The Pattern
On November 21, 2025, the campaign of Representative Mike Collins of Georgia published a video of Senator Jon Ossoff, the man Collins hopes to unseat, appearing to say this: “I just voted to keep the government shut down. They say it would hurt farmers, but I wouldn’t know. I’ve only seen a farm on Instagram.”
Ossoff never said it. The voice was synthesized. The Collins campaign did not deny that. Its defense, reported by CBS News Atlanta, was that the video was “clearly labeled as AI-generated satire,” and that the campaign embraces “cutting-edge technology” to “deliver our message directly to voters.” The label in question was a small on-screen note. Ossoff’s campaign called the video “an intentional attempt to mislead voters,” and added: “The only reason a candidate would need to use a deepfake to make up an opponent’s words is if they didn’t think they could win on their own.”
This review is about one word in that defense. Not “labeled.” A label existed. The word doing the work is “clearly.”
The second label
In March 2026, the National Republican Senatorial Committee released an ad built around an AI-generated James Talarico, the Democratic Senate candidate in Texas. The synthetic Talarico stands before a Texas flag and appears to say, on camera, in his own voice, lines assembled from social media posts he had written years earlier. In the lower right corner sit the words “AI generated,” in a font the Associated Press described as easy to miss.
The committee’s communications director, Joanna Rodriguez, said Democrats were “panicking after seeing and hearing James Talarico’s own words.”
Hold that sentence against the video. The words were his. The voice was not. The face was not. The flag, the framing, the delivery, the confession-shaped performance of a man owning each line: manufactured. A fact-checker who examines the transcript will find real quotes and be right. A viewer who misses the corner text will watch a candidate say these things and be wrong. Both outcomes are built into the same file. The real words launder the fake mouth.
Texas Republicans have already turned the tool on each other. In the Senate primary, Ken Paxton’s campaign circulated an AI video of Senator John Cornyn dancing with Representative Jasmine Crockett, narration implying coordination between them. Cornyn’s campaign answered with an AI ad of Paxton in a convertible beside women labeled as his mistresses. One party, both directions. The tool has no loyalty.
What a label is for
A disclosure can be measured two ways. By presence: whether it exists somewhere in the frame. By function: whether the person about to share the video learned, before sharing, that the confession never happened. The campaigns measure by presence. Pointed to afterward, the corner text converts a fabrication into a technicality. The label proves the label. Nothing else.
Since August 2 of this year, the European Union measures by function. Article 50 of the AI Act requires anyone deploying a deepfake to disclose the manipulation “in a clear and distinguishable manner,” at first exposure, in a form a person can perceive without special tools. The rule keeps a lane for satire and art: disclosure that does not spoil the work, still disclosure. Penalties reach 15 million euros or 3 percent of worldwide turnover. The same ad, delivered to a feed in Dublin, is a legal exposure. Delivered to a feed in Dalton, Georgia, it is a messaging strategy.
The satire lane deserves one more sentence. Satire does its work after the audience recognizes it. A corner label sized to be missed postpones the recognition until the sharing is done.
The second chance
Whether a viewer gets a second chance at the truth depends on the platform, and the largest platform has been reorganizing its second chances. In January 2025, Meta announced the end of its third-party fact-checking program in the United States, part of a global network of more than 90 independent organizations, and replaced it with Community Notes written by users. Internal risk reviews, according to the lawmakers now questioning the company, were increasingly automated. When an integrity call goes wrong now, the answer to who decided is a system. A system cannot be asked to answer.
This week, nineteen House Democrats led by Representative Kevin Mullin of California put those changes into a letter to Mark Zuckerberg. They asked for Meta’s election-integrity staffing and budget set beside its 2020 and 2024 numbers. They asked for the current policies on deepfake video, cloned voices, and impersonation accounts, the disclosure rules for paid political influencers, the removal timelines for deceptive content, the reporting channel a county election official is supposed to use, and evidence that Community Notes works. They cited the Talarico ad by name. The deadline is September 18. Election day is November 3, forty-six days after that.
The letter is signed by one party. The commitments run in more than one direction: after the Collins ad, the Ossoff campaign said it would not use deepfakes that fabricate or misattribute an opponent’s words. The pledge is on the record. So is the practice it answers.
The market
The lawmakers’ letter also cites a case with no candidate in it. For thirteen months, an account called Emily Hart posted on Instagram as a young nurse in New York who loved Trump, guns, beer, and church. She weighed in on immigration, abortion, gun rights, and Christianity. She did not exist. Her operator, a medical student in northern India working under the name Sam, had asked Google’s Gemini how to grow an account, and the chatbot steered him toward a niche it reportedly called a “cheat code”: older, MAGA-leaning American men, loyal and financially capable. Google later said the model should not have offered that opinion unprompted. He built her to the spec. Ten thousand followers inside the first month. Subscriptions on an OnlyFans-style platform tolerant of synthetic content, where AI-generated nudes of her sold under her name. A few thousand dollars a month for thirty to fifty minutes of work a day. Instagram removed the account in February 2026 for fraudulent activity. WIRED published its investigation in April.
Emily Hart was not propaganda in the old sense. No state built her. No ideology steered her. The persona was inventory, and the audience was selected on projected revenue by a consumer chatbot doing what it was asked. The people who followed her were not hated. They were priced.
What the defense covers
The Collins campaign’s defense contains true statements. A label exists in the frame. The campaign acknowledged the fabrication when asked. The NRSC’s ad does contain Talarico’s own written words. None of that is in dispute here, and a review that pretended otherwise would be doing what the corner text does: technically present, aimed somewhere else.
What the defense does not cover
“Clearly.” A disclosure sized and placed so the ad still works on whoever misses it is not a clarity practice. It is a liability practice. The label answers critics and regulators. The fabrication reaches voters.
We offer no evidence that these ads moved votes, and make no claim that they did. We make no claim about parties as such: the documented general-election deployments this cycle trace to Republican committees and campaigns, the primary exchange ran inside the Republican Party, and the record stays open on all sides. And we make no claim against satire. Satire that survives recognition is satire. The mechanism documented in The Deepfake Propaganda Pipeline runs on the interval before recognition, and the corner label sits inside that interval.
The wider pattern has a name on this site. When any image might be synthetic, every authentic image can be denied, and the people who profit are the ones with the most to deny. That ledger is documented in the liar’s dividend, and it paid out again in the case of the President’s fakes. The corner label adds a new line to it: the fabrication now carries a receipt, and the receipt is the alibi.
Verdict: The claim that the fabricated candidate ads were “clearly labeled” is Misleading. The labels were present. Presence is the entire truth of the claim, and the claim needed more than presence. Since August 2, European law measures a label by whether a person perceives it. The campaigns measure a label by whether it can be pointed to afterward. A disclosure the defense can find and the viewer can miss is not disclosure. It is the fabrication’s paperwork.
Image note: the scratch-to-reveal illustrations above are AI-generated editorial art by Deceit, made for this piece. The figure is a composite; no real person is depicted. Not documentary evidence.
