How Misinformation Spreads Online — And What Actually Stops It

In 2018, a team led by Soroush Vosoughi, Deb Roy, and Sinan Aral at MIT published the largest study of rumor spread on Twitter to that date. They followed about 126,000 news cascades, tweeted by roughly 3 million people between 2006 and 2017. False stories were 70 percent more likely to be retweeted than true ones, and true stories took about six times as long to reach 1,500 people. Removing bots from the data did not close the gap. People did the sharing, and that is the starting point for understanding what works against it.

Key takeaways

• In the 2018 MIT analysis, false stories were 70 percent more likely to be retweeted and reached 1,500 people about six times faster than true ones, even with bots removed.
• Novelty, emotion, and inattention explain much of the pattern. Most people can judge accuracy when asked, but accuracy is rarely on their mind while scrolling.
• Engagement-ranked feeds amplify what provokes reactions. Internal Facebook documents show reactions were once weighted five times more than likes.
• Super-sharers matter. In one 2020 study, 2,107 accounts produced about four-fifths of the false news shared by a panel of 664,391 voters.
• Accuracy prompts, prebunking, friction, and fast crowd-sourced notes all work a little. None works alone, and most effects fade or arrive late.

What the 2018 MIT study measured, and what it did not

The study tracked "rumor cascades": chains of retweets that start with one original post asserting a claim. Six independent fact-checking organizations judged each claim, and their verdicts agreed more than 95 percent of the time, according to MIT's summary of the work. Politics was the largest category, with about 45,000 cascades.

Three numbers from that study are worth keeping. False stories were 70 percent more likely to be retweeted. True stories needed about six times as long to reach 1,500 people. And falsehoods reached a cascade depth of 10 (ten consecutive retweets, one person passing it to the next) about 20 times faster than facts.

The limits matter just as much. The study covered Twitter only, and only claims that fact-checkers had bothered to examine. It measured spread, not belief. A false story that is shared widely is not necessarily a story that changes minds, and the authors themselves said they could not claim that novelty alone causes retweets.

Think of it like measuring which songs get played at parties. You learn what travels, not what anyone believes about the lyrics.

Misinformation, disinformation, and "fake news": three different things

Researchers separate these terms because the remedies differ. A person who shares a false claim in good faith needs a different intervention from a coordinated operation that manufactures one.

Term	What it means	Intent	Typical example
Misinformation	False or misleading content, regardless of intent	Often none	A relative reposts an invented quote without checking
Disinformation	False content spread deliberately to deceive	Yes	A coordinated network seeds a fabricated story
"Fake news"	Colloquial label, used both for fabricated sites and for coverage someone dislikes	Varies	Fabricated "news" sites versus accurate but unflattering reports

This article uses "misinformation" as the broad category and flags disinformation where intent is documented. Most of the sharing studies below involve ordinary users, not coordinated campaigns.

Why false claims travel faster

Novelty: the "did you hear" effect

The MIT team found that false stories scored as more novel than true ones, meaning they differed more from what a user had recently seen in their feed. Sharing something new signals that you are plugged in. A story that confirms what everyone already knows has little social value.

True reporting on complex events is often incremental. A correction to an earlier report, a revised casualty count, and a cautious scientific finding rarely give anyone a "you won't believe this" moment.

Emotion: surprise and disgust

The same study analyzed replies to each cascade. Replies to false stories showed more surprise and disgust, while replies to true stories showed more sadness, anticipation, and trust. High-arousal emotions push people to act, and on a social platform acting means sharing.

This is a correlation in replies, not proof that emotion caused the sharing. But it fits a long line of findings that emotionally charged posts spread more readily.

A quick analogy: the leaky bucket and the garden hose

Picture a feed as a garden hose spraying water into a field of buckets. Accurate, nuanced stories arrive as a trickle. Provocative stories arrive in a jet. Nothing about the buckets (the readers) has to be broken for the jet to fill them faster. The shape of the flow does most of the work.

The psychology underneath: three vulnerabilities

Repetition feels like truth

Psychologists call it the illusory truth effect: statements you have encountered before are judged more likely to be true than new ones. A 2015 paper by Lisa Fazio, Nadia Brashier, B. Keith Payne, and Elizabeth Marsh in the Journal of Experimental Psychology: General found the effect even for claims people knew to be false. The authors called it "knowledge neglect": people can hold the right fact and still be nudged by familiarity.

Social platforms create repetition at scale. A claim that appears in 10 different feeds in an afternoon starts to feel familiar, whatever its source. Our article on why conspiracy theories spread online shows how the same repetition loop operates in tightly knit communities.

Wanting a claim to be true

People weigh evidence differently when a claim touches their identity or group loyalty. This is motivated reasoning: reasoning that works backward from a preferred conclusion. It affects everyone, and the psychology of motivated reasoning explains why more education does not make anyone immune.

A related trap is that people judge claims by how many others seem to believe them. Social proof turns a share counter into a credibility signal, even when the counter is inflated by a few busy accounts.

Not thinking about accuracy at all

The third vulnerability may be the most hopeful. David Rand, Gordon Pennycook, and colleagues argue that many people share false content not because they believe it, but because accuracy is not what they are thinking about when they decide to share. On social media, attention goes to what will get likes, what is funny, and what signals identity.

The evidence comes from experiments in which participants rated headlines for accuracy and, separately, said whether they would share them. People often rated a false headline as inaccurate yet still said they might share it. Prompting them to think about accuracy first reduced the gap. An MIT Sloan summary of the work quotes Rand: "social media makes us share things that we would think better of if we stopped to think."

How platform design amplifies the pattern

Engagement as the ranking signal

Most large feeds rank posts by predicted engagement: clicks, likes, comments, shares, time spent. Nothing in that objective asks whether a post is true. Content that provokes reactions rises, and what provokes reactions overlaps with what is false, outrageous, or both.

Documents disclosed by whistleblower Frances Haugen in 2021 gave an unusually concrete view. According to reporting on internal Facebook documents, summarized by Nieman Lab, starting in 2017 the News Feed ranking system counted emoji reactions as five times more valuable than a "like." Employees raised concerns in the documents that this could favor controversial content, and in 2019 company data scientists reportedly found that posts drawing angry reactions were disproportionately likely to contain misinformation, toxicity, and low-quality news. Facebook later reduced that weighting; accounts of the exact schedule differ across reports.

The counterintuitive experiment: switching algorithms off

If the algorithm is the culprit, removing it should help. A 2023 Science study led by Andrew Guess tested that during the 2020 US election, with about 23,000 consenting Facebook users and 21,000 Instagram users. Moving users from the default algorithmic feed to a reverse-chronological feed cut the time they spent on the platforms.

It also changed what they saw. On both platforms, users saw more political content and more content from untrustworthy sources, while on Facebook they saw less uncivil content. Over three months, the chronological feed did not significantly change issue polarization, affective polarization, or political knowledge. A later critique argued that Meta made temporary changes to its feed during the experiment, which may complicate the comparison.

The lesson is not that algorithms are harmless. It is that "turn off the algorithm" is not a cure. Feed ranking reshapes exposure, but the demand for provocative content sits partly in the people using it.

Design features that reward speed

A platform that shows a reshare button beside every post, displays raw share counts, and autoplays the next clip makes sharing nearly effortless. Our look at infinite scroll psychology explains how these features keep attention moving, and our explainer on how the TikTok algorithm works covers recommendation systems that can place a claim in front of strangers within hours.

Who actually shares misinformation

A small group does a large share

Two Science papers by the same research tradition describe concentration. Nir Grinberg and colleagues studied a panel of Twitter accounts linked to voter records in the 2016 US election and reported that 0.1 percent of users accounted for nearly 80 percent of the fake news sources shared (Science, 2019). A 2024 follow-up by Sahar Baribi-Bartov, Briony Swire-Thompson, and Grinberg covered 664,391 registered voters during the 2020 election. Just 2,107 of them produced about four-fifths of the false news the panel shared.

These "supersharers" reached about 5.2 percent of registered voters on the platform. They were disproportionately women, older adults, and registered Republicans, and their volume did not look automated. It came from manual, persistent retweeting. If a few thousand accounts supply most of the supply, targeted platform policies on repeat offenders can matter more than broad education campaigns.

Age and ideology: patterns, not stereotypes

A 2019 Science Advances study by Andrew Guess, Jonathan Nagler, and Joshua Tucker linked a survey to Facebook activity for 1,191 respondents. Only 8.5 percent of them had shared at least one article from a fake news domain, so sharing was rare. But respondents over 65 shared more than twice as many as the next-oldest group, and roughly seven times as many as the 18-to-29 group after controlling for ideology and education. The authors called age their "most robust" finding.

The same paper found that conservatives shared more fake-news-domain articles than liberals or moderates, with caveats: the result was sensitive to coding, rested on a small number of sharers, and may reflect the slant of the fake news that happened to be available in 2016. A single election cycle on a single platform cannot settle whether any group is inherently more susceptible.

How big is the problem in the first place?

A 2020 Science Advances analysis by Jennifer Allen and colleagues combined Nielsen and Comscore panels covering January 2016 through December 2018. It found that fake news made up only 0.15 percent of Americans' daily media diet, and that news of any kind made up at most 14.2 percent. The authors cautioned that their measure captures prevalence, not impact, and suggested that ordinary news framing and news avoidance may matter as much as fabricated stories.

Finding	Source	What it suggests
0.1% of users shared nearly 80% of fake news sources	Grinberg et al., Science, 2019	Supply is concentrated
2,107 of 664,391 voters shared about 80% of false news	Baribi-Bartov et al., Science, 2024	Concentration persisted in 2020
8.5% of linked Facebook users shared a fake news article	Guess et al., Science Advances, 2019	Sharing is rare for most people
0.15% of daily media diet is fake news	Allen et al., Science Advances, 2020	Prevalence is low even if impact is large

Put together, these studies describe a problem that is concentrated and heavy-tailed, not a flood that reaches every person equally. That framing changes which interventions make sense.

What the evidence says about fixes

Researchers distinguish interventions by when they act: before exposure (prebunking), at the moment of sharing (nudges and friction), or after exposure (corrections and labels). The table below summarizes what the studies cited in this article report.

Intervention	When it acts	Evidence	Main limitation
Accuracy prompt	At sharing	Survey and Twitter field experiment, Pennycook et al., Nature, 2021	Small effects; novelty can wear off
Prebunking videos	Before exposure	Six randomized studies plus a YouTube field study, Roozenbeek et al., Science Advances, 2022	About 5 percent average gain in the field; decay over time
Read-before-retweet prompt	At sharing	Twitter's own report of 40 percent more article opens, 2020	Company-reported, not independently verified
Community Notes	After exposure	Slaughter et al., PNAS, 2025	Effect depends heavily on how fast a note appears
Fact-check corrections	After exposure	Wood and Porter, Political Behavior, 2019	Reach is limited; belief change is modest

Accuracy prompts

The most cited nudge comes from a 2021 Nature paper by Gordon Pennycook, Ziv Epstein, Mohsen Mosleh, Antonio Arechar, Dean Eckles, and David Rand. In survey experiments, asking people to judge the accuracy of one headline made them more discerning about what they said they would share. In the field experiment, researchers used bot accounts to message over 5,000 Twitter users who had previously shared news from websites known for publishing misleading content. The message asked for a rating of a random non-political headline, and according to the MIT Sloan summary, users subsequently retweeted higher-quality news.

The paper reports a statistically significant improvement, not a dramatic one, and it tested a short window after the message. You will sometimes see a precise percentage attached to this study in popular coverage. Treat it with caution unless it comes from the paper itself.

Prebunking

Prebunking borrows from vaccine logic. Expose people to a weakened version of a manipulation technique, and they recognize the real thing faster. A 2022 Science Advances paper by Jon Roozenbeek, Sander van der Linden, and colleagues reported seven preregistered studies: six randomized experiments with 6,464 participants and a YouTube field study. The videos taught five techniques: emotionally manipulative language, incoherence, false dichotomies, scapegoating, and ad hominem attacks. They used fictitious, non-political, humorous examples to avoid partisan reactions.

In the field study run by Google's Jigsaw unit, about 5.4 million US YouTube users saw a video, and almost one million watched at least 30 seconds. Among the 22,632 who answered a test question, the ability to recognize manipulation techniques rose by about 5 percent on average, according to the University of Bristol's release. The cost was about $0.05 per view of significant length.

A 5 percent gain sounds small. The researchers pointed out that typical brand-lift measurements for YouTube ads are around 1 percent, and that prebunking is cheap to run at scale. It is a public-health style measure, small per person and meaningful across millions.

Friction at the moment of sharing

In June 2020, Twitter began testing a prompt on Android that asked users whether they wanted to read an article before retweeting it. The company reported in September 2020 that people opened articles 40 percent more often after seeing the prompt, and that some did not retweet after opening. That is a company-reported figure, and the check only counted opens on Twitter itself.

Friction is the cheapest tool: a delay or a prompt placed between the impulse and the share button. Its weakness is that it targets impulsive sharing, not motivated sharing by committed partisans or professional operators.

Community notes and crowd labels

X's Community Notes system lets approved contributors attach context to posts, and notes display only when contributors who usually disagree both rate them helpful. A 2025 PNAS study by Isaac Slaughter, Axel Peytavin, Johan Ugander, and Martin Saveski tracked 40,078 posts with proposed notes from March to June 2023, of which 6,757 received helpful notes. Using synthetic control methods, they found that after a note was attached, post growth in reposts fell 46.1 percent, likes 44.1 percent, replies 21.9 percent, and views 13.5 percent, according to the paper on PubMed Central.

The over-the-lifetime effects were much smaller (reposts down 11.6 percent), because notes often arrived after most of the spreading was done. For notes attached after 47 hours, the repost reduction was only about 6.2 percent. The authors concluded that speed is the lever. The analysis is observational and covers only posts that received notes.

Corrections and the "backfire" worry

For years, a popular idea held that correcting a false belief could make people believe it more strongly. A large study by Thomas Wood and Ethan Porter, published in Political Behavior in 2019, tested 52 issues across more than 10,100 subjects and found no corrections that triggered backfire. People generally moved a little toward the facts. This does not make corrections powerful. It means that the fear of backfire should not stop fact-checkers from publishing them.

What does not work well, or is not settled

• Debunking one claim at a time. Corrections do help individual readers, but they cannot keep pace with production, and many people who saw the original never see the correction.
• Media literacy training alone. It is useful, and the better programs teach concrete habits such as lateral reading (leaving a page to check what others say about its source). But general awareness does not change behavior at the moment of a tired late-night scroll.
• Deplatforming as a cure. Removing accounts reduces their reach on the platform that removes them. Whether the audience migrates and what that does to public discourse is still debated, and this article does not rely on a single study for either side.
• Assuming one intervention scales. Most effects were measured on specific platforms, in specific years, on specific populations. Effects measured in a two-week experiment may not hold a year later.

What AI changes, and what it does not

Generative tools lower the cost of producing a plausible article, image, or voice clip. The core dynamics described above (novelty, emotion, inattention, concentrated supply, engagement ranking) are unchanged, but the cost of supply drops. Our explainer on deepfakes and how they work covers the technical side, and the article on why AI hallucinates facts explains why a chatbot can produce confident errors without any intent to deceive.

One honest caution: the studies cited here were run mostly before widespread generative AI, so their effect sizes may not carry over. Detection tools, provenance labels, and prebunking for synthetic media are active research areas without settled answers.

Common misconceptions

"Bots spread most misinformation." In the MIT study, removing bots left the gap between true and false news intact. Bots exist and some campaigns use them, but ordinary retweeting did the work in that dataset.

"Only gullible people fall for it." The illusory truth research shows that prior knowledge does not fully protect people, and the sharing-versus-accuracy experiments show that people who know better still share. Gullibility is a weak explanation for a pattern this broad.

"Misinformation is everywhere in everyone's feed." The prevalence studies describe a concentrated problem. A few accounts do a lot of the sharing, and for most people false news is a small slice of what they see. That does not mean it is harmless, since measures of exposure are not measures of effect.

"If we fix the algorithm, we fix the problem." The 2023 chronological-feed experiment changed what users saw and how long they stayed, without shifting polarization measures. Feed design matters, but demand for provocative content, group identity, and offline media all play a role. Our guide to how to tell if a science claim is too neat offers some practical habits for evaluating claims, and why headlines overstate breakthrough science shows how accurate research gets distorted on the way to a headline.

The structural problem: incentives

Researchers tend to describe the core tension in similar terms. A feed ranked by engagement serves advertisers and keeps users scrolling, and the content that does best is often the content that is least careful. Platforms can reduce the effect through ranking changes, labels, and friction, but each costs engagement, so changes have tended to come under public or regulatory pressure.

That argues for measuring what platforms do. Independent researchers need access to data, and several of the studies above (the Meta election experiments, for example) were possible only through unusual cooperation. Without access, the public has to rely on company-reported numbers such as Twitter's 40 percent figure, which cannot be checked.

How to read a viral claim: five habits with research behind them

1. Pause and rate accuracy before sharing. This is the intervention with the best field evidence, and you can apply it to yourself for free.
2. Check the source, not just the post. Leave the page and see what other outlets say about the site. Lateral reading is how professional fact-checkers work.
3. Notice your emotional reaction. Strong surprise, disgust, or anger is a cue to slow down, since those are the emotions most associated with the false cascades in the MIT data.
4. Look for the original. Screenshots, quote cards, and clips are easy to edit or take out of context. Trace the claim to a primary document.
5. Be wary of a claim you have seen many times. Familiarity feels like evidence but is not. The illusory truth effect works even when you know better.

These habits are about the individual reader. They do not replace platform design or policy, but they cost nothing and the first one has experimental support.

The takeaway

Misinformation outpaces corrections because the properties that make a claim shareable (novelty, emotion, a flattering story about your group) have little to do with accuracy, and because feeds ranked by engagement amplify them. The data also show a narrower problem than the headlines suggest: a few thousand accounts supply a large share of false news, and for most people it is a small slice of their media diet. The tools that work do so in modest, measurable ways. Accuracy prompts shift attention, prebunking trains pattern recognition, friction slows impulses, and fast notes cut spread. Their combined effect depends less on any one clever trick than on whether platforms deploy them early and let outsiders check the results.

Frequently asked questions

Q: Is "fake news" the same as misinformation? A: No. Misinformation is false or misleading content regardless of intent, while disinformation is false content spread deliberately to deceive. "Fake news" is a colloquial label used both for fabricated stories and, politically, for coverage someone dislikes, so researchers generally avoid it.

Q: Do fact-checkers make a difference? A: Modestly. Wood and Porter's tests of 52 issues found corrections moved people toward the facts and did not cause backfire. Fact-checking cannot cover everything, and people who saw a false claim often never see the correction, so it works best alongside platform measures.

Q: Are bots the main reason false news spreads? A: Not in the 2018 MIT dataset. When researchers removed bots, false stories still spread farther and faster than true ones, so human retweeting accounted for the difference. Bots and coordinated networks do exist, but they are not the whole explanation.

Q: Do older people share more misinformation? A: In one 2019 study linking survey data to Facebook activity, users over 65 shared more than twice as many fake-news-domain articles as the next-oldest group. Only 8.5 percent of respondents shared any, so the finding describes a rare behavior, and it comes from the 2016 US election on one platform.

Q: What is prebunking, and does it work? A: Prebunking teaches people to recognize manipulation techniques before they meet specific false claims. In a Google Jigsaw field study on YouTube, recognition of those techniques rose by about 5 percent on average across thousands of respondents. The effect is small per person but cheap to deliver at scale.

Q: Would a chronological feed fix the problem? A: Evidence from a 2023 Facebook and Instagram experiment says not by itself. Users saw more political and untrustworthy content and spent less time on the platforms, and measured polarization did not change significantly over three months.

Q: Could AI help combat misinformation? A: Possibly in narrow roles, such as flagging coordinated behavior or speeding up fact-check distribution, but AI also lowers the cost of producing convincing fakes. Most of the evidence in this article predates widespread generative AI, so the net effect is uncertain.

Sources

• Vosoughi, Roy, and Aral summary: MIT News, "Study: On Twitter, false news travels faster than true stories"
• Accuracy prompt: MIT Sloan press summary of Pennycook et al., Nature, 2021
• Prebunking field study: University of Bristol news release on Roozenbeek et al., Science Advances, 2022
• Age and sharing: Guess, Nagler, and Tucker, Science Advances, 2019, on PubMed Central
• Media diet: Allen et al., Science Advances, 2020, on PubMed Central
• Community Notes: Slaughter et al., PNAS, 2025, on PubMed Central
• Chronological feed experiment: Guess et al., Science, 2023, record at Princeton
• Supersharers: Baribi-Bartov, Swire-Thompson, and Grinberg, Science, 2024

Related Reading

• What Are Deepfakes — and How Do They Actually Work?
• Why Conspiracy Theories Spread Online
• The Loftus Misinformation Effect: How Memory Gets Rewritten
• False Memory Psychology: Why Confidence Is Not Proof
• How Correlation Differs From Causation
• What Is 5G — and How Is It Really Different from 4G?
• What Are VR and AR — and Where Is the Metaverse Actually Going?
• What Is Neuralink? How Brain-Computer Interfaces Actually Work
• How Drones Actually Work — and Where They're Going