The documented election-deepfake record is narrower than the discourse suggests. The most systematic public accounting — the deepfake database maintained by researchers at the University of Amsterdam and cited widely in 2024–2025 analyses — counted a few thousand election-related deepfake items across the major 2024 election cycle, with the large majority satirical or clearly labeled content, while the genuinely deceptive items that spread widely were dominated by low-tech manipulation: slowed video, miscaptioned clips, crudely edited audio. In the United States, the incident most often cited as a demonstrated effect on an election process was the January 2024 New Hampshire primary robocall using an AI voice clone of a candidate telling voters to stay home — an attempt at suppression that the Federal Communications Commission subsequently addressed by ruling that AI-generated voices in robocalls are artificial under the Telephone Consumer Protection Act, making them illegal without consent.
Which formats are actually being used?
The documented cases cluster into three tiers. Voice dominates: audio cloning requires seconds of source material, travels well on messaging apps, and lacks the visual artifacts that let viewers scrutinize video. Image and video fabrications come second — mostly nonconsensual imagery and staged satirical clips, with the satire frequently stripped of its context as it travels. The third and largest tier is what researchers call cheap fakes: authentic media re-captioned or selectively trimmed. The synthesized content gets the hearings; the cheap fakes do the distributing, because they inherit the credibility of real footage and cost nothing to produce.
Do deepfakes change votes?
Evidence of persuasion effects remains thin. Experimental studies through 2024–2025 consistently found that exposure to a political deepfake produces small, short-lived attitude shifts that decay within days, while belief corrections work reasonably well when delivered promptly. The better-documented harm is epistemic: the liar's dividend, a term coined by legal scholars Bobby Chesney and Danielle Citron — the ability of a real recording to be dismissed as fake. Campaigns caught in authentic scandals now routinely allege synthesis, and the mere availability of the technology degrades the default credibility of all recordings. That harm is not measurable as flipped votes but as a widening baseline of disbelief, and it is the harm most resistant to detection tools.
What detection and labeling approaches exist?
Three families, with honest limits. Automated detection models flag statistical artifacts of generation, but performance degrades on compressed, re-encoded social-media video — precisely the form that spreads — and open benchmarks show accuracy falling well below lab conditions on in-the-wild samples. Provenance standards, notably the C2PA specification adopted by major camera manufacturers, platform operators and news agencies, cryptographically sign authentic capture at the source rather than detecting fakes; its weakness is adoption, since unsigned real content remains the norm. Platform labeling policies — Meta's decision to label AI-generated media rather than remove it, announced in 2024 — put the flag inside the distribution channel, but enforcement depends on advertisers and uploaders self-disclosing, and deceptive actors do not. No family is useless; none is close to sufficient alone.
What has regulation actually done?
Three concrete actions anchor the U.S. record. The FCC's February 2024 declaratory ruling on AI voices in robocalls gave enforcers a bright-line tool against the New Hampshire-style audio vector, and state attorneys general followed with the robocall case's litigation. The FEC's proceeding on AI in campaign ads — prompted by a petition concerning a 2023 synthetic advertisement — concluded in 2025 with the commission declining to rewrite its regulations, finding existing fraud provisions adequate and citing statutory limits on its authority over content. And state legislatures filled the gap unevenly: by 2025 more than twenty states had enacted election-deepfake statutes, typically requiring disclosure labels on synthetic media about candidates within election windows, with several already under First Amendment challenge in the courts. The pattern is classic: communications regulators moved where their statutes reached, the campaign-finance regulator declined, and the states produced a patchwork.
What should newsrooms actually do?
The documented best practice is process, not software: verify at the source (confirm with the purported speaker or their office before publishing a viral clip), privilege provenance-signed material where it exists, treat reader suspicions of fakery as a verification request rather than a debunk request, and — the discipline most often skipped — apply the same skepticism to allegations that real media is fake. An outlet that debunks fabrications but launder liar's-dividend claims through hedged language is solving half the problem it created for itself.
For more context, read How television technology quietly redesigned political debating.
For more context, read push poll definition.
For more context, read political influencer disclosure.