The side you're on

Upworthy, a website that pioneered click-bait while branding itself a "positive news outlet", tested its headlines the way a lab tests a drug: the same story went out under different headlines to randomly chosen readers, and the clicks were counted.¹ Claire Robertson and colleagues analysed 22,743 of these tests, covering about 105,000 headline variants and more than 370 million impressions, and found that each additional negative word in a headline of average length raised the click-through rate by 2.3%, while positive words lowered it.¹

That is the old complaint about media in miniature: bad news sells. The question underneath it is whether conflict dominates because of how attention works or because of how the business works. The evidence gives a sharper answer than either. What holds attention is not conflict as such but a threat from the other side. The mind supplies a small pull toward it; the business counts the pull, and counting makes it larger.

The pull is in people, not only in one market. Stuart Soroka, Patrick Fournier and Lilach Nir put sensors on the fingers of 1,156 people in 17 countries on six continents and showed them BBC World News stories.² On average, people were more physiologically activated by negative stories than by positive ones, but the variation between individuals was large, and the authors told news producers not to "underestimate the audience for positive news content."² The bias is real everywhere and it is an average, not a law.

Negativity is also the weaker part. In 2.7 million posts by US news outlets and members of Congress on Facebook and Twitter, each word referring to the political out-group raised the odds of a share by about 67%, words about the in-group had no significant effect, and out-group language predicted shares about 4.8 times as strongly as negative language.³ That study was correlational.³ A 2025 working paper by the same group went back to Upworthy's randomised tests: for a readership that leaned liberal, negative headlines about conservatives raised clicks 14.1%, and negative headlines about liberals 1.6%.⁴

A fight holds attention when the reader is in it. The other side's misdeed is a threat to me and mine; my own side's misdeed is merely bad news.

None of this needs a cynical editor. A newsroom that tests every headline and keeps the winner is running thousands of small experiments, and an edge of a few percent, applied at every choice, is enough to tilt a front page. A feed ranked by engagement makes the same choice for every post, for every reader, all day.

In February 2023, Smitha Milli and colleagues collected two feeds for each of 806 American Twitter users: the one the engagement-based algorithm would have shown, and the most recent posts from the accounts each user followed.⁵ Compared with the chronological feed, the ranked one carried more emotionally charged content and more hostility toward the user's political out-group (by about a quarter of a standard deviation), users said it left them feeling worse about their opponents, and they did not prefer the political posts the algorithm had picked.⁵ In an exploratory test, when the researchers instead ranked posts by what users said they wanted, angry, partisan and hostile content fell, though posts users already agreed with may have risen.⁵

That is where the question splits. Attention works the way the experiments show, and no business created it. But a business paid by attention counts one kind of vote, the one cast by the part of the mind that notices threats, and the feed is built from that count. Each click is honest. It is just not the whole person.

So a medium can fill with what its own audience would vote against, without anyone being deceived, because nobody asked the other part. The conflict on the front page is partly the world and partly an election in which only the reflexes were allowed to vote.