The Mitigation Paradox: How AI-Disinformation Experts Converge on Threats but Polarize on Solutions

Abstract

Policy on AI-generated disinformation is being shaped before the expert community advising it has converged on what works. We report a structured survey of 54 experts across 21 countries (Jul 2025 - Mar 2026) designed to map disagreement, not just opinion. Threat assessments converge sharply: 78% rate election-related video deepfakes as the single most urgent risk, and every modality-domain pair is rated above 5/7. Mitigation assessments do not: media literacy is the only strategy whose distribution is genuinely bimodal (BC=0.559), and government regulation is the most-chosen "most effective" (30%) yet is rated least effective by 15% of the same panel. We term this gap between converged threat perceptions and polarized mitigation preferences the Mitigation Paradox. Group comparisons further reveal that non-EU respondents rate financial-domain voice cloning as significantly more threatening than EU respondents (p=0.024), and rate technical detection as a more effective mitigation (p=0.060). Combining disagreement metrics with LLM-assisted thematic coding of open-ended responses, we identify three epistemic communities - infrastructure-oriented, governance-oriented, and harm-proximity - whose threat models systematically diverge. The Mitigation Paradox argues for layered, portfolio-style governance rather than single-strategy commitments.