Every year since 2018, people have had one driving question for artificial intelligence expert Henry Ajder: Is this the year deepfakes will disrupt an election?
Now, as deceptive AI-generated videos, still pictures, voice recordings, and other media proliferate during the presidential race, 2024 seems to have provided an answer.
“They are already playing a role,” said Ajder, who runs Latent Space Advisory, a UK-based consulting firm. “The question now is how persuasive and how effective will they be.”
In recent months, a host of fraudulent deepfakes have circulated online and beyond: Robocalls of President Biden urging New Hampshire voters not to vote; images of former President Donald Trump surrounded by Black supporters; Tom Hanks hawking a dental plan; sexually explicit videos of Taylor Swift.
Deepfakes of politicians, celebrities, and ordinary people may seem unrelated, but their overall effect is to raise people’s awareness of the technology and its potential to blur the line between what is real and what is fraudulent. It’s a troubling dynamic in an increasingly polarized age, as deepfakes inject falsified media into the information ecosystem while also making it easier to dismiss legitimate media as fake, a phenomenon known as the “liar’s dividend.”
“These technologies are going to fundamentally undermine what many people see as one of the last bastions of shared reality,” said Ajder, “this idea that if you have a video of someone doing something or saying something that it is uncontrovertibly real.”
The rise of deepfakes comes at a precarious time, as trust in traditional media remains low and social media companies such as Meta and X have recently curtailed efforts to rein in disinformation. Meanwhile, Washington has yet to enact a unified legal framework to regulate deepfakes, and so-called deepfake detectors remain spotty at best.
It’s an anxiety-inducing moment, said Lindsay Bartholomew, exhibit content and experience developer at the MIT Museum.
Untangling that anxiety lies at the heart of “Detect a Fake,” one of several installations at the museum’s “AI: Mind the Gap” exhibition, which explores the social, cultural, and political implications of deepfakes and other forms of generative AI.
“It’s meant to address the idea that technology can manipulate what we perceive as true or false,” said Bartholomew. “But you also need to appreciate what you can do as a human, you have some agency here.”
The “Detect a Fake” installation presents a series of videos, prompting viewers to decide after each clip if what they’ve seen is authentic or a deepfake. Choose incorrectly, and the module will reveal the deepfake’s specific tells: odd reflections, funny shadows, lips that are out of synch.
The exhibit draws on the work of AI researcher Matthew Groh, an assistant professor at Northwestern University’s Kellogg School of Management who studies how people engage with AI.
“Part of what I’m trying to do is just get adults to be better at critically reasoning about synthetic media,” said Groh, who earned his PhD from the Massachusetts Institute of Technology. “There’s a lot of potential for us to get better at it. The question is, how?”
For starters, he said, it’s essential to understand how deepfakes are made. High-quality videos, such as those featured on the @deeptomcruise TikTok account, are relatively rare and hard to make, needing a lookalike actor, a technically savvy AI- and visual-effects artist, and a lot of time.
Compare that with simpler face-swapping videos, which place someone’s face on someone else’s body, or lip synching, which manipulates the mouth so it tracks with new audio. Many AI-generated still images continue to harbor telltale artifacts, misshapen hands, odd shadows or reflections, strange blurs, and other visually confusing elements that don’t make sense in the real world.
Groh said it’s unreasonable to expect us all to become digital forensics experts, but simple critical thinking about the media we encounter can go a long way toward ferreting out misleading AI-generated content.
“Consider the source,” he said. “Why are they sharing this with me? How might they want to change my beliefs? If you can just ask those kinds of questions, that’s going to do the majority of the work.”
Todd Helmus, a disinformation expert with the Rand Corporation, said we’re entering a “period of risk” where “people are not very savvy about this, and the technology has outpaced their savviness.”
“What’s probably going to be most important is combining initiatives,” he said, “so combining detectors with open source intelligence techniques and good journalism to identify quickly whether an emerging video that’s disputed makes sense.”
President Biden signed a sweeping executive order in October developing safety standards for AI. The order calls upon AI companies to share safety test results and calls upon the government to formulate guidance standards to label AI-generated content, among other measures.
Congress has introduced a host of bills to address dangers posed by AI, while a number of states have already passed legislation. In Massachusetts, state legislators are considering a number of bills, including one by State Senator Barry Finegold that would require a disclaimer for any AI-generated media that depicts a candidate.
“We’re not saying you can’t do it,” said Finegold, noting the First Amendment’s expansive protections for political speech, particularly when it concerns public figures. “What we’re saying is that within 90 days of an election, you need to have a disclaimer, just like a disclaimer you see at the end of a political advertisement on TV.”
Meanwhile, deceptive deepfakes continue to go viral: Pornographic deepfakes of Swift racked up millions of views on X before the social media site suspended the offending account in January. That same month, a robocall with a voice that sounded like Biden made as many as 25,000 calls urging voters in New Hampshire to avoid the polls, according to the call-monitoring service Nomorobo.
Last month, a doctored image released of Catherine, Princess of Wales, brought weeks of conspiracy-fueled speculation about her whereabouts to a crescendo.
“We have to be honest with people and say, ‘Well, look, increasingly, trusting what you see by default is not a prudent thing to do,’” said Ajder. “At the same time, we don’t really have a foolproof way out of the problem.”
Malcolm Gay can be reached at malcolm.gay@globe.com.