In August, the reëlection campaign of the North Carolina state senator Michael Lee, a Republican, released an ad attacking his Democratic opponent, a yoga-studio owner named Jessica Bichler. The ad, titled “Bonkers Bichler,” seems to show Bichler, a middle-aged woman with graying straight hair, doing yoga in a well-equipped, gently lit studio. The figure onscreen is not Bichler, though; it’s a hyperrealistic, animated A.I. rendering of her. The next frame shows her body twisting into an impossible knot, even as she maintains a vacant-eyed grin. Bichler told me recently that, when she first encountered the ad imagery, “I was thinking, who was even advising on this? Because it was really bizarre.” Five seconds into the ad there is a written disclaimer noting that A.I. was used to make “these silly video clips,” but the footage still deceived some local voters, who asked Bichler why she had let Lee into her yoga studio. With A.I., “people can make any image of you doing anything that they want you to be doing, and put that out in the public sphere,” Bichler said. “There is just a huge vulnerability.” (There was a whiff of hypocrisy about the ad, too: Lee and his wife have been known to frequent yoga studios themselves.)

Bichler has noticed apparently A.I.-generated imagery in other North Carolina election ads, including ones targeting the state Senate candidates April Cook and James Gailliard. Even Lee himself has been a target of a kind of impersonation; according to Bichler, a group not connected with her campaign has put up cardboard cutouts of the senator gripping handfuls of money and wearing a shirt that says “I <3 special interest.” (He also appears to have peed his pants.) Whatever the outcome of the races on Tuesday, this year’s midterms season marks a tipping point for A.I. in campaigning. AdImpact, an organization that tracks political advertising, has identified fifty-eight unique ads across the country that deployed A.I.; more than ten million dollars has been spent to air them. (Lee’s campaign spent a total of more than a hundred thousand dollars on “Bonkers Bichler” and two follow-up ads.)

This is not the first election cycle to be touched by misleading A.I. content. In January, 2024, thousands of New Hampshire voters received a recorded message from a voice that sounded just like then President Joe Biden, discouraging listeners from voting in the local primary election. In reality, it was a deepfake created by Paul Carpenter, a magician from New Orleans, using the audio-generating A.I. tool ElevenLabs, on behalf of a consultant associated with Dean Phillips’s long-shot Presidential campaign. But, as A.I. models have grown more powerful, more accessible, and more affordable, releasing fake messages from one’s opponents has become something of a default political strategy. In October, 2025, Andrew Cuomo’s mayoral campaign posted a series of A.I. spots targeting Zohran Mamdani: one depicted him as an “Austin Powers” Mini-Me to Bill de Blasio; another showed a silent Mamdani running through the streets of Manhattan with a giant set of keys and messily eating rice with his hands. Like the infamous “Trump Gaza” slopaganda video that the President shared earlier last year, these were blatant and shoddy A.I. creations, with blurry, color-saturated backgrounds and overly smooth faces. The Cuomo videos were criticized as racist and quickly taken down, but they proliferated online regardless, playing a role similar to that of extreme political cartoons: exaggerated and offensive provocations that could be played off as jokes.

Today’s A.I. ads are more easily mistaken as real. Compare the Mamdani content to a recent attack ad against the Pennsylvania governor, Josh Shapiro, by his opponent Stacy Garrity’s campaign. An A.I.-generated Shapiro, outfitted in a blue suit and casually tieless white shirt, with realistically coiffed hair and facial wrinkles, seems to testify to “deleting incriminating e-mails” and supporting “boys in girls sports.” There are no six-fingered hands here as a tell (though the imperfect voice replication might be a hint.) Lambasting an opponent’s record in office is the bread and butter of political attack ads, but candidates used to have to put their criticisms into the mouths of actors playing salt-of-the-earth voters—a plumber, say, or a factory worker. Now they can deposit any unflattering idea into the A.I.-generated mouth of the opponents themselves. Manuel Bonder, a spokesperson for Shapiro for Pennsylvania, described the ad to me as “cheap, dishonest AI bullshit” and pointed out Garrity’s support for “aggressive data center deregulation.”

Two-thirds of the A.I. ads identified by AdImpact are from Republicans, which might explain why they often fixate on aspects of personal identity, enlisting an opponent’s likeness to help depict him as a reductive trope. One such attack ad shows the Michigan U.S. Senate candidate Abdul El-Sayed playing a Risk-type board game with his sister and the left-wing streamer Hasan Piker, flashing cards saying “DEATH TO AMERICA.” “The destruction of Israel and the West isn’t just a game for them, it’s a mission,” a voice-over narrates. Another depicts James Talarico, a Democrat running for U.S. Senate in Texas, wearing a T-shirt supporting trans people and eating tofu cubes at a football game, presumably to mock his masculinity and build on false rumors that he is a vegan. (The Talarico campaign said in a statement that his opponents are “slandering him with AI deepfakes, foreign bots, and conspiracy theories.”) Yet another video depicts Senator Jon Ossoff of Georgia in a salon, getting his hair done and his nails filed, as a fictional illustration of the “makeup services” that Ossoff has reportedly paid for while in office—a sign, the ad alleges, that “he’s hiding something else.” In the ad, Ossoff appears perfectly real, as if he’d signed up for a walk-on role in his opponent’s campaign. Only a careful viewer would notice that the technicians seem to be filing his hands, not his nails.

As with the A.I. agents spreading across the internet or the A.I. software assisting in warfare, A.I. imagery in politics has gone largely unregulated. The 2025 Take It Down Act requires online platforms to remove sexually explicit deepfake content, but machine-generated political ads are treated as examples of free speech, granted the same protections as political cartoons. There have been some efforts toward halting the flood; after the 2024 fake-Biden robocall, Steve Kramer, the consultant who commissioned the ad, was ejected from the campaign and faced criminal charges of voter suppression and impersonation of a candidate, as well as millions of dollars in fines from the Federal Communications Commission. (He was later acquitted.) Not long after the fake-Biden phone message, the F.C.C. banned A.I.-generated voices in robocalls. Jessica Bichler’s campaign sent a cease-and-desist letter to Michael Lee’s campaign on the ground that the A.I. ad misrepresented her affiliation with a nonprofit. In regard to the actual A.I. impersonation, “there are no laws on the books to say he can’t do that,” Bichler said, noting that Lee voted to table an amendment to a state bill which would have required disclosure for generative-A.I. use in media. Bichler worries that even voters who know that such ads are unreal will be subtly influenced by their imagery. On the algorithmic internet, voters already live within their own bubbles and struggle to find common ground. A.I. attack ads provide further fuel for that fragmentation, Bichler said: “We’re talking across each other, but oftentimes we’re talking with such conviction. We feel like, ‘I’ve seen this with my own eyes.’ ” ♦