Confronting Misinformation Produced with Generative AI: A Meta-Analysis of Experimental Scientific Evidence

A high-level précis of this Synthesis Review can be found in the Summary for Policymakers report, "Responding to Generative AI Misinformation: Results from a Meta-Analysis of Scientific Evidence."

This report synthesizes experimental evidence on the effects of generative artificial intelligence (GenAI) misinformation on individuals’ ability to evaluate information and the effectiveness of interventions designed to mitigate its influence. It draws on a meta-analysis of experimental studies (60 effects from 24 publications and 33,801 participants) published between 2018 and 2025.

The findings indicate a divergence in public responses to different modalities of GenAI-produced misinformation (referred to in the report as GenAI misinformation), alongside substantial heterogeneity across studies. More recent research tends to report higher perceived accuracy and credibility of textual GenAI misinformation, whereas individuals’ perceptions of visual GenAI misinformation—primarily deepfakes—have become more skeptical over time. These patterns are associated with the time of data collection rather than model capacity and are accompanied by wide variation in effect sizes across contexts.

The report also evaluates two countermeasures: preventive corrective information and content labeling. In studies conducted after 2020, corrective information yields small-to-moderate (0.17 to 0.43 standard deviation units) reductions in the perceived accuracy or credibility of GenAI misinformation. Pooled estimates in this subset are significant and comparatively stable, though prediction intervals show substantial variation, especially when outcomes focus on detection rather than evaluative judgment.

Content labeling interventions vary more than corrective information. Althoughlabels are associated with modest average reductions in perceived credibility, effect sizes vary widely across modalities, designs, presentations, and contexts,and prediction intervals often include null effects.

Key conclusions from the synthesis are as follows:

1. Textual GenAI misinformation may currently pose greater persuasive risks than visual GenAI misinformation.
2. Preventive corrective information appears to be more a consistently effective countermeasure.
3. Content labeling policies yield highly variable outcomes and require careful design and evaluation.
4. The available evidence base is geographically narrow and methodologically constrained, with most studies focusing on populations in Western Europe and North America

‍

Based on these findings, policy responses should consider the following:

• Prioritize addressing text-based GenAI misinformation.
• Treat preventive corrections as a core intervention.
• Approach content labeling cautiously

‍

ISBN: 978-3-03983-015-2

DOI: 10.61452/UGTR3022

Citation: International Panel on the Information Environment [A. Herasimenka, S. Valenzuela, S. Boulianne, F. Esser, L. M. Given, S. Lewandowsky, E.M. NavarroLópez, P.N. Howard (eds.)], “Confronting Misinformation Produced with Generative AI: A Meta-Analysis of Experimental Scientific Evidence,” Zurich, Switzerland: IPIE, 2026. Synthesis Report, SR2026.2, doi: 10.61452/UGTR3022.

Appendix: Download the Online Supplemental Appendices PDF.

‍