Someone you trust sends you a video, and you cannot place it.
Be honest about that moment. Not I knew it was fake — you'd love to be able to say that. Not I believed it, either. Fifteen seconds of a politician saying something slightly worse than he usually says, in a room you can't identify, lit too evenly. You watch it four times. You zoom in on the hands, because somebody once told you to look at the hands (advice with a shelf life). Then you do what everybody does: close the app and carry the thing around unresolved.
Your first instinct? There ought to be a law. There is, now — thirty-three states have enacted one — and in those thirty seconds on your sofa, not one of them would do anything for you. Not because legislators were lazy. Because of what a statute is.
In July I wrote here that the first deepfake election was underway and the referee had never shown up. The chair in Washington is still empty. What's changed is that the state record is finally long enough to read — and it doesn't say what I expected.
Thirty-three, or thirty-one
Start with the number in my headline, because it's contested — better the disagreement than a clean figure. By Ballotpedia's annual tally, thirty-three states have enacted laws regulating political deepfakes, up from 28 in July 2025 — and the next sentence carries a clause most coverage drops: the laws passed in Massachusetts and Hawaii "are no longer in effect." Enacted and operating are different words. Public Citizen, whose tracker most of this debate runs on, counts 31 states as of July 2026, and the midterms are the first major test of any of them. Two reputable counts, two numbers, and we haven't started arguing yet.
What do these laws do? Two things, mostly. They typically dictate when disclosure of AI's use is required and often give candidates targeted by non-compliant ads recourse in the courts. A label, and a lawsuit. Not a prosecutor, not a takedown button, not a hotline. Michigan Representative Penelope Tsernoglou, who wrote two of them in 2023, expects them "tested and stretched to their limits this year."
Montana's version is as plain as these get: within 60 days of the initiation of voting, covered material must carry a disclaimer saying it "has been significantly edited by artificial intelligence and depicts speech or conduct that falsely appears to be authentic or truthful" — with exceptions, including satire or parody. Hold on to that exemption. It does more work than anything else in the statute.
And the federal referee? Absent by choice. On September 19, 2024, the Federal Election Commission voted not to open a rulemaking on the use of AI in campaign ads, adopting instead an interpretive rule pointing at its existing fraudulent-misrepresentation regulation. The Conference Board's backgrounder is blunt: that move "replaces the prospect of new rulemaking", and the FCC's July 2024 proposed disclosure rule reaches radio and television only, since the agency "does not have jurisdiction over streaming video or digital platforms."
The case everyone points to, and how it ended
Every conversation about this arrives at New Hampshire eventually.
Two days before the January 23, 2024 presidential primary, thousands of voters answered the phone and heard a voice very like Joe Biden's telling them to stay home. The consultant behind it, Steve Kramer, had paid a New Orleans magician $150 to make the recording. One hundred and fifty dollars (I'm not rounding).
New Hampshire charged him under statutes that say nothing about artificial intelligence: eleven felony counts of voter suppression, each punishable by up to seven years, and eleven counts of impersonating a candidate. In June 2025, a jury acquitted him of every count.
So the criminal scoreboard reads zero. The money is a different story — and neither half is a conviction.
The Federal Communications Commission adopted a six-million-dollar forfeiture order against Kramer on September 26, 2024 — not for using AI, but for violating section 227(e) of the Communications Act and section 64.1604 of its rules. Caller ID. The order describes a voice "artificially created to sound like President Biden," but the violation it names is about who the call claimed to be from. And a forfeiture order is a civil finding of liability to pay, not a criminal judgment; I couldn't establish from a source I'd stand behind whether the six million was ever collected, so I won't say it was.
Then the carrier: Lingo Telecom, which carried 9,581 of those calls, agreed to pay $1,000,000 and adopt a compliance plan to settle the FCC's investigation into its "apparent violations" — the Commission's own word — of the caller-ID authentication rules. A consent decree resolves a matter; it doesn't find liability.
Why does that matter? Section 227(e) exists because of the Truth in Caller ID Act of 2009, approved December 22, 2010 — signed years before anyone could clone a voice from a podcast. Read what it forbids: transmitting "misleading or inaccurate caller identification information with the intent to defraud, cause harm, or wrongfully obtain anything of value." Not a technology. A lie told for gain — and conduct doesn't go out of date when the tooling improves.
The scoreboard, one courtroom at a time
Here's where it gets uncomfortable for anyone who wanted these laws to work (if you nodded along to there ought to be a law, that includes you).
California. On August 29, 2025, Judge John Mendez struck down AB 2839, calling it "well-intentioned but constitutionally infirm" and prescribing counter speech and rigorous fact-checking instead. Nine days earlier the same court killed the companion statute, AB 2655, on different grounds: preemption by Section 230, the judge expressly not reaching the First Amendment. One unconstitutional, one preempted — not the same thing. And neither is the end: the state appealed to the Ninth Circuit, and a California labeling law remains in effect.
Hawaii. On January 30, 2026, in Babylon Bee, LLC v. Lopez, Judge Shanlyn Park held Act 191 unconstitutional, finding its "reckless disregard of the risk of harming" standard muddies the line between compliance and noncompliance. Trade coverage records her calling the law presumptively invalid and unconstitutionally vague. Then the coda: Alliance Defending Freedom, which represented The Babylon Bee, says the state settled in May 2026 and paid more than $118,000 in attorneys' fees. A state that legislated against deceptive media, paying a satire site's lawyers. (Satire, it turns out, writes itself.)
Montana. AI-altered mailers from a conservative committee drew three complaints to the Commissioner of Political Practices. All three will be dismissed under the satire-or-parody exemption, Commissioner Chris Gallus writing that "further factual development is not necessary." That is an executive-branch officer disposing of complaints filed with his own office, not a court construing the statute. On September 16, 2026, a court construed it. Senior U.S. District Judge Susan Watters granted a preliminary injunction against SB 25 — the suit filed in May 2026 by Bozeman attorney Matthew Monforton, arguing the mailers are protected speech — finding the state had "failed to prove that the purported risks of generative AI technologies could not be adequately managed" by something less restrictive. But the deeper objection is the one to sit with. Because the disclaimer duty attaches to synthetic media that injures a candidate's reputation, the law treats "identically deceptive media differently based solely on whether its message is weaponized or deployed to elevate a candidate" — viewpoint-based discrimination, and so "presumptively unconstitutional." A statute written against deceiving voters turned out, on inspection, to be a statute against deceiving them unkindly.
Before you file that as a fourth law struck down, read the remedy. It isn't one. The same ruling reaches only Dan Bartel and his committee, leaving SB 25 in force for everybody else in Montana — Watters citing Trump v. CASA, where the Supreme Court held that the Judiciary Act of 1789 gives federal courts no power to issue universal injunctions. The statute stands; two litigants are exempt from it. And they told the court what the exemption is for: absent the threat of enforcement, the committee planned to send more AI-generated mailers in October, this time against Democratic legislative candidates.
Minnesota. This one gets written up as a win for the laws; it isn't one yet. X Corp challenged the state's statute, and on December 2, 2025, Judge Laura M. Provinzino denied the motion and dismissed the Section 230 count without prejudice for lack of Article III standing, then stayed the case pending Kohls v. Ellison in the Eighth Circuit; X had failed to show it faced an enforcement action. The First Amendment questions weren't answered — postponed.
Add it up: three state laws struck down — two on constitutional grounds, one on federal preemption, with California's on appeal — a fourth enjoined for two litigants and nobody else, and Minnesota's challenge still stayed. Not one of these matters has produced a penalty against anybody.
So the laws are a failure? That isn't what the record says either
Now let me argue against my own headline.
The laws getting killed aren't the typical law. Chris McIsaac of the free-market R Street Institute — no instinct there to defend speech regulation — notes that most states have yet to see any legal challenge to the more common approach, which requires only a disclosure alerting viewers that AI was used in deceptive election media; his snapshot counted 26 states, "compared to just five in 2023." The prohibitions lose in court. The labels mostly haven't been tested.
The defenders aren't making a silly argument. Public Citizen's Ilana Beller, who maintains that tracker, told the same Arizona reporting that these laws "really zero in on the harmful content" — and by her reading, most AI material this cycle has been labeled or plainly satirical. You can't disprove that by pointing at courtrooms.
And the frame may be measuring the wrong game. David Becker of the nonpartisan Center for Election Innovation & Research cautions that AI is not radically changing the misinformation environment, because the most powerful driver of election misinformation remains false claims from prominent people, not the software. If he's right, a perfect deepfake statute addresses a modest slice of the problem.
There's even measurement pointing that way. Researchers studying the 2025 Canadian election found 5.86 percent of the election-related images they analyzed were deepfakes, but the harmful ones drew only 0.12 percent of all views on X — "deepfakes were present in the election conversation, but their reach was modest." That's Canada, not an American midterm, so I'm not transplanting it — only leaving it as the honest possibility that a lot of legislative energy went toward a small fire.
South Korea ran the other experiment. Its bill comes due in people.
If you want the maximalist version, you don't have to imagine it.
In December 2023, a revision to the Public Official Election Act banning AI-deepfake campaign videos during the election season cleared a special parliamentary committee carrying "a maximum seven years in prison or a fine of up to 50 million won ($38,100)." Under the resulting Article 82-8, South Korea's National Election Commission explains, no one may produce, edit, distribute, screen, or post deepfake videos for election campaigning purposes from 90 days before the election day to the election day. It took effect on January 29; the commission stood up an AI identification team and expanded its monitoring staff. Flat ban, criminal penalties, a defined window, no disclosure escape hatch — more or less the law American reformers keep asking for.
So what does it buy? Between January 29 and February 16, 2024, the commission reported busting 129 deepfakes "deemed to violate the laws on elections of public officials" with a task force of 72 officials and experts; "most of them have since been taken down." Mind the two soft words: deemed, and most. In the 2025 presidential race, police set up teams at 278 precincts nationwide. Before the June 3, 2026 local elections, an AFP reporter watched a commission monitor and 18 colleagues click through Instagram, YouTube and other platforms by hand; the government has hired hundreds of staff; a forensic specialist acknowledged the response has "many limits."
Sit with what that is: detection and takedown — monitors, task forces, precinct desks, content removed. The penalties on the books are real — up to seven years — but in sources I'd stand behind I found no reported conviction under Article 82-8. What I can prove is the labor: people, at screens, clicking.
The constitutional objection isn't absent there either. Open Net Korea filed a constitutional complaint on October 1, 2025, arguing that under this provision all such materials are prohibited regardless of whether the content is true, whether consent was given, or whether it was labeled a deepfake. It's pending. And Lasse Schuldt, at Verfassungsblog, makes the scholarly version: the article criminalizes deepfakes per se during election periods, where prior rules "penalized probable harm – not fake news as such"; the Korean and Singaporean laws, he writes, restrict speech "also in cases where the content in question causes no harm."
Which is nearly word for word the objection American judges sustained in Sacramento and Honolulu — except Korea runs the experiment while it argues, and we argue instead.
Forty-eight hours before the polls close
Just imagine the home stretch of a close race.
Monday, 9 p.m. A ninety-second clip lands on three platforms at once: a Senate candidate in a parking garage, saying something ugly about her own district. No disclaimer — which, in a disclosure state, is precisely the violation. Her campaign does the correct thing and sues. Now somebody must draft, file, find a judge, and persuade that judge to restrain political speech in the election's final hours. Assume the fastest hearing ever granted: Wednesday. The polls closed Tuesday at eight. And the account calls itself a parody account, which is how Montana's mailers walked out.
Then the move that worries me more than any synthetic video: a candidate caught on genuine footage saying a genuine thing looks into a camera and says that's AI — and a good share of the electorate has reason to believe him, because we've spent three years teaching everyone that the reason exists.
Push it to 2028. Generation moves onto the handset, a convincing clip costs nothing, and the volume stops being something 18 people can click through. Then one of two things must be true: enforcement by machines, every false positive a censored citizen — or enforcement by people, a payroll on the scale South Korea runs, times fifty states. I don't have a clean answer — only a preference for knowing the bill before we sign.
The people who have thought hardest about this don't agree, and the split isn't left versus right
Bartlett D. Cleland, a senior fellow at the Pacific Research Institute, states the case for my headline more sharply than I would dare. In the same Washington Examiner op-ed that carried David Becker's caution, he writes that "No AI statute was required to reach the conduct. The law focused on the deception, and the deception was enough." Fair — but note what his piece doesn't say: the state prosecution he points to ended in acquittal. The old statutes were charged, and they lost in front of the jury too.
Clay Calvert, a First Amendment scholar at the American Enterprise Institute, gives the most useful version of the objection, because it isn't "do nothing." Reading the Hawaii ruling, he notes the judge agreed the state had a compelling interest in regulating political deepfakes — "an essential democratic function" — and struck the law anyway because "counter speech appears to be a viable, less restrictive alternative": crowd-sourced fact checking, state counter-speech, a database tracking materially deceptive content. Judge Mendez prescribed much the same in California.
Which leaves the question nobody has priced. Who staffs that database? Who pays the fact-checkers at 11 p.m. the Monday before an election? Korea answered with a payroll. We answered with a citation.
What does this mean for you?
Find out which kind of law your state has, and read its exemptions. A disclosure law gives you a label and a wronged candidate a court date — not a referee on election night.
Don't forward the thing you can't place. It's the only lever that works at the speed of the problem. You are the distribution — screenshot it, note the time, and sit on it.
Report it to the state, and expect slowness. Oregon's Secretary of State is investigating whether a former congressional candidate violated its disclosure law — maximum penalty $10,000, no finding yet. The channel exists; it runs on a legal calendar, not a news cycle.
If you run a campaign, prepare counter-speech, not just a filing. A pre-recorded rebuttal, a verification page, a spokesperson reachable at midnight. The injunction, if it comes, comes after the count.
Watch the exemptions, not the prohibitions. Satire and parody swallowed Montana's complaints whole. Whatever your state's carve-out is, that's where the fight happens.
Remember the old laws are still there. Fraud, impersonation, voter suppression, lying about who is calling. They are what the New Hampshire case ran on, and they will still be on the books in 2028.
The lesson, as I see it
A statute is a sentence. Enforcement is a payroll. Nearly everything that has gone sideways here lives in the gap between those two — and that gap isn't a partisan failing, it's an engineering fact legislators of every stripe underestimate.
Look again at the one instrument that produced anything here: written in 2010, aimed at a lie rather than a tool. It fit the 2024 robocall exactly, because lying about who is speaking has been illegal for a very long time. Laws aimed at conduct age well. Laws aimed at a tool age at the speed of the tool.
November is the first general election with all of this on the books, and my honest forecast is that most of these laws won't be used. Watch which ones produce something, not which ones get cited. If a year from now the answer is still "the old deception statutes, plus a room of people clicking," we'll have learned something worth more than thirty-three press releases: the hard part was never the writing. It was the twelve hours after the clip lands, which no legislature has yet found a way to legislate.
If you're carrying a video you can't place, pass this to whoever sent it to you and compare notes — that conversation moves faster than any statute. The HAIA Foundation lives in the gap between what a law says and what it can do, and the rest of it turns up here.