Deepfake labels won't fix election ads. Accountable sponsors might.

A giant generative AI version of Jair Bolsonaro, Brazil's former president, told a São Paulo rally the truth about itself. "What you see here is not me," the avatar said at the launch of his son Flávio's presidential bid, The Next Web reported. "This is a simulation of my image and my voice using artificial intelligence." Then it campaigned for Flávio, in the name of a man serving twenty-seven years for plotting a coup, and a perfect label did nothing to slow it down.

What it did draw was the who-question. Supreme Court Justice Alexandre de Moraes gave Bolsonaro forty-eight hours to say whether he had authorized the video, and raised a possible return to prison. His lawyers said he hadn't, which makes who was speaking an even sharper question.

Ireland experienced the reverse. Days before its 2025 presidential vote, AI-generated RTÉ News bulletins falsely said Catherine Connolly had withdrawn from the election. An analyst told RTÉ that similar videos carried a "synthetic" label on YouTube and none on Facebook. The questions Ireland's electoral commission gave voters included "Who is telling me this? Why are they telling me this?"

Every AI label answers one question: "how was this made?" Voters decide what to believe with a different one: "who sent it?" As The Next Web put it: "A label tells you a video is fake. It does not tell you whether a jailed politician should be allowed to speak through it."

The label that mostly didn't show up

The midterms were on 3 November, and the "how" label mostly sat this cycle out. By mid-October the Wesleyan Media Project had found at least 220 ads made or enhanced with AI, across thirty-four states. Its report doesn't break them down by office. News reports describe AI ads in a Kentucky House primary, by the losing candidate's account, and in Tennessee's governor's race.

Only 27% carried a disclaimer about AI, and no federal rule requires one. That count includes uses beyond deceptive impersonations, so it's not a measurement of attempted deception. Spending behind them topped $137 million by Wesleyan's count, up from less than $50,000 in 2024.

Where states legislate, they mostly choose the label. The National Conference of State Legislatures counted thirty-one states with election deepfake laws in June, and twenty-eight rely on disclosure. "The fabricated ad can remain in circulation with that warning," writes Stefani Langehennig, an AI policy scholar at the University of Denver.

Labels do have some impact. An American Association of Political Consultants Foundation brief, quoted by the Plain Dealer, called the disclaimer "a cognitive speed bump" that made viewers more skeptical. In a study Langehennig cites, labels lowered belief, but warnings that simply identified content as AI-generated "did little to change how willing people said they were to engage with it, including sharing it." She notes the research didn't test whether disclosure laws change votes.

The foundation's poll of 1,568 likely voters, reported by the Washington Post, shows what people judge. Around nine in ten called an AI video of an opponent saying something they never said "completely unacceptable," regardless of whether it was disclosed. They'd made up their minds about the act.

Openly AI, and it still works

Openly AI-generated content gets its power from whoever sends it. In December 2023, the opposition Pakistan Tehreek-e-Insaf (PTI) closed a virtual rally with a four-minute AI voice clone of its jailed figurehead, Imran Khan, captioned as the "AI voice of Imran Khan based on his notes." AFP and Reuters reported more than 4.5 million views for the rally across Facebook, X and YouTube.

In Turkey, the opposition Cumhuriyet Halk Partisi (CHP) gave its jailed Istanbul mayor, Ekrem İmamoğlu, an AI voice at a March 2025 rally in Maltepe, BirGün reported, without saying whether the video carried a label. The message said he wasn't afraid and wouldn't back down. Like Bolsonaro's avatar, these drew their authority from the jailed figure they spoke for.

Fakes aim at the same signal. In Bangladesh's 2026 election, Global Voices reported on a study of seventy-two fact-checked AI cases, and 28% put false quotes in politicians' mouths through photocards mimicking trusted outlets. Seeing the quotes under several trusted brands, the piece's author writes, "made the fabricated quotes more credible than isolated false claims." The RTÉ bulletins worked the same way. What they forged was a newsroom's face.

Shows AI Generated Imagery

1 of 5

The fake Biden robocall that hit New Hampshire in January 2024 forged two senders at once. Steve Kramer, a Democratic operative, cloned Biden's voice and spoofed the phone number of Kathy Sullivan, a former state Democratic official, CyberScoop reported. The FCC fined him $6 million, and a jury acquitted him of all the criminal charges in June 2025, the Associated Press reported.

A name is all the law can ask for

The law that asks the right question already exists. Ohio requires candidates and campaign committees to use the words "paid for by" followed by the name of the entity, and it bars a "false, fictitious, or fraudulent" name. Little in it makes the name tell you anything.

Sometimes the line never arrives. The Ramaswamy campaign's ad featuring Amy Acton's estranged mother leaked without one, the Statehouse News Bureau reported, and Acton opponents and Ramaswamy supporters circulated it within hours, before the campaign confirmed it was theirs.

Whether a leaked copy breaks the rule is an open question. Ohio Secretary of State Frank LaRose says his office is reviewing complaints about the ad, WSYX reported, and he wouldn't comment on the specifics. Straight Arrow News put it plainly while the sponsor was still unclaimed: "Who is behind the video matters, both ethically and legally." The who-question applies however an ad was made.

Sometimes the name arrives and says very little. In Ohio's governor's race, Signal Ohio reported, the Democratic Governors Association backs a group it calls Ohioans for Lower Costs, and the Republicans' offshoot is RGA Right Direction PAC, three letters and a slogan. One of Ohio's first AI ads, which put Sherrod Brown's face on an actor, came from the Ohio Flyer PAC, and Federal Election Commission records tie that PAC to Sen. Jon Husted's joint fundraising committee, the Plain Dealer reported.

Europe shows where the line stops. The EU requires political ads to say who paid for them, among other rules, and Meta's answer was to stop running them there, citing limits on targeting and legal uncertainty. Politicians can still post organically: "They just won't be able to amplify this through paid advertising." The sponsor line covers the paid slot, and sharing goes on without it.

The vouch does the convincing

In a social feed, belief follows the person who shared it. In a Media Insight Project experiment with 1,489 American adults in late 2016, people saw a Facebook-style post of a diabetes story. It was shared by a public figure who often talks about health, one they'd already rated as trusted or not. Trust in the post was "determined less by who creates the news than by who shares it."

A trusted sharer with an unknown outlet beat a reputable outlet shared by someone untrusted, 49% to 32% on whether it got the facts right, and readers were readier to pass it on (39% to 23%). About half could recall who shared the story, but only about two in ten could name where it came from. That's news rather than ads, and it's ten years old.

When people stop to check, they say they go to outlets they already trust. In the Reuters Institute's 2025 Digital News Report, 38% said they'd look there first, ahead of official sources (35%), fact-checkers (25%) and social media (14%). In the US, that meant each side's own: CNN on the left, Fox News on the right.

The same channel carries bad sources just as smoothly. In July 2024, a fake video using a clone of Kamala Harris's voice went up on X labeled as a parody. Elon Musk reposted it eight hours later with "This is amazing," a laughing emoji and no such label, PetaPixel reported, and NBC News put his post at more than 150 million views. Reactions split over whether it was parody, and neither report measures what viewers believed, but the audience came through the sharer and the label stayed behind.

Jimmy Narang, describing his WhatsApp-style experiment in a UC Berkeley interview, says receivers "update as if stories shared by their friends are ~66% likely to be true," while people share true and false stories at about equal rates, and about a quarter of stories get shared even when the sharer gives them less than a 50% chance of being true.

So forwarding something doesn't equal vouching for it, regardless of whether it's true or false. I previously made this argument about testimonials: proof got cheap, and what still counts comes from someone with credibility at risk.

If labels are going to be required, I think the "who" label is the one worth strengthening. An Election Commission of India advisory for the Bihar vote already joins the AI label to the name of the entity responsible, and Full Fact proposes one UK marker naming the promoter, on whose behalf it's published, and that it was altered. None of the research here shows that changes a vote. The sharer research points at the forward more than the sponsor, and the sponsor's name is the part of that chain a law can reach.

A label tells you how an ad got made. You'll decide whether to believe it by who sent it, so the sender's name is the one that has to cost something.