Abstract
This study investigates the impact of warning labels on the sharing of political videos, particularly in the context of potential deepfakes. Using an online experiment with 578 US participants, the research compares "prebunking" (warnings before viewing) with "debunking" (warnings after the decision to share). Results indicate that prebunking significantly reduces sharing probability, while debunking is largely ineffective once a user has committed to an action. The findings highlight the importance of intervention timing in digital misinformation management.
Introduction
As AI-generated deepfakes become increasingly sophisticated, the line between reality and fabrication in political discourse is blurring. This phenomenon creates a "liar’s dividend," where authentic footage can be dismissed as fake, and a "misinformation epidemic," where fabricated content is shared as truth. While previous research has demonstrated that warnings can help individuals identify fake content, their impact on actual behavior—specifically the act of sharing—remains less understood. This study explores whether the timing of a warning (prebunking vs. debunking) influences the decision to share potentially manipulated political content.
Methodology
The researchers conducted an online experiment with 578 US participants (267 female, mean age 43.59), recruited via Prolific and balanced between Democrats and Republicans. Participants were shown high-quality, authentic videos of their respective party leaders (Joe Biden for Democrats, Donald Trump for Republicans). Although the videos were real, they were presented in a context where they could be perceived as deepfakes.
The study employed a 1-factor, 3-level between-subjects design:
- Control: No warning provided.
- Prebunking: A warning ("This video may contain altered content") appeared during video playback.
- Debunking: The warning appeared only after the participant clicked a button indicating an intent to share.
Data were analyzed using a Generalized Linear Mixed Model (GLMM), controlling for partisanship and treating specific videos as random effects.
Results
The timing of the warning proved critical to its effectiveness. Prebunking showed a statistically significant reduction in sharing behavior compared to the control group (Odds Ratio [OR] = 0.17; p < .001). In contrast, debunking—issued after the user had already decided to share—showed no significant difference from the control group.
As shown in Figure 1, the probability of sharing was also influenced by the user's perceived accuracy of the video. For participants who judged the video as "fabricated," the prebunking warning was particularly powerful, significantly lowering the probability of sharing compared to those who were "unsure" (OR = 3.72; p = .027). Interestingly, because the study used real videos, the results also highlight a "skepticism trade-off," where warnings can suppress the sharing of legitimate information if users are primed to be overly skeptical.
Discussion
The findings suggest a "commitment trap": once a user has psychologically committed to sharing a video, a subsequent warning is often ignored. Prebunking is more effective because it sets a skeptical frame before the emotional or cognitive impact of the video takes full effect. However, the suppression of real videos suggests that platforms must balance misinformation mitigation with the risk of delegitimizing authentic discourse.
Future research should include actual deepfakes to determine if warnings can help users distinguish between truth and lies, rather than simply suppressing sharing across the board. Additionally, exploring the long-term effects of "warning fatigue" will be essential for practical implementation on social media platforms.
References
Chen, J., Sugano, B. T., Frik, A., Nguyen, H. H., Igarashi, T., Echizen, I., & Sasahara, K. (2024). The effects of warning messages on political video sharing behaviors. Proceedings of the 88th Annual Convention of the Japanese Psychological Association.
Summary updated