Tesla's lawyers said Elon Musk's recorded words might be a deepfake

How the liar's dividend got its name

Before the 2016 election, a technologist named Aviv Ovadya gave a talk to Bay Area colleagues that he called "Infocalypse." He was not warning about a particular fake. He was warning that the tools to fabricate convincing audio and video were arriving, and that once they did, the damage would not stop at any single forgery. The question he kept returning to, in the reporter Charlie Warzel's account, was "what happens when anyone can make it appear as if anything has happened, regardless of whether or not it did?"¹ His subject was synthetic media: audio, images, or video produced or altered by software rather than captured from the world.

Ovadya put the stake plainly in his own writing. In the Harvard Business Review in 2018, he argued that the coming wave "won't just impact our elections. It will impact our ability to trust just about anything we see and hear."² A year later, in MIT Technology Review, he and a co-author held that the same tools that entertain "can also enable deception with global ramifications." They rejected the shrug that there has always been deception. What had changed, they argued, was the degree and the impact.³

The sharpest version of the worry was named in 2019 by two law professors, Bobby Chesney and Danielle Citron, in a California Law Review article, "Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security."⁴ A deepfake, in their sense, is synthetic media that convincingly shows a real person saying or doing something they did not. Their article named two distinct dangers. The first is the flood of fakes: convincing forgeries produced cheaply enough to deceive people in bulk. The second is subtler, and it was their own contribution. Fakes do harm even when no fake is made. Once the public knows a recording can be faked, anyone caught on a real recording can escape it by claiming it is fake. They called the payoff the liar's dividend, and they noticed a trap inside it: the better the public understands deepfakes, the more readily that denial is believed.

What the 78 catalogued AI uses in 2024's elections turned out to be

Dozens of countries held national elections in 2024, the first such year with cheap generative tools in general reach. The prediction was a flood. Two Princeton researchers, Sayash Kapoor and Arvind Narayanan, read every one of the 78 uses of AI in elections the WIRED AI Elections Project had catalogued.⁵

Of the 78, 39 carried no intent to deceive: satire, labeled campaign art, illustration. Of the 22 uses by candidates or supporters, 19 aimed to improve campaign materials rather than mislead. The genuinely deceptive examples could mostly have been made without AI for a few hundred dollars. And the fakes that spread furthest were not AI at all but cheap fakes, meaning real footage relabeled or crudely edited by hand. By the News Literacy Project's tally, cited in the same analysis, cheap fakes ran to roughly seven times the AI ones.⁵ That figure is theirs, and I am reading it secondhand. No study of the year, theirs included, found a single AI fake that measurably moved a result.

Kapoor and Narayanan are the serious skeptics of the whole alarm. Their point is that the flood was never the real mechanism. Misinformation, they argue, is pulled by demand rather than pushed by supply. People seek out claims that flatter what they already believe, so producing such claims more cheaply changes little. Whether an AI can actually out-argue a person is a separate question, and an earlier lesson took up the one trial that has tested it.

Their own honesty cuts the other way. Their count of deepfakes was small. That "could either indicate that AI deepfakes are a much smaller problem so far than anticipated," they note, "or that the database has many missing entries."⁵ So the case rests on a tracker that may undercount. But undercounting cannot rescue the strong prediction, which was about swung votes, and no measured swing has turned up.

In court, the "it's a deepfake" defense has already been tried

The second worry has fared differently, because it does not need a flood. It needs one person with something to hide and a way to call a recording fake. That move has already reached a courtroom. In Huang v. Tesla, a wrongful-death suit over a fatal 2018 Autopilot crash, the plaintiffs wanted to question Elon Musk about recorded 2016 statements in which he touted Autopilot's safety. Tesla's lawyers argued the recordings might be deepfakes, and that Musk therefore should not be compelled to sit for questions about them.⁶ Judge Evette Pennypacker refused. She called the argument "deeply troubling," and read their position as holding that "because Mr. Musk is famous and might be more of a target for deep fakes, his public statements are immune." She ordered a narrow three-hour deposition, limited to authenticating whether he had said the words. This is the liar's dividend in the exact form Chesney and Citron feared: real, recorded words waved away as possibly fake to dodge accountability.

A looser instance came from politics. In August 2024, Donald Trump posted that a photograph of the crowd at a Kamala Harris rally was AI-generated, and that no crowd had been there. The rally drew about 15,000 people by the campaign's count, documented by wire agencies. The forensics researcher Hany Farid found no sign the image was generated, and PolitiFact rated the claim its lowest, "Pants on Fire."⁷ It fits the pattern, but differs from the Tesla case: the photo showed a rival's crowd, not Trump's own recorded conduct, a weaker fit for the paradigm of escaping one's own acts.

The tactic is also older than deepfakes. The Brennan Center's review records earlier instances of the same denial. A Spanish minister dismissed real images of police violence in Catalonia as "fake photos" in 2017. Trump himself called the Access Hollywood tape "not authentic" in 2017.⁸ I am reporting those as the Brennan Center reports them, not from each underlying record. They change the reading of Chesney and Citron's claim in one way. Deepfakes did not invent the cry of fake. They lower its cost and raise its plausibility, a change of degree rather than a new move. The Brennan Center's own answer is to make authenticity easy to prove: provenance standards that tie a file to a verifiable origin, newsrooms that pre-empt fakes, and norms against candidates who falsely cry fake.⁸

The one experiment found it weakest where the fear is sharpest

The move being tried is not the move working. The Tesla judge rejected it, and the Trump claim was debunked. Neither tells us whether the denial changes what audiences believe. Only one controlled study has tested that. Kaylyn Jackson Schiff, Daniel Schiff, and Natalia Bueno ran five pre-registered survey experiments with more than 15,000 participants. Each showed people a politician hit by a scandal, then a false claim that the evidence was misinformation.⁹ They tested two versions of the denial: informational uncertainty, which says the evidence might be fake and no one can be sure, and oppositional rallying, which says enemies fabricated it. Both beat apologizing or staying silent, and both worked across party lines.

The payoff turned on the scandal's format. The denial raised a politician's support when the scandal was reported in text. It was largely ineffective when the scandal came with video.

Chesney and Citron feared the dividend most for genuine audio and video, the material deepfakes can imitate. The only experiment to test the tactic found it worked least well there. In the authors' summary, "false claims of misinformation do pay a liar's dividend, but primarily (if not exclusively) when used against scandals reported in text format."⁹ The Tesla and Trump episodes show the move being attempted. This study is the reason not to assume it lands on an audience.