The Algorithm That Makes Everyone Look Furious

Conceptual illustration of a person looking at a phone while arguing faces appear in a window reflection.

AI-generated editorial illustration.

A furious comment section can tell you what attracted attention. It cannot, by itself, tell you what the country thinks. Research helps explain the gap—and why blaming everything on “the algorithm” is too simple.

Open a political post and the same world seems to appear: everyone is furious, everyone despises the other side, and nobody will reconsider anything. Scroll long enough and ordinary disagreement begins to look like permanent civil war. But the screen is a selection of human behaviour, arranged by a technical system and interpreted by a fallible reader. Each stage can change the picture.

This article examines peer-reviewed research on sharing, outrage, repetition, corrections and feed experiments. It separates observed associations from experiments that can support causal conclusions. The evidence includes findings published in 2026, rather than treating older Facebook studies as the final word on today’s platforms.

1. Anger has an audience advantage

A 2017 study examined 563,312 tweets about three polarising issues. Moral-emotional language was associated with greater diffusion: approximately 20% more expected retweets per additional moral-emotional word in the original analysis. The association was especially pronounced within ideological networks. This was an analysis of naturally occurring messages, so it did not establish that inserting a particular word would make any individual post go viral. Brady et al. (2017): moral-emotional language and diffusion.

The newer evidence matters. A 2025 preregistered replication and meta-analysis covered 27 studies and approximately 4.82 million observations. Its pooled estimate was smaller: around 13% more expected shares per additional moral-emotional word, with a 95% confidence interval corresponding to roughly 6%–20%. The relationship survived further testing, but its magnitude was not a universal “20% rule”. Brady et al. (2025): preregistered replication and meta-analysis.

Replication changes the headline number

Estimated relative increase in expected sharing per additional moral-emotional word.

Original study, 2017: 20%
Pooled estimate, 2025: 13%
Bars start at zero. Separate analyses, not a time trend. The pooled 95% interval is approximately 6%–20%. Observational associations; no guaranteed boost for a particular post. Sources: Brady et al. (2017): moral-emotional language and diffusion; Brady et al. (2025): preregistered replication and meta-analysis.

2. Attacking the other side can outperform praising your own

Rathje and colleagues analysed political material on Facebook and Twitter. Terms referring to the political out-group were associated with increased sharing; the reported estimate was a 67% increase in the odds of being shared for each additional out-group term. That is an odds estimate, not a 67-percentage-point increase in probability. It also does not show that every mention of an opponent is abusive. Rathje et al. (2021): out-group animosity and engagement.

The practical implication is uncomfortable. A communicator may receive stronger visible rewards for defining an enemy than explaining a workable solution. A sensible funding proposal needs qualifications. An accusation needs a villain and a share button. That is an editorial inference from the incentive pattern, not proof of what motivated any named politician.

3. People can learn which version of themselves gets rewarded

A 2021 study combined two preregistered observational Twitter studies—7,331 users and 12.7 million tweets—with two behavioural experiments involving 240 participants. Positive feedback for expressions of outrage predicted further outrage expression, and the experiments supported a social-learning account. People also responded to the norms of their networks. Brady et al. (2021): social learning and outrage.

Expression and emotion are different things. Writing more angry posts does not necessarily mean feeling more anger. A person can learn that a sharper performance attracts approval without undergoing a matching emotional transformation. Nor does social learning make all outrage fake: anger about abuse, corruption or injustice can be justified.

A possible feedback loop

Provocative post → visible reactions → learned audience expectations → another provocative post

Conceptual diagram based on social-learning research. It is not a measured model of every user, and it does not establish the contribution of a platform’s ranking system. Brady et al. (2021): social learning and outrage.

4. We can mistake a loud expression for a furious population

In 2023, researchers compared how much outrage authors reported feeling with how much observers perceived in their posts. Observers tended to overestimate the outrage, and the research linked those perceptions with exaggerated beliefs about hostility between groups. This helps explain how reading political material can make the social world feel more antagonistic than its authors’ reported feelings would suggest. Brady et al. (2023): overperception of outrage.

There are also basic sampling problems. A comment thread includes people who encountered a post, chose to respond and remained visible under that thread’s sorting or moderation. It excludes many readers who said nothing and everyone who never saw it. The number of comments does not repair those selection stages.

A thousand comments saying the same thing may be politically interesting. They are not a probability sample of a constituency. An emoji count does not reveal whether someone agrees with the claim, mocks it, dislikes its author or simply wants to bookmark the argument. Before inferring a national mood, ask who was eligible to appear in the data and who is absent.

5. Repetition can make a claim feel more credible

Experiments by Pennycook, Cannon and Rand found that prior exposure increased the perceived accuracy of fabricated news headlines. This is one example of the illusory-truth effect: familiarity can influence judgments of truth. The paper also identified limits, including extreme implausibility. It does not imply that repetition makes everyone believe anything. Pennycook, Cannon and Rand (2018): repetition and perceived accuracy.

For a reader, the useful distinction is between recognition and verification. “I have seen that figure everywhere” tells us that a message travelled. It does not tell us whether its denominator, date or source is correct. Ten accounts can repeat one unsupported calculation without producing ten independent pieces of evidence.

This applies to our own side too. A satisfying rebuttal deserves the same checks as the allegation it answers. If a graphic compares one country’s quarterly GDP growth with another country’s multi-year GDP per head, visual polish cannot make the comparison meaningful.

6. But do algorithms actually change political opinions?

A major experiment published in Science in 2023 assigned consenting Facebook and Instagram users to chronological feeds instead of the usual ranking during the 2020 US election period. The intervention substantially changed their platform experience. However, over the three-month study, it did not significantly change issue polarization, affective polarization, political knowledge or other key attitudes examined. Guess et al. (2023): Facebook and Instagram feed experiment.

That finding rules out a sweeping claim that altering a feed must quickly change all political attitudes. It does not establish that ranking has no effect anywhere, that a different intervention would fail, or that long-term exposure is harmless. A null result has a defined setting, duration and set of measurements.

A 2026 Nature paper adds a different result. In a seven-week experiment conducted on X in 2023, switching users from chronological to algorithmic feeds shifted some policy and current-affairs opinions in a more conservative direction. Switching the algorithm off produced no comparable reversal. Neither direction significantly changed affective polarization or self-reported partisanship. The main post-treatment sample contained 4,965 participants. Gauthier et al. (2026): political effects of X’s feed algorithm.

The authors also found changes in account-following behaviour that help explain why switching off a feed may not undo its earlier effects. These results concern particular US users and the X system studied in 2023. They cannot establish the size of an effect on British voting in 2026. The responsible conclusion is that platform, exposure history and outcome measure matter.

Two feed experiments: different settings, different results
StudyIntervention and settingMain distinction
Guess et al., 2023Chronological versus ranked Facebook/Instagram feeds; 2020 US election; three monthsNo significant change in several measured political attitudes
Gauthier et al., 2026Switching X feeds on/off; US users in 2023; seven weeksSwitching on shifted some opinions; no significant change in partisanship or affective polarization
Different outcomes should not be combined into a single “social media effect”. Guess et al. (2023): Facebook and Instagram feed experiment; Gauthier et al. (2026): political effects of X’s feed algorithm.

7. Correcting people is not automatically pointless

The familiar claim that facts simply make people dig in is too broad. Wood and Porter tested corrections across five experiments involving more than 10,100 participants and 52 potentially contentious issues. They found no corrections that triggered the factual backfire effect they were testing. The work was published in Political Behavior. Wood and Porter: The Elusive Backfire Effect.

A correction can improve a factual belief without changing a party preference, an identity or someone’s willingness to admit error publicly. Those are different outcomes. Silence after a correction is not a measurement of failure; an angry reply is not proof that every reader rejected the evidence.

Ridicule needs similar care. It may entertain allies, but the studies above do not establish that insulting a stranger is an effective persuasion technique. If the aim is public accountability, a precise correction gives bystanders something usable: the original claim, the source, the mismatch and the corrected comparison. Humiliation is a separate objective.

8. A small pause can help

A 2021 Nature study found that prompting people to consider accuracy could improve the quality of news they subsequently shared, including in a Twitter field experiment. A 2022 analysis of 20 experiments involving 26,863 participants examined the replicability and generalizability of accuracy prompts. These interventions are evidence against the idea that everyone sharing misinformation is knowingly lying or completely indifferent to truth. Pennycook et al. (2021): accuracy prompts; Pennycook and Rand (2022): 20 accuracy-prompt experiments.

A prompt is not a complete solution to propaganda, financial incentives or deliberate deception. It is a practical point of friction. Before sharing, ask: is this accurate, and what would let another person check it? That question can be useful even when the post agrees with everything you already believe.

9. What a serious response looks like

For readers: open the source, check the publication date and identify what was actually measured. Separate percentages from percentage points, totals from rates and observed changes from forecasts. A screenshot preserves a claim but may omit the context that determines whether it is fair.

For publishers: link the underlying document beside the relevant sentence. Label calculations and editorial inferences. Preserve corrections where readers can see them. If a claim is false, explain the specific error; if intent is unproven, do not pretend the evidence establishes deliberate lying.

For platforms: a defensible research agenda would test changes to ranking, sharing friction and contextual information against several outcomes. More clicks, more time spent, more accurate beliefs and less hostility are different objectives. A proposal should say which outcome it seeks and what evidence would count as success.

For political communicators: attention is a poor substitute for an answer. A claim about welfare spending needs the calculation and population denominator. A claim about migration needs the correct category and period. A claim about crime needs a consistent measure. Repeating an accusation louder does not resolve missing evidence.

10. What this evidence cannot tell us

These studies do not diagnose the motives, intelligence or mental health of people in a screenshot. They do not show that every viral post is false, that every angry person is manipulated, or that all political camps behave identically. Much of the evidence comes from US settings and particular versions of platforms. Language, regulation, political institutions and technical systems change.

Large datasets can estimate an association precisely while still leaving a causal question unresolved. Experiments provide stronger causal evidence for their tested intervention, but have limits involving recruitment, compliance, duration and generalisation. Replications can retain an effect while revising its size. Those qualifications belong in the explanation, not hidden after the headline.

Watch the researchers explain

These videos provide accessible background. The written studies above support the specific findings; an older talk should not be treated as covering later results.

Molly Crockett: “Moral outrage in the digital age”. Oxford Martin School source page. Historical background, predating the 2023 and 2026 feed experiments.
Sander van der Linden discusses misinformation with the American Psychological Association. Episode and transcript.

The screen is evidence of what reached you

The anger can be real. The wrong is assuming that the people on screen are a representative picture of everyone else. Human choices, social rewards, ranking and interpretation can all shape that impression. Their contributions need to be tested rather than guessed.

A comment section can show us an argument. It takes better evidence to show us a country.

Research and source links

  1. Brady et al. (2017): moral-emotional language and diffusion
  2. Brady et al. (2025): preregistered replication and meta-analysis
  3. Rathje et al. (2021): out-group animosity and engagement
  4. Brady et al. (2021): social learning and outrage
  5. Brady et al. (2023): overperception of outrage
  6. Pennycook, Cannon and Rand (2018): repetition and perceived accuracy
  7. Guess et al. (2023): Facebook and Instagram feed experiment
  8. Gauthier et al. (2026): political effects of X’s feed algorithm
  9. Wood and Porter: The Elusive Backfire Effect
  10. Pennycook et al. (2021): accuracy prompts
  11. Pennycook and Rand (2022): 20 accuracy-prompt experiments

Research checked 20 November 2026. Figures are attributed to their original studies. Visualisations are explanatory and identify their sources; the cover is an AI-generated editorial illustration.

Leave a comment