When Britons encountered misinformation, fact-checks, prebunking messages, or fake headlines, did our synthetic counterparts respond the same way? Matching a survey’s headline numbers is relatively easy; reproducing an intervention’s effect is harder. We took four published UK experiments, sealed our design, simulated them on our UK General Population society, and compared the results.
1. Misinformation Erodes Vaccine Intent
In 2020, Loomba and colleagues showed 3,000 Britons five social media posts spreading misinformation about COVID-19 vaccines. The share saying they would definitely accept a vaccine fell by 6.6 percentage points in the released data and 8.2 in our synthetic population, comfortably within the authors’ reported confidence interval of 3.8–9.0 points.
The simulation also captured how misinformation works: it rarely converts supporters into opponents, but nudges them into doubt. Among Britons initially saying “Yes, definitely”, 81% stayed there, 14% softened to “unsure but leaning yes”, and 4% went further. The synthetic proportions were 82%, 14%, and 4%.
The five posts were not equally damaging. A claim about vaccine-trial monkeys did the most harm; a Bill Gates conspiracy poster was among the least damaging. The synthetic public ranked all five in the same order, although the bottom three were closely grouped in both populations.
There was a mismatch in the control group. Factual posts produced no significant shift among 1,000 Britons, while our synthetic group moved about four points towards vaccination, suggesting a small bias towards vaccine-positive content.
Firm vaccine supporters moved alike after 5 misinformation posts
Bar chart. Of those who began “Yes, definitely”, 81.4% of human and 82.2% of synthetic respondents stayed there; 14.3% and 13.7% moved to “Unsure, lean yes”; 3.0% and 2.8% to “Unsure, lean no”; 1.3% and 1.4% to “No, definitely not”.
Which posts did most damage: identical rank order in both populations
Dot plot of each post’s mean “less inclined to vaccinate” score, on a scale of −2 to +2, for human and synthetic respondents. Both populations rank the vaccine trial monkeys post first (0.41 human, 0.40 synthetic), then mRNA vaccine “alters your DNA” (0.34, 0.38), a whistleblower’s “97% will become infertile” (0.23, 0.33), “99.6% survival rate, why a vaccine?” (0.21, 0.26), and the Bill Gates “film” poster last (0.20, 0.15).
2. Fact-Checks Work, Then Fade
Carey and colleagues (2022) followed a YouGov panel across three waves. Half read fact-checks of four false COVID-19 claims; half read unrelated articles. Belief in the claims fell by 45% within the human fact-check group, calculated from the replication data, and 43% in ours. Against the control group, the paper reported a 45% reduction; ours was 46%.
Belief in every claim fell in both populations. The claim that China built the virus as a bioweapon, the most believed in both, fell by 39% among Britons and 37% in our panel. The hydroxychloroquine fact-check was among the strongest in both.
The largest proportional mismatch concerned a claim that a Bill Gates-funded group patented the virus: belief fell by 15% among Britons and 59% among synthetic respondents. But hardly anyone believed it initially. Three quarters of Britons and nine in ten synthetic respondents already called it “not at all accurate”. The average shifts were about a twentieth and an eighth of a point on a four-point scale. With such low starting belief, small movements produce large percentages. Synthetic respondents’ explanations confirmed that the fact-check mostly reinforced an existing rejection.
Every targeted claim fell after a fact-check, in both panels
Bar chart of the fall in belief after the fact-check, human panel then synthetic panel. Gates-funded group patented the virus, −15% and −59%; China created it as a bioweapon, −39% and −37%; hydroxychloroquine cures or prevents it, −61% and −54%; antibiotics prevent and treat it, −49% and −44%.
Two months later, the human effect was no longer distinguishable from zero and, by our calculation, had shrunk to 8% of its original size. Synthetic respondents do not live through the weeks between surveys, so we modelled time explicitly. At the third wave, each respondent either recalled the correction and earlier answers or answered afresh, using a recall rate drawn from Guess and colleagues (2020).
The synthetic effect also shrank to 8%. This match depends on the imported forgetting rate: a published estimate of how corrections fade, applied to our panel, reproduced Carey’s result. That makes campaign timing, alongside content, something a synthetic public can help explore.
3. Reproducing Prebunking’s Small Effects
Prebunking exposes people to manipulation techniques, with a warning, so they recognise them later. Basol and colleagues (2021) tested this with British respondents who read UNESCO’s #ThinkBeforeSharing infographics or, in the control group, played Tetris. Participants rated social media posts for manipulativeness and willingness to share.
Recomputed for the 498 UK respondents in these two groups, the infographics increased perceived manipulation in fake posts by 0.21 standard deviations and reduced willingness to share by 0.12, the latter not statistically significant. Our population produced effects of 0.21 and 0.10, respectively. These were small signals, closely reproduced.
A week later, neither population was reliably more sceptical of new fake posts. The human effect was 0.08 and not significant; the synthetic effect was smaller still.
The infographics produced the same 3 effects in both populations
Bar chart of the effect relative to control in Cohen’s d, human then synthetic respondents. Fake posts rated more manipulative, +0.21 and +0.21; less willing to share fake posts, −0.12 and −0.10; still more sceptical one week later (new fake posts), +0.08 and +0.03.
Even before treatment, the populations agreed on which of the 18 posts looked manipulative: rank correlation was 0.87. Both placed conspiracy posts near the top and fake-expert posts lower down.
9 fake posts, judged alike by both populations
Scatter plot of each fake post’s mean manipulativeness rating on a scale of 1 to 7, human on the horizontal axis and synthetic on the vertical, with a dashed line marking equal ratings. The three conspiracy posts (microchips in the vaccine, mRNA “alters your DNA”, and testing as “DNA collection”) sit at the top right, near the line. The fake-expert posts sit lower: the Nobel laureate’s “virus is manmade” furthest below the line, the immunologist’s 20-second breath test furthest above it.
4. Which Fake Headlines Fool Britons?
The Misinformation Susceptibility Test (MIST-20), developed by Cambridge researchers and co-authored by our CEO James He, asks respondents to classify ten real and ten fake headlines. In a July 2020 UK quota sample of 1,227 people, the fake headlines that fooled the most Britons also fooled the most synthetic respondents: rank correlation was 0.87. Britons correctly rejected 7.4 fake headlines on average; our panel rejected 7.9.
The biggest gap was “Left-Wing Extremism Causes ‘More Damage’ to World Than Terrorism, Says UN Report”. Four in ten Britons believed it, compared with two in ten synthetic respondents. Our platform’s comment analysis suggested why: synthetic respondents saw its sensational phrasing as inconsistent with the UN’s usual language.
Both populations showed a political gradient, with respondents on the right less likely to spot the fake, but the synthetic gap was narrower. Here, greater institutional literacy protected the synthetic public from a headline that fooled many Britons.
Which fake headlines fool Britain: the same ones fool the synthetic panel
Dot plot of the share of human and synthetic respondents correctly calling each of ten fake headlines fake. Both populations reject the weather manipulation, Ebola, and disease-spreading headlines most often, above 84%. The widest gap is the left-wing extremism headline, rejected by 61% of human and 80% of synthetic respondents; the stock-price headline is the one humans reject least often, at 59%.
5. Mapping Susceptibility Across the Country
Loomba and colleagues also measured misinformation susceptibility among 16,477 people across the UK, finding that regional ability to spot fake news predicted vaccine uptake. We compared their regional estimates with our synthetic map, reweighting respondents to regional age, sex, graduate, and ethnic-minority profiles. Scotland was excluded from this analysis.
The broad patterns agreed, with a rank correlation of 0.80. The South East was the strongest English region in both; the West Midlands, North East, and Northern Ireland were below average in both.
Veracity discernment across the UK
MIST-20 score by region: headlines correctly identified, out of twenty, above or below each population’s UK average.
Two maps of England, Wales, and Northern Ireland and a dot plot: each region’s MIST-20 score, in headlines correctly identified out of twenty, above or below its population’s UK average, for human respondents in 2021 and for synthetic respondents. The South East is highest in both, +0.4 and +0.3. The West Midlands, North East, and Northern Ireland are below average in both. The widest gap is London: −0.6 among human respondents, 0.0 among synthetic ones.
What This Makes Possible
Across four published experiments and evidence from more than 25,000 Britons, our UK society reproduced the human direction, ordering, and, on operationally important effects, size of response. Where the simulation fell short, we could identify the gaps and often explain them through respondents’ own words.1 (footnote) A validated synthetic society can tell you in advance which narratives will gain purchase, how much of the damage a fact-check will recover and how quickly that recovery fades, how large an effect a prebunking message should be expected to have, and where in the country vulnerability runs higher.
Provenance
The design, instruments, code, and panel were hashed and time-stamped before the first survey call.
Notes
- 1.
Indeed, our society represented the UK in 2026 and believed false claims somewhat less than the 2020 samples, so we focused on relative differences rather than absolute rates.
References
Basol, M., Roozenbeek, J., Berriche, M., Uenal, F., McClanahan, W. P., & Linden, S. V. D. (2021). Towards psychological herd immunity: Cross-cultural evidence for two prebunking interventions against COVID-19 misinformation. Big Data & Society, 8(1), 20539517211013868.
Carey, J. M., Guess, A. M., Loewen, P. J., Merkley, E., Nyhan, B., Phillips, J. B., & Reifler, J. (2022). The ephemeral effects of fact-checks on COVID-19 misperceptions in the United States, Great Britain and Canada. Nature Human Behaviour, 6(2), 236–243.
Guess, A. M., Lerner, M., Lyons, B., Montgomery, J. M., Nyhan, B., Reifler, J., & Sircar, N. (2020). A digital media literacy intervention increases discernment between mainstream and false news in the United States and India. Proceedings of the National Academy of Sciences, 117(27), 15536–15545.
Loomba, S., De Figueiredo, A., Piatek, S. J., De Graaf, K., & Larson, H. J. (2021). Measuring the impact of COVID-19 vaccine misinformation on vaccination intent in the UK and USA. Nature Human Behaviour, 5(3), 337–348. (medRxiv: doi.org/10.1101/2020.10.22.20217513)
Loomba, S., Maertens, R., Roozenbeek, J., Götz, F. M., Van Der Linden, S., & De Figueiredo, A. (2023). Ability to detect fake news predicts sub-national variation in COVID-19 vaccine uptake across the UK. medRxiv.
Maertens, R., Götz, F. M., Golino, H. F., Roozenbeek, J., Schneider, C. R., Kyrychenko, Y., Kerr, J. R., Stieger, S., McClanahan, W. P., Drabot, K., He, J., & Van Der Linden, S. (2023). The Misinformation Susceptibility Test (MIST): A psychometrically validated measure of news veracity discernment. Behavior Research Methods, 56(3), 1863–1899.