Deepfakes in elections. Impact and how to combat them.
2024 was the first election cycle where AI-generated audio reached voters at scale. What we saw, what we caught, and what changes for the next cycle.
What happened in 2024
The category opened with the New Hampshire primary robocall in January, where an AI-generated voice of President Biden urged voters to save their vote for November. The robocall reached thousands of phones before law enforcement intervened. We were one of several teams that flagged the audio as synthetic on first listen.
From there, similar incidents surfaced in roughly a dozen democracies through 2024: Slovakia, Indonesia, India, the UK, Pakistan. The pattern was consistent.
Patterns we observed
- Late-cycle release. Deepfaked audio dropped 48 to 72 hours before voting, when the news cycle is too short for rebuttal to compete with the original claim.
- Distribution through unofficial channels. WhatsApp groups, encrypted messaging, low-friction sharing. The clip had reach before any platform could moderate.
- Plausible deniability. Some campaigns ran the synthesis themselves; others tolerated supporters who did. Attribution was hard.
- Targeted demographics. The deepfakes were not for everyone. They were targeted at specific voter segments (older voters, language minorities) least likely to have detection tools.
What we caught
We flagged the New Hampshire Biden robocall within hours. We flagged a deepfaked clip of a UK opposition leader during the 2024 general election within an afternoon. We caught more than 30 election-related submissions during the cycle, ranging from confirmed synthesis to confirmed authentic.
The numbers we cannot give you: the ones that ran past us entirely. The category is asymmetric. Attackers ship at scale; detectors are reactive.
Combatting deepfakes in elections
- Rapid response. Detection within hours, not days. Election integrity teams now run continuous monitoring on social channels.
- Provenance standards. Real campaign audio should carry verifiable provenance (signed metadata, blockchain hashes, or similar). The C2PA standard is the most credible candidate.
- Pre-bunking. Voters who are told in advance that deepfakes are likely become more skeptical of late-cycle audio. Public service campaigns work.
- Detector access for journalists. Every newsroom covering an election should have a detector account. We make ours free for verified journalists.
Ensuring election integrity
The detector is a small part of a larger system. It cannot stop the deepfake from being made. It can only verify a specific recording once it surfaces. The harder work is building voter resilience: public awareness, rapid platform moderation, and provenance infrastructure that makes verified audio trivially distinguishable from unverified.
2024 was the first cycle where this happened at scale. 2026 will be worse. The work is now permanent.