Prebunking vs debunking: what the experiments show

Debunking is a correction issued after a false claim has circulated: a fact-check names the claim and gives the evidence against it, one claim at a time. Prebunking comes first. It shows people a weakened example of a manipulation technique, such as a fake expert or a conspiracy frame, together with a warning, so that they recognise the technique when they meet it later. Roozenbeek, Basol and van der Linden, writing in Behavioral Scientist in February 2021 (opens in a new tab), ground the approach in inoculation theory: as a weakened pathogen prompts the body to make antibodies, a weakened persuasive argument builds resistance to later manipulation.

Debunking needs a claim to answer and a channel back to the people who saw it. Prebunking needs a forecast of the techniques an audience will meet, and because it targets the technique, one message can in principle cover claims that have not appeared yet; in the replication below, though, neither population was reliably more sceptical of new fake posts a week later. The table compares two of the four UK experiments that Artificial Societies replicated.

Source: Artificial Societies, We Accurately Simulated Misinformation, Fact-Checking, and Prebunking Experiments, 23 September 2026. Artificial Societies recomputed some human figures from the original data.

A fact-check reduced belief in a false claim by just under half in the UK experiment by Carey and colleagues (2022), and Artificial Societies’ synthetic panel reproduced the size of that fall. Carey’s team followed a YouGov panel across three waves: half read fact-checks of four false COVID-19 claims and half read unrelated articles. Against the control group, the paper reported a 45% reduction in belief among Britons; Artificial Societies’ synthetic panel produced 46%.

Belief in every targeted claim fell in both populations; the most believed, that China built the virus as a bioweapon, fell by 39% among Britons and 37% in the synthetic panel. The paper adds that the fact-checks reduced belief especially among the groups most vulnerable to the claims, and that they had minimal spillover onto the accuracy of related beliefs. A correction works on the claim it names, and little else.

The largest proportional mismatch concerned the claim that a Bill Gates-funded group patented the virus. Belief fell by 15% among Britons and 59% among synthetic respondents, but almost nobody believed the claim at the start: 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, and on a base that small, tiny movements become large percentages. A percentage fall is only readable next to the starting level of belief.

In the one panel that measured it, the effect of a fact-check had almost gone two months later. In Carey and colleagues’ 2022 experiment, the human effect was then no longer distinguishable from zero and, by Artificial Societies’ calculation, had shrunk to 8% of its original size. The paper’s abstract reports that the reductions did not persist even after repeated exposure, and its authors conclude that the effects of fact-checks “are ephemeral”. On that evidence, repeating a correction does not make it last.

Synthetic respondents do not live through the weeks between waves, so we modelled time explicitly. At the third wave, each synthetic respondent either recalled the correction and its earlier answers or answered afresh, with a recall rate drawn from Guess and colleagues (2020). Artificial Societies’ synthetic effect also shrank to 8%. That match depends on the imported forgetting rate, which the simulation was given and did not discover. Where a published decay rate exists, a synthetic public can help a team explore timing: how much of a correction will remain by a vaccination drive or a polling day.

One exposure to prebunking infographics produced a small effect that did not last a week, in real Britons and in Artificial Societies’ synthetic population alike. Basol and colleagues (2021) tested two prebunking interventions against COVID-19 misinformation: Go Viral!, a five-minute browser game, and UNESCO infographics. In the UK data the replication used, British respondents read UNESCO’s #ThinkBeforeSharing infographics or, as a control, played Tetris, and then rated social media posts for manipulativeness and willingness to share.

Recomputed by Artificial Societies for the 498 UK respondents in those two groups, the infographics raised the perceived manipulativeness of fake posts by 0.21 standard deviations and reduced willingness to share by 0.12, the second figure not statistically significant. Artificial Societies’ synthetic population produced 0.21 and 0.10. An effect in standard deviations (Cohen’s d) compares the shift in the average with the usual spread of ratings, so 0.21 is about a fifth of that spread; Artificial Societies’ post calls these 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, and the synthetic effect was 0.03.

In the original paper, the format mattered. The game’s effects on perceived manipulativeness and on confidence remained significant for at least one week, and the authors call the infographics’ effects descriptively smaller than the game’s (Basol et al., 2021). Artificial Societies replicated only the infographic arm, so these results say nothing about the game.

When misinformation moves vaccine supporters, it moves them into doubt rather than opposition, and Artificial Societies’ synthetic population reproduced that pattern in the vaccine experiment by Loomba and colleagues (2021). In that 2020 study, 3,000 Britons saw five social media posts spreading misinformation about COVID-19 vaccines. The share who would definitely accept a vaccine fell by 6.6 percentage points in the data the authors released and by 8.2 in the synthetic population. Among Britons who began at “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 synthetic population also ranked the five posts in the same order of damage as real Britons, although the bottom three sat close together in both. A claim about vaccine-trial monkeys did the most harm and a Bill Gates conspiracy poster among the least, which fits the original paper’s finding that scientific-sounding misinformation was more strongly associated with declines in intent. Factual posts produced no significant shift among 1,000 Britons, while Artificial Societies’ synthetic control group moved about four points towards vaccination, which suggests a small bias towards vaccine-positive content. A team testing a pro-vaccination message should therefore read any simulated gain against the synthetic control group, never against zero.

The same synthetic population judged fake headlines much as Britons did. On the Misinformation Susceptibility Test (MIST-20), co-authored by Artificial Societies’ CEO James He, the headlines that fooled the most Britons in a July 2020 UK sample of 1,227 people also fooled the most synthetic respondents (rank correlation 0.87, where 1 is the same order). Regional scores ranked similarly against Loomba and colleagues’ 2023 estimates (0.80, Scotland excluded).

Artificial Societies’ recommendation, drawn from the Carey and Basol experiments rather than from a test of the choice itself, is to debunk when a specific false claim is already circulating among people you can reach, and to prebunk when you can forecast the technique before the claims arrive. Carey and colleagues’ 2022 fact-checks cut belief by 45%, especially among the groups most vulnerable to the claims, but the effect had almost gone two months later. For a prebunking warning, such as one about fake experts ahead of a vaccination campaign, expect a single round of infographics to move ratings by about a fifth of a standard deviation, as in Artificial Societies’ replication of Basol and colleagues (2021).

Exercise Pegasus already rehearses this. The national pandemic exercise, led by the Department of Health and Social Care and the UK Health Security Agency and scheduled for September to November 2025, includes the objective to “test the strategic response to disinformation and misinformation” (NHS England, 16 July 2025 (opens in a new tab)). An exercise tests the response on paper; our national resilience write-up sets out what the replicated experiments add to that planning.

You can test a correction or a prebunking message on a synthetic population before it reaches the public. Artificial Societies’ replication of four UK experiments, against evidence from more than 25,000 Britons, shows how close that comes, using the published studies’ design:

1. 1.

   Build or select the population; the replication used the UK General Population society, a synthetic model of the British public.

2. 2.

   Seal the design before the first run: the design, instruments, code and panel, hashed and time-stamped.

3. 3.

   Run a treatment arm and a control arm. In the vaccine experiment, the control arm exposed the synthetic drift of about four points towards vaccination, so it is never optional.

4. 4.

   Model time explicitly if durability matters, with a recall rate taken from published evidence.

5. 5.

   Read the reasons: synthetic respondents’ explanations are how the replication accounted for its gaps on the Gates-patent claim and on a fake UN headline.

For a live reputational event, see crisis communication simulation; for comparing drafts of a statement, message testing; for officials and opinion leaders, public affairs research.

A synthetic population reflects the year it represents. Artificial Societies’ UK society represented Britain in 2026 and believed false claims somewhat less than the 2020 samples did, so the replication compared relative differences, not absolute rates of belief. Four in ten Britons believed a fake headline about a UN report on left-wing extremism, against two in ten synthetic respondents, whose comments read its sensational phrasing as unlike the UN’s usual language; the synthetic political gradient on fake headlines was also narrower. In Loomba and colleagues’ 2023 regional data, real Londoners scored about half a headline below the national average on MIST-20, where synthetic Londoners scored at it.

Durability is the part the simulation could not supply for itself. In the fact-check experiment, the fade matched because a published forgetting rate was imported, so a new kind of intervention, with no published estimate of its decay, has no rate to import. A synthetic public that has never lived through the two months between waves can reproduce a fade it is given, and it cannot yet be relied on to find one nobody has measured.