Elections & Democracy

AI, Deepfakes & Election Disinformation

2023— present

Political manipulation is ancient. Cheap, convincing synthetic media is three years old — and 31 state laws written against it have never faced a general election.

About thirty states wrote laws against election deepfakes. November 2026 is the first real test of whether they work. Generative AI made synthetic audio and video cheap and convincing almost overnight, and legislatures responded fast: 31 states now have laws addressing deepfakes in political communications, most requiring disclosure rather than banning outright. But enforcement has been thin, California's and Hawaii's versions were struck down on First Amendment grounds, and the FEC has issued no rule. This covers what the technology can now do, what the state laws actually require, the constitutional problem with regulating political speech, and how to evaluate a suspicious clip.

States with an election deepfake law in force

as of Aug 3, 2026 (Public Citizen tracker); was 30 in May

States with any deepfake statute

all but Ohio and New Mexico, up from 46 in January (Ballotpedia)

Fine in the 2024 New Hampshire AI robocall

$6M

$1,000 × 3,000 spoofed calls, doubled for egregiousness (FCC)

Most common pre-election window

days; the range runs 30 (Texas) to 120 (California)

Federal statutes on AI in political ads

FEC declined a rule in 2024; the FCC never finished its own

Read the horizon before the charts

Political manipulation is as old as politics. Fabricated quotes helped decide 1800; darkroom compositing was routine by 1900; the 1972 Canuck letter and a hundred selectively edited tapes are all the same impulse. What changed is cost. A convincing impersonation of a specific person used to require money, skill and time; it now requires seconds of reference audio and a few dollars. That is why every measured series on this page starts in 2019 or later — the first two state statutes were passed that year, and the phenomenon they aim at barely existed before 2023. There is no long baseline to chart here, and drawing one would be a lie about the data.

What came before — context, not a series

1796 — 1860

The partisan press

Newspapers were openly owned by factions and parties. The 1800 Adams–Jefferson race featured fabricated quotes, invented biographies, and a widely printed false report that Jefferson had died. Nothing about the content was synthetic; the distortion was editorial and the distribution was slow.

1898 — 1940s

Retouching and the doctored photograph

Darkroom compositing was well understood by the 1890s. Stalin's censors airbrushed purged officials out of official photographs through the 1930s. The skill barrier was high, the output was static, and a single image took a trained technician days.

1950s — 1980s

Broadcast-era dirty tricks

The 1964 "Daisy" ad, the 1972 Canuck letter, selective edits of tape. Manipulation moved at broadcast speed and through a small number of gatekeepers who could be identified and, in the Fairness Doctrine era, regulated.

1990s — 2016

Photoshop, then platforms

Desktop image editing put compositing in every campaign office by the late 1990s. Social platforms then removed the distribution bottleneck. By 2016 the constraint on political deception was no longer reach or editing — it was the cost of making a fake that survived scrutiny.

2023 —

The synthesis break

Generative models collapsed that last cost. A convincing clone of a specific person's voice now takes seconds of reference audio, minutes of work, and a few dollars. That is the discontinuity this page measures, and it is why every series here starts in the 2020s rather than the 1800s.

The AI-specific record · 2023 — present

Hover any milestone for context.

IncidentLaw / enforcementCourt rulingAgency action

Earliest entry: 2023. The technology existed earlier, but there is no documented use of generative synthetic media in a U.S. federal campaign before the RNC's April 2023 spot.

What it can do now, by modality

VoiceSolved, and cheap

Commercial voice cloning needs seconds of clean reference audio — trivially available for any candidate, election official or local anchor. The New Hampshire robocall is the documented benchmark: the operator who made it told CNN it took about 20 minutes and he was paid roughly $150. Phone audio is the worst case for a listener, because bandwidth compression hides exactly the artefacts a person would otherwise notice.

TELL · Breath and room tone. Clones often have unnaturally even pacing, no lip-smacks, and a background that does not change when the speaker moves.

VideoGood, and improving fastest

Text-to-video systems now produce short clips with synchronised speech, plausible lighting and coherent camera motion. The practical limits are duration and continuity: quality is highest in the first seconds and degrades as a clip runs, and physical interactions — hands, objects being passed, text on signage — remain where errors concentrate. That matters politically, because a damaging clip only needs to be eight seconds long.

TELL · Watch the boundaries: hairlines, glasses, jewellery, teeth, and anything the subject touches. Freeze on a frame where the head turns.

ImagesEffectively unlimited

Still images have been past the detection threshold for ordinary viewers since 2023 and are the most common form in real campaigns — AI billboards, fake crowd shots, composite mailers. Because a still carries no motion or audio cues, almost every artefact-based test people have learned is unavailable.

TELL · Reverse image search first. Provenance beats forensics on stills: check whether the file carries Content Credentials at contentcredentials.org/verify.

Text and personasThe underrated one

The volume problem is not any single fake video — it is thousands of coherent accounts producing locally plausible political text. Meta reported blocking 590,000 requests to generate images of the 2024 presidential and vice-presidential candidates in the month before Election Day, but the harder measurement problem is synthetic accounts that never generate an image at all.

TELL · Look at the account, not the post: creation date, posting cadence, whether it engages with replies.

How much accuracy detectors lose in the wild

Drop in AUC for open-source state-of-the-art detectors, in-the-wild deepfakes vs. the academic benchmarks they were tuned on.

The benchmark is 45 hours of video, 56.5 hours of audio and 1,975 images, collected from 88 websites in 52 languages.
Commercial detectors and models fine-tuned on the in-the-wild data beat off-the-shelf open-source models — but still do not reach the accuracy of human forensic analysts.
The paper's conclusion is the one that matters for election use: academic benchmarks are out of date and not representative of real-world deepfakes. A vendor accuracy figure quoted without naming the test set tells you nothing.
On in-the-wild speech sets, some detectors fall to near chance. Treat any single automated verdict on a political clip as evidence, never as proof.

Source: Deepfake-Eval-2024 (arXiv 2503.02857). Earliest available year: 2024 — this is the first benchmark built from deepfakes actually circulating in the wild, so no earlier comparison exists.