Political ads made with AI break into the midterms campaign

The ad opens with James Talarico, the Democratic candidate for the U.S. Senate from Texas, seated at a desk, apparently rewriting the Book of Genesis from the Old Testament. “In the beginning,” Talarico says, hunched over as he writes, “God was nonbinary. And was not Christian. And God said: ‘Let there be six genders’; and there were.”

Although the voice and image appear real, the scene is implausible. The video — distributed by the campaign of his rival, State Attorney General Ken Paxton — quotes remarks Talarico did make in the past, but places them in an unbelievable context. There are also odd details, such as the candidate starting by writing with a pen and ending by drawing his own crowned face in crayons atop an open Bible that appears to be on the first page. Doubts dissipate at the 20-second mark, when a notice at the bottom of the screen clarifies: “Certain video content in this ad was generated with artificial intelligence.”

This is just one of more than 200 political ads generated with AI that have circulated in the United States during the midterm election cycle, according to a count by the Wesleyan Media Project (WMP), a university research lab that tracks political advertising spending in the country.

“We are adding new ads we identify as AI-generated to our database virtually every day,” Travis Ridout, one of WMP’s directors, says by video call. In early September the total stood at 164 ads and an estimated $80 million in spending; a month later Ridout estimates the number surpassed 200 and there was more than $20 million in additional spending, bringing the total to over $100 million. And the weeks of most aggressive advertising are still to come.

Those millions of dollars are a drop in the bucket compared with overall campaign ad spending, which this week was estimated at more than $10 billion, according to media-tracking firm AdImpact. The figure includes both official campaigns and outside interest groups, primarily, but not exclusively, political action committees, which face looser spending and reporting rules.

Ridout notes that it is precisely the latter groups that most commonly appear to be behind AI-generated commercials, although “there are many different campaigns using this tactic.”

That relatively low spending reflects that AI is an accessible, inexpensive tool that allows less-resourced candidates to produce creative ads without the high costs of a traditional shoot. Experts say it remains to be seen whether this will truly level the playing field in political contests.

What is already evident is that use is strikingly varied. Some campaigns have generated entire videos, including deepfakes — hyperrealistic impersonations of real people — of candidates; in others, only certain scenes or shots are AI-created. But there can also be uses that are virtually undetectable, such as certain image or sound editing processes.

For example, in an ad where Donald Trump cooks a burger bearing a congressman’s face, the manipulation is obvious, but there are other edits and image tweaks that can be practically imperceptible.

That makes WMP’s count only an approximation, since identification of ads that use AI is done manually and campaigns do not always flag them explicitly. The tally also excludes material produced by individuals unaffiliated with campaigns that circulates on social media and sometimes goes viral.

Ridout considers the latter “quite impossible to measure,” because unlike paid ads there is no spending to quantify it, so there is no way to know whether those who produce it follow ethical standards different from campaigns. Still, it helps illustrate how a political communication phenomenon is spreading whose effects are not yet understood.

A bipartisan majority of Americans holds a negative view of AI’s rapid advance, according to several polls conducted in recent months. And regarding AI-generated ads, many say they should carry disclosures, Ridout says. But when his team showed people, in small experiments, real ads with deepfakes, “almost everyone could tell it was AI-generated.”

His team also ran a test suggesting that AI-made ads do not seem to affect trust in political advertising overall, which has been one of the major fears since the first deepfakes emerged. Ridout cautions that the experiment is limited and it is too early to draw definitive conclusions about societal-level effects.

What worries him is something else. While cruder impersonations are obvious, what happens after people see them? “What we don’t know is whether a voter sees that and knows it isn’t real, does it still have an impact on how they think?” he asks. “‘I know it’s not real, but it sticks in my head and can affect how I behave afterward.’ That’s the real question for me.”

The phenomenon is so new that there is not yet federal legislation on the matter, although 31 states have laws that regulate AI-generated political advertising in different ways. In some states it is mandatory to disclose that an ad was made with AI and, in three, there are certain bans on deepfakes; but the patchwork of dozens of statutes has been behind several controversies in recent months.

In Tennessee, that ambiguity has been laid bare. Since July 1, a law, passed unanimously by both chambers, requires that ads using AI include a notice at the beginning and the end of the ad.

This summer, in the Republican primary for the 6th District, Johnny Garrett released an ad showing his rival, Van Hilleary, receiving a wad of cash from billionaire Michael Bloomberg. Hilleary responded with another ad showing Garrett sleeping on duty and attending a gay pride parade. Neither included an on-screen AI notice. Both campaigns argued the state law did not apply to them because they were competing for a federal office, although the statute does not explicitly clarify that point.

In the state’s gubernatorial race, an outside group attacked Republican Marsha Blackburn with an AI ad that showed her alongside members of a pharmaceutical lobby without a message declaring the technology’s use. In that case, her campaign sent a cease-and-desist letter alleging defamation, and several television networks pulled the ad.

The federal gap, however, does not appear close to being filled, even though Ridout believes campaigns themselves would welcome a national rule: “Everything would be more consistent about what they can and cannot do,” he says. Attempts to ban deceptive AI-made ads or to require disclosures have not advanced in Congress or at the Federal Election Commission (FEC).

Moreover, experts say regulating political speech faces major legal hurdles, especially when commercials reuse words the candidate actually said, as in Paxton’s ad against Talarico.

For now, regulation has been, in a sense, self-imposed. Ridout says he is not seeing many truly deceptive deepfakes, and assumes the average voter, on seeing an ad, is able to tell it is not real. In other words, implausible and visibly false montages like the Talarico ad abound. “I think there’s a fear of backlash if you go too far with this, and maybe that’s why we see it more in outside groups, which are held less accountable than the candidates themselves.”

Humor has also become a shield. Many ads resort to satire or parody, a form of speech with greater legal protection that, in general, is exempt from the notice requirements of state laws. “Campaign lawyers are happier if they can label something as satire,” Ridout explains. It is less clear whether audiences find it funny.

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition