Election deepfake laws meet their first real election

In June, Michigan Senate Minority Leader Aric Nesbitt posted an AI-generated video of Governor Gretchen Whitmer being hit by a tractor. It carried a label. The same month, former congressman Mike Rogers, running for Michigan's open Senate seat, released an AI-generated video casting himself as a Hulk-like crime fighter. That one did not. Same state, same technology, same election — one disclosure, one none.

Michigan has an election deepfake law. So do thirty other states. On 3 November, for the first time, every one of them meets a general election at the same moment.

The interesting question is no longer whether states will regulate synthetic political media. They already have. It is whether a disclosure rule can be enforced against content whose origin nobody can establish — because that, not the drafting, is where these statutes break.

Thirty-one states, and no federal floor

Public Citizen, which has tracked this legislation bill by bill since 2023, counted election deepfake laws on the books in 31 states as of July 2026. They are not one law thirty-one times. Minnesota and Texas prohibit political deepfakes inside a window before election day. Maryland's restriction runs year-round. Most of the rest take the lighter path: require a disclosure when synthetic media appears in political advertising, and give a candidate depicted without one a way into court.

Above that patchwork there is nothing. The Federal Election Commission has not adopted rules specific to AI in campaign advertising, and reporting on the 2026 cycle describes a commission split along partisan lines and unable to produce guidance. So the operative law is thirty-one definitions of "materially deceptive," thirty-one sets of exemptions, and thirty-one enforcement bodies — governing a medium in which a video crosses all fifty states in an afternoon.

The courts took the bans and left the labels

California went furthest, and lost. AB 2839 restricted the distribution of materially deceptive election media. On 29 August 2025, Judge John A. Mendez of the Eastern District of California entered summary judgment against it and permanently enjoined its enforcement, holding that it imposed content-, viewpoint-, and speaker-based restrictions on core political speech and could not survive strict scrutiny. His reasoning is worth reading as a constraint on every legislature that follows:

The antidote is not prematurely stifling content creation and singling out specific speakers but encouraging counter speech, rigorous fact-checking, and the uninhibited flow of democratic discourse.

Nine days earlier the same judge had struck down AB 2655, the companion statute directing platforms to block or label election deepfakes, on federal preemption grounds under Section 230 — without needing to reach the First Amendment at all. Hawaii's SB 2687, enacted in July 2024, was also struck down. Montana's law is under challenge now.

What survives that reasoning is narrow and specific. Not you may not publish this, but if you publish this, say so. Compelled disclosure is a lighter burden than prohibition, which is exactly why most legislatures were already there.

Notice what the surviving approach demands, though. A ban asks a court to judge content, which courts do all the time. A disclosure mandate asks something harder: it asks whether a particular file was synthetic, and who made it that way. That is a question about origin — and on that question a court has no better instrument than the rest of us.

A disclosure rule is only as good as its evidence

Follow an actual complaint and the same three failures show up every time.

In Oregon, a former Republican congressional candidate posted unlabeled synthetic videos of Representative Janelle Bynum after losing his May primary. The Secretary of State opened an investigation. His answer was that Oregon's synthetic media law is unconstitutional. In Arizona in July, a congressional candidate sued a super PAC over unlabeled billboards carrying AI-generated images. In Kentucky, one PAC ran a satirical AI ad about Representative Thomas Massie and another ran an unlabeled one about the primary's winner.

The satire carve-out swallows the rule. Michigan and Oregon both exempt parody and satire, and so do many of the others. Attack advertising has been arguing it is satire since long before diffusion models existed. A statute that exempts satire and regulates deception has handed the defendant the framing of the case.

Nobody can establish the artifact. By the time a complaint is filed, the video in question has been re-uploaded, transcoded, clipped into a vertical crop, and screenshotted into a group chat. Whatever disclosure metadata the original carried — if it carried any — is gone. The thing under dispute is no longer the thing that was published, and the regulator is left arguing about a copy of a copy.

The clock beats the process. In January 2024, political consultant Steve Kramer sent a robocall using a cloned Biden voice telling New Hampshire voters to skip the primary. It is the most clear-cut, most investigated, most widely reported synthetic-media incident in recent American politics. He was acquitted of the criminal charges in 2025, while facing a multimillion-dollar FCC fine. If that case took the better part of two years to half-resolve, an unlabeled ad dropped six weeks out is adjudicated, at best, long after the seat is filled.

Scott Babwah Brennen, who studies political ad transparency at NYU, framed the open question about labeling requirements directly: "Are they going to be effective? And are they up to the task of addressing the problems that we're seeing?" Michigan Representative Penelope Tsernoglou, who backed her state's law, expects the answer this cycle: "We are going to see them tested and stretched to their limits this year."

Detection cannot referee an election

The instinct is to reach for a detector. Run the ad through a classifier, get a number, enforce against the number.

That fails on contact with a courtroom. A classifier returns a probability derived from patterns in its training data, and the probability moves when the generator changes, when the clip is recompressed, when it is filmed off a screen. No state statute says "liability attaches above 80% likelihood," and none will, because a percentage from a proprietary model is not a fact a respondent can cross-examine. We have written at length about why provenance beats detection; election law is the cleanest illustration. The standard is preponderance or better. The tool offers a hunch.

Worse, detection has exactly the error profile you least want in politics. A false positive brands a real recording a fake — which is the accusation a campaign most wants to make about inconvenient footage, and now it arrives with a number attached.

What a record made at creation would give a regulator

Invert the problem. Instead of asking a machine to guess after the fact, ask what an honest campaign could write down at the moment of creation.

A record like that has four properties an enforcement body actually needs. It names the origin: which model, which tool, which account, cryptographically signed, so a disclosure becomes a claim someone is accountable for rather than a line of ad copy. It is bound by a perceptual fingerprint — derived from what the content looks like, not from its bytes — so the vertical crop that landed in the group chat still matches the original entry. It sits in an append-only public log, timestamped, so the record demonstrably predates the complaint instead of being assembled in response to it; that is the same structure Certificate Transparency uses, and we have written about why an append-only log matters. And checking it is free and needs no account, so a journalist, an opposing campaign, and a Secretary of State's office all query the same record and get the same answer. You can run a verification against a file right now.

The same record cuts in the other direction, which is the part campaigns underrate. A campaign that stamps its genuine footage at capture has something concrete to point at when a fabricated version circulates. That is the only real defense against the liar's dividend — the increasingly common move of dismissing an authentic recording as AI. Denial is cheap when nothing was ever recorded about the original.

What this does not fix

An operative who fabricates a video of a candidate is not going to stamp it. Provenance is not a content filter and never will be. The absence of a record proves nothing about a file except that the file has no record — and any system that treats "unstamped" as "fake" has reinvented the false-positive problem it was supposed to solve. That asymmetry is the honest limit of the whole approach, and we have been explicit that Content Credentials solve half the problem.

It does not resolve satire either. A labeled parody is still a parody, and whether a reasonable viewer was deceived stays a judgment for humans.

What it changes is who carries the uncertainty. Today the fabricator and the honest campaign sit in the same evidentiary position: neither can prove anything about where their video came from, so disputes collapse into assertion against assertion, settled by whoever is louder before election day. A provenance layer does not silence the fabricator. It gives everyone else something to hold up.

What a campaign can do before 3 November

None of the above helps a communications director who has an ad to ship this week. Three things do.

Stamp your own footage at capture. Every rally recording, every candidate interview, every piece of B-roll. The cost is near zero and the payoff is asymmetric: it does nothing on an ordinary day and everything on the day a manipulated clip of that same event circulates. You cannot retroactively create a record of the moment a camera was rolling, which means the decision has to be made before you know you need it.

Disclose in the artifact, not just the caption. A caption survives exactly one re-upload. Whatever the applicable state law demands as a visible label, pair it with a signed machine-readable record, because the copy that reaches a Secretary of State's office will not be the copy you posted.

Log the opposition's ads as you find them. Register the fingerprint of the file you actually downloaded, at the time you downloaded it. It will not prove the ad was synthetic. It does prove which artifact you are complaining about and when it was in circulation — which is the evidentiary question that otherwise decays into a copy of a copy.

Campaigns already do the analogous thing for broadcast: they keep tapes. The digital equivalent has simply never been available as a routine practice.

Six weeks

Thirty-one legislatures have written down the rule. None of them built the thing that would make it checkable, and the first general election to test all of them at once is six weeks away.

The statutes were the easy part. A law about where a video came from cannot be enforced in a system that never wrote it down.

Further reading: Tracker: state legislation on deepfakes in elections (Public Citizen) · State AI deepfake laws face first big test in 2026 midterm elections (Arizona Capitol Times) · Kohls v. Bonta (Columbia Global Freedom of Expression) · Midterms put deepfake election laws to test (Pluribus News) · EU AI Act, Article 50