van der Linden, S., Louison-Lavoy, D., Blazer, N., Noble, N. S., & Roozenbeek, J. (2026). Prebunking misinformation techniques in social media feeds: Results from an Instagram field study. Harvard Kennedy School Misinformation Review. https://doi.org/10.37016/mr-2020-193
Summary
This research note reports a field experiment testing whether psychological inoculation (“prebunking”) against emotional manipulation can be scaled in a real social media scroll feed. Working with Google Jigsaw and Reality Team, the authors served a 19-second prebunking video as a Story Feed ad to 375,597 UK Instagram users aged 18–34, then used the platform’s own poll sticker to measure whether treated users could better spot fearmongering in a fictitious headline. The study argues that inoculation efficacy—well established in lab and YouTube settings—transfers to the ecologically valid, high-scroll environment of Instagram, and that such campaigns can be deployed cheaply and at scale. Treated users were roughly 21 percentage points more accurate at identifying emotional manipulation, an effect that persisted at a five-month follow-up, and they clicked through to learn more at about three times the control rate.
Key Contributions
- One of the first real-world field demonstrations of inoculation efficacy in an Instagram scroll-feed environment, extending prior YouTube-based work.
- An innovative quasi-experimental methodology using Instagram’s poll sticker and birth-month targeting to evaluate interventions where true randomization is unavailable.
- Evidence of unusually long-lasting (five-month) inoculation effects for a single-technique prebunking message, informing the literature on inoculation decay.
- Demonstration that prebunking shifts digital behavior (click-through), not just recognition, suggesting downstream behavioral relevance.
- Practical cost benchmarks and guidance for policymakers, NGOs, and platforms aiming to scale prebunking.
Methods
- Field experiment on Instagram targeting UK users aged 18–34, running a 19-second prebunking video using a weakened-dose fearmongering example (“Yoga linked to terrifying full body cancer”) as a Story Feed ad (Feb 5–11, 2025).
- Quasi-experimental assignment: control group defined by birth month (April, July, October), since the Instagram ad manager does not permit true randomization.
- “As-treated” design: users had to watch ≥50% of the video to enter the treatment watchlist.
- Outcome measured via a single binary Instagram Poll Sticker asking users to identify the manipulation technique, with the correct answer placed left to require intentional swiping.
- Initial poll responses (n = 806) collected 24 hours to 10 days post-exposure; a follow-up poll was run five months later (July 2025).
- Pilot A/B tests optimized video hook/hold rates; analysis used two-proportion z-tests / chi-square tests with Cohen’s h and 95% CIs.
Findings
- Treatment group identified emotional manipulation 59.55% of the time vs. 38.21% for controls—a 21.4 pp lift (X²(1) = 35.87, p < .001, h = 0.43).
- Control baseline recognition (38%) was below chance, indicating poor default ability to detect fearmongering.
- At five months, treatment accuracy remained higher (66.39% vs. 43.98%; +22.21 pp; X²(1) = 31.38, p < .001, h = 0.45).
- Click-through rate was roughly triple in the treatment group (0.31% vs. 0.11%; X²(1) = 84.61, p < .001), persisting at follow-up.
- Treatment-to-watchlist conversion was 12.05%; poll response rate was about 1%.
- Cost-per-thousand impressions of $8.25 was below Meta averages, demonstrating low-cost scalability.
Connections
This paper sits within the message-intervention strand of the misinformation literature and speaks directly to work on AI- and dialogue-based debunking such as Costello2024-bg, which similarly tests scalable, real-world corrective interventions. Its concern with emotional-manipulation susceptibility and platform-level countermeasures also connects it to studies of misinformation diffusion and platform dynamics like DeVerna2025-dl and Pierri2025-hm. The remaining papers under these topics address distinct facets of information disorder without a direct methodological or theoretical tie to prebunking.
Podcast
A research-radio episode discusses this paper: 🎧 MP3 · Spotify · Apple Podcasts