The Digital Whale

Social media and communication

Who actually sees and shares false content

Fake news exposure and sharing are concentrated in a tiny minority of users and amount to a fraction of a per cent of what people consume.

Misinformation is not evenly distributed and it is not abundant. Once researchers stopped counting posts and started measuring people, the same result kept appearing: a very small number of accounts see almost all of it, a smaller number share almost all of it, and across everything Americans read and watch it comes to a fraction of one per cent. Sharing rises steeply with age. None of that makes the problem imaginary — concentrated exposure can still do damage, and the most exposed are not a random sample — but descriptions of a population swimming in falsehood have no support in the measurement studies. The literature on correcting it is divided about whether anything works. For a workplace-oriented example of technology measurement in practice, see workplace stereotyping.

Exposure is concentrated in about one user in a hundred

The foundational study matched registered voters to their Twitter accounts and counted what reached their timelines.

2019

Roughly 1% of users accounted for 80% of all exposures to fake news sources, and 0.1% of users were responsible for 81% of the fake news shared.

Direction: No detectable effect. Strength of evidence: Strong.

Grinberg, Joseph, Friedland, Swire-Thompson & Lazer, Science, 201916,000+ US registered voters matched to Twitter accounts, 2016

Caveat One platform in one election, and the result depends on a source-level blacklist definition of "fake news" that teams operationalise differently.

Sharing is rare and rises sharply with age

A parallel Facebook study had panellists install a data-sharing app, so sharing was observed rather than recalled.

2019

Only 8.5% of surveyed Facebook users shared any article from a fake news domain during the 2016 campaign, and sharing rose with age from 3% of 18–29s to 11% of the over-65s.

Direction: No detectable effect. Strength of evidence: Strong.

Guess, Nagler & Tucker, Science Advances, 2019About 1,300 opt-in US panellists who installed a data-sharing app, 2016 campaign

Caveat Researchers could see posts but not feeds, so this measures sharing rather than exposure, and the panel was opt-in.

The age gradient sits awkwardly with the framing of misinformation as a young person's problem, and is why checking what a search turns up belongs in adult education as well as school.

As a share of what people consume, it is a rounding error

The strongest corrective to alarmist volume estimates is a proper denominator: false content counted against everything a person reads and watches in a day.

2020

Fake news accounts for only 0.15% of Americans' daily media diet, and news of all kinds accounts for at most 14.2%.

Direction: No detectable effect. Strength of evidence: Strong.

Allen, Howland, Mobius, Rothschild & Watts, Science Advances, 2020US panels measured across mobile, desktop and television

Caveat A time-share denominator makes any category look small, and the panels predate short-video feeds.

That denominator includes daytime television, which flattens everything. The figure bounds exposure rather than proving harmlessness.

The corrections literature disagrees with itself

Two meta-analyses of comparable quality reach different conclusions, and the difference looks like subject matter rather than method. On political beliefs, fact-checking moves people.

2020

Fact-checking shifts political beliefs by d = 0.29, substantially attenuated by pre-existing belief and ideology and by truth-scale ratings rather than full refutations.

Direction: Decrease. Strength of evidence: Mixed.

Walter, Cohen, Holbert & Morag, Political Communication, 2020Meta-analysis of 30 experiments, N = 20,963

Caveat Almost all included studies measure belief immediately after exposure in a survey experiment, so durability and behavioural effects are untested.

On science-relevant misinformation, a larger synthesis finds nothing distinguishable from zero.

2023

Corrections of science-relevant misinformation were on average not effective, at d = 0.11 with a confidence interval from −0.04 to 0.26 and p = 0.142.

Direction: No detectable effect. Strength of evidence: Strong.

Chan & Albarracín, Nature Human Behaviour, 2023Meta-analysis of 75 reports, 245 effect sizes, N = 53,320

Caveat Corrections did work on non-polarised topics and with audiences familiar with both sides, so the null conceals real heterogeneity.

Both can be true: correction appears to work where a topic is not yet a marker of identity and to fail where it is, a narrower prescription than the fact-checking industry implies.

Inoculation raises scepticism more reliably than accuracy

Prebunking — showing people manipulation techniques before they meet them — has more industrial backing than any rival, and its field trial was funded by an interested party.

2022

Prebunking videos improved recognition of manipulation techniques by around 5% among those answering a single follow-up question about 18 hours later.

Direction: Increase. Strength of evidence: Mixed.

Roozenbeek, van der Linden, Lewandowsky et al., Science Advances, 2022About 5.4 million YouTube users, 22,632 answering the follow-up item

Caveat Funded by Google Jigsaw, which had a commercial interest in the outcome; the field measure is a single multiple-choice item answered by under 0.5% of viewers.

Independent replication of the best-known inoculation game found worse than a null: people became more sceptical of everything.

2023

The "Bad News" game did not improve discrimination between true and false headlines, with a Bayes factor of 8.43 favouring no effect, and made participants rate true headlines as false almost as much more often (d = 0.51) as false ones (d = 0.58).

Direction: No detectable effect. Strength of evidence: Strong.

Graham, Skov, Gilson, Heise, Fallow, Mah & Lindsay, Journal of Cognition, 2023Four independent replications, N = 353

Caveat The stimulus set was unbalanced at 8 true against 24 false items, which the authors note may itself encourage blanket scepticism.

Raising distrust of accurate reporting is not obviously an improvement. The Reuters Institute put trust in news across 48 markets at 37% in its 2026 Digital News Report, its lowest since 2015.

The AI wave has been counted more than it has been measured

The first systematic look at real elections found less synthetic content than expected: a search of the 2024 UK, French and EU votes identified 27 viral cases of AI-generated disinformation and no evidence that any of it measurably affected a result (Stockwell, CETaS / The Alan Turing Institute, 2024). It captured only content above a virality threshold, and the harms it did document — deepfake harassment of female politicians, death threats — were real whatever the vote effects. Site counts come mostly from vendors: NewsGuard reported 3,749 AI content farm news sites across 16 languages as of 23 June 2026, against 49 when it began tracking in May 2023, but it sells detection products, does not publish the domain list, and a count of sites says nothing about readership — the same caution that applies to the vendor figures behind what people know about online privacy.

The short version

  • About 1% of Twitter users took 80% of exposures to fake news sources in 2016, and 0.1% did 81% of the sharing (Grinberg et al., 2019).
  • Sharing is uncommon and strongly age-graded, from 3% of 18–29s to 11% of over-65s on Facebook (Guess et al., 2019).
  • Fake news comes to 0.15% of the American daily media diet across all screens (Allen et al., 2020).
  • Corrections move political beliefs at d = 0.29 but average zero on science topics (Walter, 2020; Chan & Albarracín, 2023).
  • Nearly all of it rests on 2016-era platforms and same-day survey outcomes, so almost nothing is known about short-video feeds or about effects lasting months.