Last updated: 31 October 2026.
Misinformation spread looks like an epidemic from a distance: a claim appears, moves through a population, peaks and fades, and some people seem more resistant than others. The comparison is useful because epidemiology has well-developed tools for thinking about contagion, immunity and targeting. It is also dangerous when taken literally, because sharing is a choice, not an infection. This article builds a small compartment model of misinformation, runs a toy simulation that shows why a few highly active sharers change the picture, reviews what empirical studies have found, and sets out where the analogy breaks.
Key takeaways
- A compartment model (susceptible, exposed, sharing, immune) captures the basic loop: more sharers produce more exposure, which produces more sharers.
- What matters is not average reach but its distribution. Two networks with the same average reach can produce very different tails of large cascades.
- Empirical work finds concentrated sharing: in one large study of Twitter in 2016, 0.1% of users accounted for nearly 80% of shares from fake-news sources.
- Interventions act on different parts of the structure: prebunking changes susceptibility, platform design changes contact and delay, and targeting changes who the spreaders are. Each has trade-offs.
The compartment model
Epidemiologists divide a population into compartments and describe the flows between them. A common form for misinformation uses four:
- Susceptible (S): has not met the claim and could be persuaded by it.
- Exposed (E): has seen the claim but has not yet decided to act on it.
- Sharing (I): is passing the claim on.
- Immune (R): will not pass it on, because they saw through it, ignore it, or have been corrected.
Three flows connect them. Contact moves people from S to E at a rate proportional to the number of sharers they are connected to. A delay moves some of the exposed to the sharing group. And sharers move to immune as they lose interest, are corrected, or attention shifts. A fourth, dashed flow moves people directly from S to R before they are ever exposed, which is what prebunking tries to do.
Figure 1. Four compartments with a reinforcing loop through contact. Prebunking moves people to immune before exposure, and a small number of supersharers carry a large share of the sharing.
The structure is a reinforcing feedback loop, as described in reinforcing and balancing feedback loops, limited by a balancing loop: as the susceptible pool shrinks, each sharer finds fewer new people to reach. That is why such spread curves are S-shaped: slow start, rapid middle, saturation.
The reproduction number, translated
The key quantity in an epidemic is R0, the average number of new cases caused by one case in a fully susceptible population. The translation for misinformation is the average number of new sharers produced by one sharer. It equals the average number of people a sharer exposes times the probability that an exposed person shares.
Two consequences follow directly from the arithmetic:
- The threshold is 1. If each sharer produces fewer than one new sharer on average, chains die out. If more, the claim can spread widely.
- Immunity needed. If R0 is 3, the share of people who must be immune to stop sustained spread is 1 − 1/3, about 67%. If R0 is 1.5, it is about 33%. Claims differ enormously in R0, so one blanket intervention level does not fit all.
Why averages mislead: a toy simulation of supersharers
Real sharing is not uniform. A few accounts have very large audiences and share frequently, while most people share rarely. To see what that does, I ran a simple branching-process simulation with two network structures that have the same average reach of about 10 people per sharer, and the same 9% chance that an exposed person shares. In both, the average number of new sharers per sharer is about 0.9, so cascades on average die out.
- Uniform network: every sharer reaches exactly 10 people.
- Heavy-tailed network: 99.9% of sharers reach 8 people, and 0.1% (the supersharers) reach 2,000.
I started 200,000 cascades from a random sharer in each. This is a toy model with no real network and ignores depletion of the audience, so it illustrates a mechanism rather than predicting reality.
| Measure | Uniform | Heavy-tailed |
|---|---|---|
| Average cascade size | 10.0 | 10.0 |
| Cascades reaching 100 or more | 1.7% | 0.37% |
| Cascades reaching 1,000 or more | 0.001% | 0.17% |
| Largest cascade | 1,095 | 21,893 |
The averages are identical, yet the heavy-tailed network produces about 170 times more cascades of 1,000 or more, and the biggest one is twenty times larger. Typical cascades are smaller (fewer reach 100), because most sharers reach few people, but the rare ones are enormous. The lesson is that mean reach hides what matters for risk, a point familiar from power laws and heavy tails. A policy judged on the average sharer will miss the cases that cause the harm.
What the empirical studies show
Falsehood travels differently
Soroush Vosoughi, Deb Roy and Sinan Aral analysed about 126,000 news cascades tweeted by roughly 3 million people between 2006 and 2017, using the verdicts of six fact-checking organisations. They found that false stories spread farther, faster, deeper and more broadly than true ones, and were about 70% more likely to be retweeted. True stories took about six times as long to reach 1,500 people. Removing bots did not close the gap, so human sharing was the main driver. The authors proposed novelty as an explanation and were explicit that they could not claim to have shown that it causes the effect.
Sharing is concentrated
Nir Grinberg and colleagues studied registered voters on Twitter during the 2016 US presidential election. They reported that 1% of individuals accounted for 80% of exposures to fake-news sources, and 0.1% accounted for nearly 80% of shares from those sources. Those most engaged with such sources were more likely to be conservative-leaning, older and highly engaged with political news. This is the empirical counterpart to the toy model: concentration in a small group, with real consequences for where interventions should aim.
Prebunking can build resistance
Jon Roozenbeek, Sander van der Linden and colleagues tested short animated videos that teach people to recognise manipulation techniques such as scapegoating, false dichotomies and incoherence. In controlled experiments, people who watched them were roughly twice as good as controls at spotting some of these techniques. In a field test on YouTube, where about 5.4 million US users saw a video, recognition rose by about 5% on average among respondents. The authors acknowledge limits: only two videos were tested in the field, the measurement used fictional posts shortly after viewing, and the method targets tactics rather than specific false claims.
Where the epidemic analogy breaks
- Sharing is a choice with motives. People share to signal identity, to warn others, for humour, or because they think the claim is true. Viruses have no motives. Reasons for sharing differ, so a single transmission probability is a crude summary.
- Complex contagion. Some behaviours need reinforcement from several contacts before a person adopts them, unlike a virus that needs one exposure. Experiments by Damon Centola have shown that such behaviours spread differently across networks than simple contagion. Clustered networks can help them spread.
- Immunity is leaky and topic-specific. Someone sceptical about one kind of claim may be credulous about another. Immunity also fades.
- Labels are contested. Deciding what counts as misinformation involves judgement, and fact-checks can be wrong or disputed. A model needs the assumption that some claims are false; applying it to real content requires someone to decide which, with all the trade-offs that follow.
- The platform is part of the system. Ranking, recommendation and sharing features change contact rates and delays. The recommender loops covered in filter bubbles and recommender loops shape who is exposed, and the optimisation targets feed back into what spreads, as in Goodhart’s law in feedback systems.
Reading the levers through the model
| Lever | Part of the model it acts on | Trade-offs to weigh |
|---|---|---|
| Prebunking and media literacy | Moves people S to R, lowers the chance an exposed person shares | Effects are modest and fade; targets tactics, not every claim |
| Friction before sharing (prompts, delays) | Adds delay and lowers the sharing probability | Can irritate users; applies to true content too |
| Limiting reach of repeat spreaders | Cuts the tail of the reach distribution | Requires accurate identification and raises speech and fairness questions |
| Ranking changes | Lowers contact rate for flagged content | Depends on classifier accuracy and transparency |
| Corrections | Moves sharers and exposed to R | Often arrive late and reach fewer people than the original |
A recurring mismatch is timing. Spread is fastest early, when the reproduction number is highest and the susceptible pool is full. Corrections usually come after the peak. This is a delay problem, and interventions that act before or at the start, such as prebunking and friction, have an inherent advantage in the model. That advantage must be weighed against their costs; the model alone cannot decide those trade-offs.
How to use the model in practice
- Estimate distribution, not just averages. Measure how reach and sharing frequency are distributed, and report percentiles.
- Look for the tipping behaviour. Find which claims have a reproduction number near the threshold. Small changes in friction or ranking can move those below it, as in the discussion of a tipping point.
- Match the intervention to the compartment. Prebunk where susceptibility is high, add friction where speed is the problem, target repeat spreaders where concentration is high.
- Test for side effects. Run the same intervention on true content to see what it costs.
- Monitor the loop. Track the effect on measured spread and keep watching, because spreaders and audiences adapt.
Frequently asked questions
Is misinformation really like an epidemic?
Partly. The structure of contagion, immunity and super-spreading carries over and gives useful quantities such as the reproduction number. The analogy fails where motives, identity and platform design matter, so use it as a model of structure, not as a literal claim about people being infected.
Who are the superspreaders?
In the study of Twitter during the 2016 US election, a very small group of accounts, 0.1% of users, accounted for nearly 80% of shares from fake-news sources. Studies on other platforms and periods differ, so check the data for the context you care about.
Does prebunking work?
Controlled experiments show improvements in people’s ability to recognise manipulation techniques, and a large field test found a small average gain. Effects decay and the evidence on lasting changes in sharing behaviour is more limited.
Why do corrections often fail to catch up?
Because spread is fastest early and corrections arrive later to fewer people. In the model the reproduction number is highest when the susceptible pool is full, so the window for effective action is short.
Conclusion
Misinformation spread is best understood as a network process with a reinforcing core, a delay, and a heavy-tailed distribution of spreaders. The compartment model gives a shared language for where to intervene, and the toy simulation shows why averages hide the risk. The empirical studies support the picture of concentration and of false content travelling faster, while prebunking evidence shows modest gains. The honest summary is that the model clarifies the structure but not the values: deciding what to restrict, and at what cost, remains a human decision.
See the glossary entries for misinformation spread, feedback loop and network effect. For a related cascade perspective, read cascading failures.
Sources and further reading
- Vosoughi, S., Roy, D. and Aral, S., “The spread of true and false news online”, Science, 2018, summary from the release.
- Grinberg, N. et al., “Fake news on Twitter during the 2016 U.S. presidential election”, Science, 2019.
- Roozenbeek, J. et al., “Psychological inoculation improves resilience against misinformation on social media”, Science Advances, 2022, University of Cambridge summary.
- Centola, D., “The spread of behavior in an online social network experiment”, Science, 2010.
Discussion