apparently because people had started following activist accounts and went on following them afterward.
On the measure this article is about, though, the result matched the Facebook and Instagram experiment: neither switching the algorithm on nor switching it off significantly affected affective polarization or self-reported partisanship.
Where the argument actually splits
The two bodies of work are not in conflict. They answer different questions, and the popular version of the argument runs them together. One asks what spreads. The other asks what happens to the person it spreads to.
The answer to the first is that conflict spreads, reliably and by a wide margin, and that at least one platform’s ranking was demonstrably feeding people more incivility than they would otherwise have seen. The answer to the second, so far, is that when you take the ranking away, the hostility people report toward the other side does not move.
What that does not license
A null result is not an all-clear, and it would be dishonest to read it as one.
Both experiments are short, run on people already shaped by years of the thing being switched off, which is a hard design problem rather than a flaw anyone could fix. Both measure individuals rather than the shape of public argument, and those are not the same object. The X findings are tied to that platform, that period and its owner’s preferences, which the authors state explicitly. The Meta experiment ran during an election, which is not an ordinary stretch of time. And the X study found the feed changing what people believed, which is a serious finding in its own right and not a consolation.
What survives is narrower than the slogan and more useful. These systems do reward conflict, measurably and at scale. Whether they are the reason you cannot stand the other half of the country is a question that has now been asked carefully twice, and the answer that keeps coming back is not yes.