What Happened
On September 16, 2026, Senior U.S. District Judge Susan Watters granted a preliminary injunction against enforcement of Montana's SB 25, the state's 2025 law restricting AI-generated "deepfakes" in election communications. The order, in Accountability in State Government v. Knudsen, came after former state legislator Dan Bartel and his political action committee sued over enforcement complaints filed against their mailers — AI-altered images depicting legislators holding pride flags and pronoun pins, and pumping cash from a gas station pump to criticize a gas-tax vote. Watters found the law likely violates the First Amendment because it draws a line the Constitution does not permit: it restricts deepfakes that damage a candidate's reputation while leaving deepfakes that flatter one alone.
The Law Montana Passed
SB 25 passed with bipartisan support — Democratic votes plus a majority of Republicans — and was signed by Gov. Greg Gianforte. It requires a disclosure statement on any AI-generated or digitally altered political communication distributed within 60 days of an election that is intended to "injure the reputation of the candidate or party or otherwise deceive a voter," with enforcement running through Montana's Commissioner of Political Practices and civil, potentially criminal, penalties attached. News broadcasts got a carve-out if they disclosed AI use in context.
Steelmanning the Legislature
Montana lawmakers were responding to a real problem. Generative AI has made convincing fabricated video, audio, and images cheap to produce and hard for an average voter to spot in the final weeks of a campaign, when there is little time for fact-checks to catch up before ballots are cast. A voter shown a fabricated clip of a candidate the week before an election has no realistic opportunity to learn it was fake before voting. Disclosure-based rules — as opposed to outright bans — are also the least restrictive tool available in that they don't suppress the speech, they just label it. That is a coherent, good-faith rationale, and it's why versions of this law have had support across party lines in several statehouses.
Where the Statute Broke
The problem, as Watters read it, is that SB 25 didn't stay content-neutral once it moved from disclosing AI use to targeting AI use that is unflattering. Because the statute only reaches deepfakes that damage a candidate, a favorable fabrication — an AI-smoothed clip making a candidate look more competent, more likeable, more electable — carries no disclosure obligation at all. Watters wrote that the law "treats identically deceptive media differently based solely on whether its message is weaponized to damage or deployed to elevate a candidate," which is the textbook definition of viewpoint discrimination: the government favoring one side of a debate over the content of the message. Under Reed v. Town of Gilbert and the line of cases it anchors, viewpoint-based speech restrictions get strict scrutiny, and Montana, the court found, hadn't shown the law was the least restrictive way to address a problem that existing tools — defamation law, election-fraud statutes, platform labeling — already partly cover.
Not an Outlier
Montana is now the third state to see an AI-election-speech law of this shape hit the same wall. A federal judge in the Eastern District of California struck down AB 2839 on August 29, 2025, in Kohls v. Bonta, finding it discriminated by content, viewpoint, and speaker alike; that ruling is now on appeal at the Ninth Circuit. A Hawaii law, Act 191, was separately found unconstitutionally vague in January 2026. Three states, three courts, one recurring defect: legislatures keep writing deepfake rules around the harm a fabrication causes a candidate rather than the fact that it's fabricated — and every version drafted that way has been treated as picking sides.
The Narrow-Injunction Wrinkle
Watters didn't block SB 25 statewide. Citing the Supreme Court's June 2025 ruling in Trump v. CASA curbing universal injunctions, she limited relief to Bartel and his PAC by name. That means Montana can still enforce SB 25 against everyone else in the state, even though a federal court has already found its core mechanism constitutionally suspect. Every other Montana campaign, candidate, or outside group now has to either self-censor under a law one federal judge has flagged as likely unconstitutional, or bring its own lawsuit to get the same protection Bartel just won. That is a genuinely uncomfortable position for the state to leave the law in, and it's the kind of outcome CASA invites more of: constitutional defects established case-by-case rather than resolved once for everyone the statute reaches.
The Fix Is Available
None of this means states are powerless. A disclosure law that applies to all materially deceptive AI-generated political content — flattering or damaging alike — would not have this defect, because it wouldn't be picking a side in the underlying political fight. That's a narrower drafting choice, not a weaker one. Montana, and the legislatures watching this ruling, don't need to abandon transparency requirements for AI in campaigns. They need to stop writing them as if only attacks on a candidate count as deception.
Key Takeaways
- A federal court found Montana's SB 25 likely unconstitutional because it disclosed AI-altered content only when it hurt a candidate, not when it helped one.
- Montana joins California and Hawaii as the third state to see an AI-election-speech law fail on First Amendment grounds within roughly a year.
- Because the injunction is narrow — following Trump v. CASA — SB 25 remains enforceable against everyone in Montana except the named plaintiffs, leaving the law's constitutional status unresolved statewide.