Prebunking reached 375,597 Instagram users. Whether it lasted five months rests on a poll under 1% answered.

The question

Psychological inoculation, or "prebunking," is one of the most studied ideas in misinformation research: show people a weakened example of a manipulation technique before they meet the real thing, and they should recognise it later. The evidence base is substantial, but almost all of it comes from the laboratory or from controlled online panels. The open question is whether the effect survives contact with an actual social media feed, where people scroll fast, attention is scarce, and no researcher controls what happens next.

A new field study puts that question to a real platform. Sander van der Linden, Jon Roozenbeek and colleagues, working with Google Jigsaw and the NGO Reality Team, ran a 19-second prebunking video as a Story Feed advertisement to 375,597 Instagram users in the United Kingdom, then measured whether the treated group could better spot emotional manipulation in a headline. It is a genuine step out of the lab. It is also a study whose strongest claims rest on a much smaller and more fragile foundation than its reach suggests, and the distance between those two things is the interesting part.

The paper appeared in the Harvard Kennedy School Misinformation Review, a peer-reviewed journal rated Q1 on SCImago for Social Sciences (miscellaneous) in 2025.

What they did

The team targeted users aged 18 to 34 over six days in February 2025 with a short video that forewarned them about fearmongering, using a deliberately weak example: a fake headline reading "Yoga linked to terrifying full body cancer." Users who watched at least half the video became eligible for a follow-up poll delivered through Instagram's own poll-sticker feature, a single binary question asking which manipulation technique a headline used.

Two design choices matter for reading the results. First, this was not a randomised trial. True random assignment is not possible inside Instagram's ad manager, so the researchers approximated it by assigning users to the control group based on their self-declared birth month, with April, July and October sent to control. Second, the control group by design did not watch the video. The comparison is therefore between people who watched an inoculation video, restricted to those who watched at least half of it, and people who did not.

What they found

On the immediate poll, 59.55% of treated users correctly identified the manipulation technique against 38.21% of controls, a lift of 21.4 percentage points (95% CI 14.6 to 28.0). It is a statistically significant effect of medium size. Five months later a second poll found the gap essentially unchanged, 66.39% against 43.98%, a lift of 22.21 points. Treated users were also about three times more likely to click a link to "learn more," though in absolute terms that meant 0.31% of them clicking against 0.11% of controls.

Read quickly, the story is clean: a cheap video, deployed at scale, improved manipulation recognition, and the improvement held. The cost claim is well supported, at roughly $825 per 100,000 impressions. The rest deserves a slower reading.

What the design does not establish

The reach figure, 375,597 users, is not the number the findings rest on. The immediate result comes from 806 poll responses, 403 in each group. Those responses represent about 0.85% of the treatment watchlist and 0.34% of the control pool. The outcome measure, in other words, is built from the under-1% of a self-selected audience who both watched the video and then chose to answer a poll. People who watch an inoculation video to the halfway mark and then stop to answer a quiz about manipulation are plausibly not a random slice of the 375,597; they are the more engaged and more motivated ones. The authors say so directly.

The five-month result needs the most care, because it is the one most likely to be repeated as "the effect lasts." The follow-up was not the same people re-tested. It was a fresh cross-section, and the authors state plainly that they do not know whether any original respondents are in it. Attrition also ran in opposite directions between the groups: treatment responses fell from 403 to 244 while control responses rose from 403 to 432. So "stable for five months" is a statement about two group averages measured on different samples five months apart, not evidence that inoculated individuals retained anything. That is a materially weaker claim than durability, and the abstract's phrase "effects persisting for five months" blurs the two.

Then there is what the outcome actually measures. Instagram's poll sticker allows one binary question, so recognition was tested with a single item, two options, on one manipulation technique. Chance performance is 50%, which makes the control group's 38% baseline itself striking and also limits how much a single item can carry. Identifying a technique in a quiz is, as the authors note, only the first step toward resisting misleading content. The study has no data on whether anyone rated real posts as less credible, shared less, or changed any behaviour beyond one link click.

Provenance is worth stating rather than implying. The study was funded by Google Jigsaw, which is also the main institutional promoter of prebunking, three authors are employed by the partner NGO, and two received Jigsaw funding. This does not invalidate anything; the materials are on the Harvard Dataverse and the analysis is transparent. But when an intervention's principal advocate funds the field test of that intervention, independent replication carries more weight than usual.

Bizarus interpretation

The genuine contribution here is methodological, not the effect size. Getting an inoculation intervention onto a live commercial feed and evaluating it with the platform's own polling tools, cheaply, is a real advance over simulated feeds and lab panels, and it is the part other groups can build on. The recognition effect is consistent with a large existing literature, so it is not surprising, and its size on a two-option item is hard to translate into anything about real-world resistance.

What the study does not yet show is the thing people most want it to: that a short video meaningfully protects ordinary users, over months, against misinformation they actually meet. Recognition in a quiz answered by the most engaged 1% is a long way from that. This is a recurring shape in intervention research worth internalising. There is a large denominator in the recruitment funnel and a small, self-selected one in the outcome, and the impressive number belongs to the former while the causal claim rests on the latter.

What remains unanswered

Does the effect appear when the outcome is resistance rather than recognition, measured on more than one item and one technique? Does it hold for the disengaged majority who scroll past, and not only the minority who stop to watch and answer? And does it replicate when someone other than the intervention's funder runs the test? Until then, the honest summary is that prebunking can be deployed at scale cheaply, that it moves a recognition measure among the engaged, and that "it lasts five months" is not something this particular design can support.

Source

van der Linden, S., Louison-Lavoy, D., Blazer, N., Noble, N., & Roozenbeek, J. (2026). Prebunking misinformation techniques in social media feeds: Results from an Instagram field study. Harvard Kennedy School (HKS) Misinformation Review, 7(1). https://doi.org/10.37016/mr-2020-193. Journal quartile verified Q1, SCImago Social Sciences (miscellaneous), 2025. Study data: Harvard Dataverse, https://doi.org/10.7910/DVN/YHXOUK.