Beyond textual disinformation: Comparing the effects of textual disinformation to AI-generated and video-based visual disinformation across different issues

Summary

This pre-registered experiment (N=982, U.S.) tests whether visual disinformation is more persuasive than textual disinformation, comparing three modalities — text, AI-generated still images, and decontextualized real videos — across two issues (the MH370 disappearance and the Russian invasion of Ukraine). Contrary to widespread assumptions that “seeing is believing” and that generative AI poses unprecedented persuasive threats, Hameleers and van der Meer find that visual disinformation does not uniformly outperform text. Only for the salient, polarized Ukraine issue did decontextualized video prove more credible and engaging than text, and AI-generated imagery rarely outperformed either. Fact-checks reduced credibility roughly equally across all modalities.

Key Contributions

• One of the first direct experimental comparisons of textual, AI-image, and decontextualized-video disinformation within a single design.
• Empirically tempers alarmist narratives about generative AI by showing that low-tech video decontextualization can be more persuasive than AI-generated stills.
• Identifies issue salience, polarization, and availability of plausible authentic footage as boundary conditions for visual disinformation effects.
• Shows fact-checks work across modalities, suggesting correction strategies need not be modality-specific (though explaining visual deception techniques remains useful).
• Clarifies moderating roles of conspiracy mentality, media distrust, and self-perceived media literacy.

Methods

Pre-registered between-subjects online experiment with a 3 (modality: text / AI image / decontextualized video) × 2 (correction vs. none) + 3 control design, with topic (MH370 vs. Ukraine) varied within-subjects in randomized order. N=982 U.S. adults from a Kantar opt-in panel (Dec 5–11, 2023). Stimuli were mocked-up news articles: a false missile-strike claim for MH370 and a fabricated Putin–Zelenskyy peace deal for Ukraine. DVs were credibility, engagement intentions, and discrete emotions on 7-point scales; moderators included media distrust, conspiracy mentality (Bruder et al.), and self-perceived media literacy. Analyses used ANOVAs, OLS regressions with interactions, and Wald coefficient comparisons.

Findings

• For MH370, no modality differences in credibility, emotion, or engagement; all disinformation was rated less credible than authentic control content.
• For Ukraine, decontextualized video was significantly more credible and engaging than textual disinformation; differences from AI imagery were smaller or marginal.
• Emotional responses did not differ by modality for either issue.
• Conspiracy mentality amplified credibility of MH370 disinformation across modalities but not for Ukraine.
• Media distrust modestly boosted credibility of textual MH370 disinformation.
• Self-perceived media literacy had little moderating effect, except slightly reducing engagement with textual MH370 disinformation.
• Fact-checks reduced credibility for both issues, with no significant modality interaction.
• Exploratory: right-leaning participants showed less emotional response and lower belief in some messages; issue-specific partisan beliefs increased credibility.

Connections

This study speaks directly to ongoing debates about generative-AI-driven disinformation and visual manipulation, complementing concerns raised in Spampatti2026-kx and the broader resilience framework developed by Humprecht2025-ml. Its finding that fact-checks work across modalities resonates with prebunking and correction work such as van-der-Linden2026-jt and DeVerna2025-dl, while its tempering of AI-alarmism aligns with the prevalence-focused critiques in Budak2024-ef and Simon-style arguments echoed across Starbird2025-jj and Marwick2025-ov. The role of conspiracy mentality as a moderator connects to audience-susceptibility work like Frischlich2025-vn.

Podcast

A research-radio episode discusses this paper: Listen