Prebunking misinformation techniques in social media feeds: Results from an Instagram field study

Summary

This research note reports a field experiment testing whether psychological inoculation (“prebunking”) against emotional manipulation can be deployed at scale within Instagram’s native ad environment. Partnering 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 measured manipulation-recognition via the platform’s poll sticker on a fictitious fearmongering headline. Treated users were 21 percentage points more accurate than controls at spotting emotional manipulation, the effect held at a five-month follow-up, and treated users clicked through to learn more at roughly three times the control rate — establishing that brief, cheap, in-feed inoculation ads can produce durable and behaviorally consequential gains outside the lab.

Key Contributions

• One of the first real-world demonstrations of inoculation efficacy inside the Instagram scroll feed, extending prior YouTube-based prebunking field trials.
• A novel quasi-experimental methodology that exploits Instagram’s poll sticker (for outcome measurement) and birth-month targeting (as a control-assignment workaround for the absence of native randomization in the ad manager).
• Evidence of unusually durable single-technique inoculation effects, with the lift undiminished at five months.
• Evidence that prebunking shifts not just recognition but downstream behavior (click-through), suggesting genuine engagement rather than mere demand effects.
• Practical cost benchmarks ($8.25 CPM, below Meta averages) and operational guidance for NGOs, platforms, and policymakers running scaled prebunking campaigns.

Methods

A field experiment ran February 5–11, 2025, targeting UK Instagram users aged 18–34. A weakened-dose fearmongering example (“Yoga linked to terrifying full body cancer”) anchored a 19-second video produced after A/B pilot testing of color, tone, length, and music. Because the Instagram ad platform does not permit randomization, control assignment was based on birth months (April, July, October). Treatment was “as-treated”: users had to watch ≥50% of the video to enter the watchlist. Outcomes were captured via a binary Instagram poll sticker asking users to identify the manipulation technique used in a headline, with the correct answer placed on the left to require intentional swiping. Initial responses (n = 806) were collected 24 hours to 10 days post-exposure; a five-month follow-up poll (July 2025) collected n = 676 further responses. Analyses used two-proportion z-tests and chi-square tests with Cohen’s h effect sizes.

Findings

• Treated users identified fearmongering correctly 59.55% of the time vs. 38.21% for controls — a 21.4 pp lift (χ²(1) = 35.87, p < .001, h = 0.43).
• Control baseline accuracy fell below chance (38%), indicating poor default sensitivity to emotional manipulation.
• At five months, the gap was effectively unchanged: 66.39% treatment vs. 43.98% control (χ²(1) = 31.38, p < .001, h = 0.45).
• Click-through to “learn more” was ~3× higher in treatment (0.31% vs. 0.11%), persisting at follow-up (0.274% vs. 0.101%).
• Treatment-to-watchlist conversion was 12.05%; poll response rates were ~0.85% (treatment) and 0.34% (control).
• Cost per thousand impressions was $8.25, below platform averages, supporting low-cost scalability.

Connections

This study extends the inoculation/prebunking literature into Instagram’s native ad environment and complements platform-scale work on the dynamics of misinformation exposure such as Gonzalez-Bailon2024-rq and Mosleh2024-op, as well as broader assessments of the misinformation problem space like Budak2024-ef. Its emphasis on resilience-building also speaks to comparative work on societal vulnerability to misinformation, e.g. Humprecht2025-ml. The use of an ad platform as both delivery and measurement infrastructure connects methodologically to other field-experimental approaches in the topic cluster, though most adjacent papers here focus on misinformation supply, narratives, or detection rather than user-side intervention.

Podcast

A research-radio episode discusses this paper: Listen