Rage Bait: The Effect Is Real, and Smaller Every Time It Is Measured

Rage bait is content built to make you angry so that you engage with it. The term is recent; the practice is not; and unlike most things in this field, it has been measured at enormous scale.

The measurements say the effect is real. They also say it is roughly half the size of the number that entered public conversation - and we know that because the original authors ran a preregistered replication and a meta-analysis, and published the smaller figures themselves.

That sequence is the most interesting thing here, and it is a better story than the one about the algorithm.

The finding, and its decay

Brady and colleagues (2017), in PNAS, analysed 563,312 tweets across three polarising issues and reported that each additional moral-emotional word increased diffusion by 20%.

Note the construction, because it is routinely misreported: per additional word, not per post. And the effect was bounded by group membership - stronger within liberal networks and within conservative networks than across them. Moral outrage travelled inside camps, not between them.

Then came the critique.

Burton, Cruz and Hahn (2021), in Nature Human Behaviour, reanalysed the claim using out-of-sample prediction, model comparison and specification-curve analysis. Their result: the moral contagion model performed no better than an implausible “XYZ contagion” model - a deliberately meaningless predictor. If nonsense fits as well as your theory, your theory is not doing the explaining.

Most fields would stop there, or fight. Instead the original group ran the test.

Brady, Rathje, Globig and Van Bavel (2025) published a preregistered direct replication with 849,266 observations plus a meta-analysis of published and unpublished studies.

Estimate	Effect per additional moral-emotional word
Original (2017)	+20%
Preregistered replication (2025)	+17% (IRR 1.17)
Meta-analysis: k = 27, N = 4,821,006 observations	+13% (IRR 1.13, 95% CI [1.06, 1.20])

The effect survives. The headline does not. Twenty per cent became thirteen, with a confidence interval that comfortably excludes zero.

If you want a single example of how a viral social-science number behaves under scrutiny, this is it - and it reflects well on everyone involved.

What actually drives sharing is not outrage but the other side

Rathje, Van Bavel and van der Linden (2021) analysed 2,730,215 posts from news accounts and US members of Congress across Facebook and Twitter.

Posts referring to the political out-group were shared roughly twice as often as posts about the in-group, with each out-group term raising sharing odds by about 67%.

The comparison that matters:

• Out-group language was 4.8 times stronger than negative-affect language as a predictor of sharing
• Out-group language was 6.7 times stronger than moral-emotional language

That second line demotes the mechanism from the 2017 study within the same field. Out-group references also predicted “angry” reactions specifically, while in-group language predicted “love.”

So the sharper claim is not that anger spreads. It is that content about people you dislike spreads, and anger is one of the things that comes with it.

The one genuinely experimental dataset

Almost everything above is observational. One study is not.

Robertson and colleagues (2023), in Nature Human Behaviour, used an archive of 22,743 randomised controlled trials - real A/B tests run by a publisher on live headlines - covering about 105,000 headline variations and 5.7 million clicks across more than 370 million impressions.

• Each additional negative word: +2.3% click-through rate (β = 0.015, SE = 0.001, z = 17.42, 99% CI [0.013, 0.018])
• Each additional positive word: -1.0% click-through rate (β = -0.008, SE = 0.001, z = -9.24, 99% CI [-0.010, -0.006])

That is causal evidence, at scale, that negativity in a headline earns clicks. It is also, note, 2.3% per word - a real effect that is not remotely the same size as the cultural anxiety attached to it.

The authors’ own limitations are unusually candid. The publisher used clickbait headlines throughout, limiting generalisability. The conclusions apply to stories, not to users. The emotion dictionary was validated at only r = 0.303, which is weak. And headlines performed worse later in the archive’s timeline, suggesting wear-out - the technique decays as audiences habituate.

A 2025 robustness reanalysis found the negativity effect held while a secondary claim did not - a partial replication, which is the ordinary fate of large observational-adjacent findings.

Outrage as social learning, at a modest size

Brady, McLoughlin, Doan and Crockett (2021), in Science Advances, tested whether social feedback teaches people to be angrier. 7,331 users and 12.7 million tweets observationally, plus two behavioural experiments of 120 participants each.

Feedback did predict subsequent outrage - b = 0.03 and b = 0.02, both p < .001. In plain terms, roughly a 2-3% increase in outrage expression per 100% increase in feedback. Real, and small.

The moderation result is the one worth carrying: users in ideologically extreme networks were less sensitive to social feedback (b = -0.02, p = .004; b = -0.05, p < .001). Reinforcement matters least exactly where outrage is highest - which means social learning cannot be the main explanation for the angriest corners of the internet.

The authors flag that their outrage classifier is 75% accurate, that the design is observational, and that the experiments used simulated networks with 120 people.

What this does not establish

That platforms deliberately optimise for anger. No study here examines a ranking system’s objective function. What has been measured is what spreads, not what a company chose.

That rage bait is making society angrier. Everything on engagement is correlational except the headline A/B tests, and those measure clicks, not attitudes. The filter bubble evidence is a useful check here: four randomised platform experiments changed exposure enormously and attitudes barely at all.

That the effect is stable. The Upworthy data show wear-out, and the meta-analytic estimate fell as the evidence base grew.

A clean mechanism. Moral-emotional language, out-group language and negativity are three different predictors from three different studies, and the largest comparison ranks the mechanism from the most famous study last.

We have not printed confidence intervals for the 2017 study or the out-group figures, because the sources do not state them.

A boundary worth stating

This site has a separate article on the curiosity gap, which draws on the same Upworthy A/B archive. The two are asking different questions of it. The curiosity gap is about withholding information; rage bait is about negative framing. Robertson’s negativity finding and the concreteness finding in that article are independent results from a shared dataset, and neither substitutes for the other.

What follows

The effect is real, causal for clicks, and about 2.3% per negative word. That is worth knowing precisely, because the precise version does not support the apocalyptic version.

What spreads is out-group content more than anger as such. If you want to predict what will travel, the better question is not “how angry is this” but “who is this about.”

It wears out. The technique performed worse over time in the one dataset large enough to see it. That fits the pattern in this programme’s other work on manipulation: confirmshaming is detected by 71% of people and shifts behaviour by about five points; infinite scroll was deployed globally with no experiment behind it. Visible techniques get discounted.

And notice what the field did. A famous number was published, was seriously challenged, and the original authors ran a preregistered replication and a meta-analysis that revised their own estimate downward by a third. That is what a healthy literature looks like - and it is why the 13% figure deserves more trust than the 20% one ever did.

Rage bait is one of several mechanisms shaping what reaches you - see the persuasive design guide for the wider set, the filter bubble for what algorithmic curation does and does not do, and the curiosity gap for the other half of the headline literature.