Cheap AI ‘deepfakes’ give new twist to US political attack ads

Looking up from his Bible, James Talarico turns to the camera. “I always think of myself as a Christian who hates Christianity,” says the Democratic Senate nominee from Texas. “America,” he continues, “is not a Christian country”.

Talarico’s comments — some of which he has recanted — are real. But the television advertisement in which he appears to proudly recite them is generated by AI and paid for by his Republican opponent Ken Paxton.

The ad, in a crucial US Senate race, is one of the “deepfakes” that have been appearing on American airwaves in the run-up to the November midterms, the first major US election in the age of cheap and convincing generative AI.

Talarico, a clean-cut former teacher and Presbyterian seminarian, has been depicted reading old social media posts about how “radicalised white men” are a terrorist threat and handing out ice creams in a van advocating “permanent sex changes for children”. He even appeared in one attack ad as a masked doctor apparently intent on mutilating a young girl.

Talarico is not the only politician whose image has been manipulated by AI. One is depicted as a menacing Batman “Joker” while others are shown getting expensive beauty treatments, dancing in a keffiyeh, going down a waterslide in a Hawaiian shirt and welcoming Vladimir Putin and Iranian clerics into their homes.

“We’re seeing a lot more [AI ads] than we have ever seen before,” said Erika Franklin Fowler, a director of the Wesleyan Media Project, which last month identified 164 political ads — “almost certainly an undercount” — generated or enhanced with AI, on which roughly $80mn was spent.

They run the gamut from obvious satire to deepfakes “where you have candidates saying or doing things that they didn’t say or never did”, she added. “It’s not always clear to viewers . . . that the images being depicted are not real.”

More than 30 US states have legislated to restrict AI use in political advertising or to force disclosures when it is deployed. But many of these laws have been challenged as unconstitutional, and there is uncertainty over whether they govern ads run in support of or against federal candidates.

As a result, almost 70 per cent of the ads found by Wesleyan did not have a disclaimer that disclosed the use of AI tools, according to researchers.

Some campaigns have been relying on the technology to generate ads that would once have been prohibitively expensive.

One of the first AI ads in this cycle was a 30-second spot created by FP1, media consultants to Alan Wilson, a Republican running to be governor of South Carolina. It featured a fictional TV game show in which contestants were asked to decide in front of a large audience whether another conservative candidate had criticised Trump or “extreme” anti-abortion measures.

“To go ahead and stage that, to fill an auditorium . . . and have actors playing contestants and the host, that’s a six-figure proposition,” FP1’s creative director Trent Wisecup told consultants at an American Association of Political Consultants (AAPC) event this week. “And we did it in an afternoon with AI.”

However, despite a steep increase in the use of AI in political advertising — mostly by Republican groups, no doubt taking a lead from their president, who frequently shares AI-generated videos online — the practice has not become as widespread as many in the field had feared.

“I think the fears were overblown,” said Eric Wilson, a Republican strategist. AI was largely being used for “comedic effect”, he added, citing a video in which a progressive candidate is depicted in a yoga pose, tying herself in knots.

Most of the AI-generated ads that had featured in midterm races, Wilson claimed, were a version “of a type of ad we’ve been running for years” in which an opponent is attacked with a selection of their own comments.

“We’re just updating that with new technology to now be able to depict the person saying the thing that we knew that they said.”

Wilson’s main concern is the use of deepfakes by foreign adversaries to influence the American electorate. An online group known as Antibot4Navalny has identified social media videos they attribute to Russian disinformation network Matryoshka, in which celebrities appear to wade into political debates.

In one, the actress Sarah Jessica Parker appears to say “Democrats should never hold power in our country”. Another features a lifelike Julia Roberts saying she does not support the party’s “destructive gender policies”.

There is no evidence that the videos have been widely viewed or had much impact, but AI “helps [adversaries’] playbook of sowing chaos”, Wilson said.

Creators of AI-generated videos, whether or not they have malign intent, are hitting roadblocks. Many mainstream AI tools now prohibit the generation of famous figures, and there is growing evidence that voters themselves are put off by the use of AI in ads.

In a survey run on behalf of the AAPC in September, almost 75 per cent said they thought it was unacceptable for AI to generate images or video in a political ad. More than 90 per cent said they were concerned that deepfakes could change how other people voted in November.

FP1’s Wisecup, who has ruled out creating deepfakes that put words in candidates’ mouths, said the AI disclaimers that some states and social platforms mandate had become “kind of like a siren to voters to be suspicious of [an ad]”.

Recommended

Campaigns “are looking at the public unease with AI and are saying: ‘well, you know, I don’t want to put AI in my ads’”, he added.

While the chair of Republican National Committee chair has defended the use of artificial intelligence in some attack ads, Democrats who have been targeted by deepfakes are also making a virtue of not using AI themselves.

In an ad released this month, the real-life Talarico accuses his opponent in Texas of lying about his record in “fake” AI videos.

“No, I’m not a member of the Chinese Communist Party. No, I don’t support sex-change surgeries for kids,” Talarico says in the pared-back clip. “I don’t have to use AI, because the truth about Ken Paxton is bad enough.”