Two days before Florida’s August 18 Republican primary, a Hillsborough County court order stopping a super PAC’s attack mailers is being fought over a single technical question: did generative artificial intelligence make the image, or did a person with Photoshop? On the answer hangs the first civil test of Florida’s AI political ad disclaimer law, and possibly the template for how twenty-eight states enforce theirs.
The case is Mike Beltran for Congress v. Conservative Solutions for America, case number 26-CA-008578, filed August 7, 2026 in the Thirteenth Judicial Circuit. The campaign committee of Michael Beltran, a former state representative running in Florida’s 14th Congressional District, alleges that the PAC blasted Hillsborough County voters with texts and printed mailers showing Beltran holding a “DEFUND THE POLICE” sign in front of riot-gear police and raised fists. The verified complaint calls the image “absolutely false” and says it was made “in whole or in part through the use of generative artificial intelligence (or digital manipulation that produces the same effect)” without the disclaimer Florida law requires.
A judge agreed enough to issue an order stopping distribution, Florida Politics reported. The PAC’s response, filed in circuit court, is the reason this case matters beyond one primary: its brief argues the image was not created with generative AI at all, that “traditional tools like Photoshop were used,” and that no reasonable person would mistake the photo illustration for a real event. An emergency hearing on dissolving the order was set for this week. The statute’s fate now rides on the difference between two buttons in the same application.
What the complaint and the statute actually say
According to the verified complaint, the texts went out on or about August 3, 2026, followed the same week by printed mailers with the same or substantially identical image. The mailers, as quoted in the filing, declare that “MIKE BELTRAN’S LIBERAL RECORD WON’T KEEP HILLSBOROUGH FAMILIES SAFE,” that he “sided with defunding the police,” and that he voted against a budget funding law enforcement and against a School Guardian program. Beltran, per the complaint, never held such a sign, never posed for the photo, and never authorized it. The filing devotes three pages to rebutting the underlying vote characterizations, quoting his 2024 budget statement and 2019 reporting on his opposition to arming teachers.
The complaint alleges the PAC received all of its funding from a now-defunct state committee tied to Kevin Steele, Beltran’s chief rival among eight primary candidates, and is run by Steele’s fundraiser and consultant. Those are allegations the PAC has not answered in a responsive pleading, though its spokesperson, Stafford Jones, has defended the mailers publicly. The winner of the primary faces Democratic U.S. Representative Kathy Castor in November.
The legal hook is section 106.145, Florida Statutes, enacted as HB 919 in 2024 and effective July 1 of that year. The statute fires when four elements line up: a political advertisement or electioneering communication contains content “created in whole or in part with the use of generative artificial intelligence,” the generated content “appears to depict a real person performing an action that did not actually occur,” and the content was created with intent to injure a candidate or deceive on a ballot issue. When all of that is true, the ad must prominently state, word for word: “Created in whole or in part with the use of generative artificial intelligence (AI).” Print requires bold type of at least 12 points. Video requires the disclaimer to stay readable throughout and fill at least 4 percent of the picture height. The complaint alleges the text message carried the wrong disclaimer and the mailers carried none.
Notably, the campaign is not suing for defamation and seeks no damages. Count I asks for a declaration that the communications violate section 106.145. Count II asks for a temporary and permanent injunction. That restraint is strategic, and it exposes the statute’s odd enforcement architecture, which we return to below.
The definitional boundary: what counts as generative AI
The PAC’s defense sounds like a factual quibble. It is actually an attack on the statute’s load-bearing wall. Section 106.145(1) defines generative artificial intelligence as “a machine-based system that can, for a given set of human-defined objectives, emulate the structure and characteristics of input data in order to generate derived synthetic content including images, videos, audio, text, and other digital content.” The disclaimer duty attaches only to content created with such a system. The trigger is the tool class, not the deceptive result.
Now map real tools onto that definition. A manual Photoshop composite, in which a designer cuts Beltran’s head from a real photo, pastes it onto another body, clone-stamps the seams, and hand-matches the color grading, is human-directed pixel arithmetic. No model emulates training data to synthesize content. On a plain reading, that workflow sits outside the definition, and the mailer needs no AI disclaimer no matter how false its message.
But “we used Photoshop” answers nothing, because Photoshop has shipped a diffusion model inside it since 2023. Generative Fill and Generative Expand, powered by Adobe’s Firefly, mask a region and synthesize replacement pixels from a trained model conditioned on a text prompt, which is precisely a system that emulates the structure of input data to generate derived synthetic content. The same application is a darkroom or a diffusion front-end depending on which button the designer pressed. A composite whose crowd, sign, or seam-blending came from Generative Fill is inside the statute even if every other step was manual, because the statute says “in whole or in part.”
The complaint’s drafters saw this coming. Paragraph 8 hedges that the image was made with generative AI “or digital manipulation that produces the same effect.” That parenthetical is the tell: the statute contains no same-effect clause. If the PAC proves a fully manual workflow, the equivalent deceptiveness of the output does not rescue the claim, because the legislature regulated a production method, not a category of lie. This is the drafting seam the case exposes. A law aimed at synthetic deception keys liability to tool identity, and tool identity is invisible in the finished artifact.
The proof problem
Which button was pressed is a discovery question, and the artifacts are specific. A layered working file with masked generative layers looks nothing like a hand-built composite with dozens of manually blended layers. Adobe attaches C2PA Content Credentials, a cryptographic provenance manifest, to Firefly generative outputs by default, though the metadata rarely survives flattening and export, so its absence in the mailer file proves nothing. Edit histories, vendor invoices, and the designer’s own testimony fill the gap. Forensic detectors that classify images as AI-generated from pixels alone remain unreliable enough that no court should rest an element on them, a limitation consistent with what we found examining how confidently AI systems misreport their own outputs.
The burden allocation therefore decides real cases. If the campaign must prove generative AI was used, it needs the PAC’s production files, which it cannot get before an August 18 primary. If pointing to a photorealistic depiction of an event that never happened shifts the practical burden to the defendant to document its workflow, the statute has teeth on election timelines. Nothing in section 106.145 speaks to this, and whatever the Hillsborough court does at the emergency hearing will be the first data point.
The PAC also contests a second element. Its position, per Florida Politics, is that no one would mistake the mailer for a legitimate photo of Beltran holding an actual sign, and that the image was a photo illustration of what his votes meant. The statute requires that the content “appears to depict a real person performing an action that did not actually occur.” Whether that test is objective photorealism or reasonable-viewer deception is unbriefed territory. Political speech has a long tradition of protected caricature and hyperbole, and the PAC will lean on it. Against that, a photorealistic image of a named candidate holding a specific fabricated sign is at least triably different from an obvious cartoon, and one judge has already found the campaign’s showing sufficient for an order.
Why this small case is the national test
As of January 2026, twenty-eight states had enacted laws targeting synthetic media in elections, per Ballotpedia’s tracking. The designs split into two families: prohibition regimes that ban or create liability for deceptive synthetic content, and disclosure regimes that permit the content but compel a label. The distinction is now the whole constitutional ballgame.
California built the prohibition kind, and it is dead. In Kohls v. Bonta, Judge John Mendez of the Eastern District of California preliminarily enjoined AB 2839 on October 2, 2024 and struck it down on August 29, 2025, holding that the law discriminated by content, viewpoint, and speaker, was not narrowly tailored, and ignored counter-speech as a less restrictive alternative. He had already rejected the companion labeling-mandate-for-platforms law, AB 2655. Minnesota’s deepfake statute is under challenge in a suit brought by X and xAI. The prohibition model keeps losing.
Florida’s law is the other family. Section 106.145 bans nothing. It compels a short factual statement about production method, structurally closer to the sponsorship disclaimers that have survived in election law for decades than to California’s speech restrictions. Compelled disclosure gets a more forgiving First Amendment test than content prohibition, which is exactly why the Beltran case matters: if a disclosure-only statute holds up in live litigation, the twenty-eight-state patchwork converges on the Florida model. If it fails, on definitional vagueness or as-applied compelled speech grounds, states are left with almost nothing that survives.
There is one more structural oddity. Section 106.145(4) enforces the disclaimer with a first-degree misdemeanor, punishable by up to a year in jail and a $1,000 fine, and authorizes any person to file a complaint with the Florida Elections Commission under section 106.25, with expedited hearings. What the statute never mentions is a private right to sue in circuit court. The Beltran campaign routed around that gap through Chapter 86 declaratory judgment and general equity jurisdiction. Whether a court can enjoin speech under a statute whose named enforcement paths are criminal prosecution and administrative complaint is itself a live question, and the PAC’s motion to dissolve gives the court its first chance to answer. A judge granted the order anyway, which either means the vehicle works or means the issue was not yet briefed.
What the record does not establish
Everything about the image’s production is currently allegation and press-reported argument. The complaint pleads generative AI on a hedge. The PAC’s Photoshop claim comes from a brief described by Florida Politics, which this publication has not independently reviewed, and the underlying production files are not public. The court’s order stopping distribution was reported by press and does not yet appear in the docket index we retrieved on August 16, a reminder that docket databases lag filings by days. No responsive pleading from the PAC appears in that index either. The funding allegations connecting the PAC to the Steele campaign are the complaint’s, and Steele’s operation has not been sued or heard from in this action.
It is also worth stating the defense’s strongest version plainly. If the image was hand-composited, the statute does not apply, the order should dissolve, and the campaign’s remedy for a false mailer is the one politics has always provided: counter-speech, plus defamation law if the requirements can be met. If the “appears to depict” element demands that reasonable viewers take the image as authentic, an exaggerated protest tableau may fail it. And if compelled AI disclaimers ever reach a federal court on an as-applied challenge, Kohls shows the judiciary is not reflexively friendly to synthetic-media election laws, even well-intentioned ones.
What happens next
The emergency hearing on dissolving the order was expected before the August 18 primary, and its outcome decides only the injunction, not the statute. Count I’s declaratory claim survives the election either way, which means the definitional question, whether the statute’s generative AI definition reaches the contested workflow, can be litigated on a full record with discovery into the production files. A parallel Florida Elections Commission complaint remains available to any person under section 106.25 and would generate an administrative record on the same facts. Whichever track moves first will produce the first adjudicated reading of a state political AI disclaimer definition anywhere in the country.
The larger lesson is already visible. Legislatures drafting synthetic-media laws keyed to tool identity have built statutes whose central element cannot be observed in the regulated artifact. Every future case under the Florida model starts with the same forensic dispute this one did. Statutes keyed to deceptive effect avoid that problem and die on First Amendment review, as California learned. The Beltran case is the first courtroom where that design trade-off gets cashed out, and it is being decided on an emergency timeline with a primary two days away. This is how the doctrine in our state court AI litigation series keeps getting made: not in appellate grandeur, but in county courthouses, under deadline.
Frequently asked questions
What does Florida’s AI political ad law require?
Section 106.145 requires political ads containing content made in whole or in part with generative AI, where that content appears to depict a real person doing something that never occurred and was created with intent to injure a candidate or deceive on a ballot issue, to state prominently: “Created in whole or in part with the use of generative artificial intelligence (AI).” Print disclaimers must be bold and at least 12 points.
What is the penalty for missing the disclaimer?
The person identified as paying for, sponsoring, or approving the ad commits a first-degree misdemeanor, punishable by up to a year in jail and a $1,000 fine, in addition to civil penalties under the election code. Any person may also file a complaint with the Florida Elections Commission, which must provide expedited hearings.
Does the law cover Photoshop edits?
It depends on which Photoshop features were used. A manual composite built by hand appears to fall outside the statute’s definition of generative AI. Photoshop’s Generative Fill and related Firefly features are diffusion systems that synthesize content, which fits the definition. The statute applies if generative AI made the content in whole or in part.
What happens if the image was not made with AI?
Then section 106.145 does not apply, no disclaimer was required, and the order stopping distribution should not stand on that basis. The statute regulates a production method rather than falsehood itself, so a deceptive image made entirely by hand falls back on defamation law and counter-speech.
Methodology: This analysis is based on the verified complaint in Mike Beltran for Congress v. Conservative Solutions for America, case 26-CA-008578, filed August 7, 2026 in the Thirteenth Judicial Circuit and retrieved on August 16, 2026 from a commercial index of state trial court dockets, together with the text of section 106.145, Florida Statutes, the 2024 legislative history of HB 919, the rulings in Kohls v. Bonta, and contemporaneous reporting by Florida Politics on the restraining order and the PAC’s response. Court filings described here are allegations or arguments, not findings. Nothing here is legal advice. For the closest companion dispute over synthetic depictions of real people, see our coverage of the Omni Hotels AI likeness suit.