On January 21, Patricia Gingrich was about to sit down to dinner when her landline rang. New Hampshire voters picked up the phone and heard a voice telling them not to vote in the upcoming presidential primary.
“I heard that and I thought, oh, that’s typical of Joe Biden,” Gingrich told Al Jazeera, “but the fact that he was saying to save his vote and not use it in the next election made me think there’s no way Joe Biden would say something like that.”
The voice may have sounded like the President of the United States, but it wasn't the president: it was a deepfake generated by artificial intelligence (AI).
Experts warn that deepfakes – audio, video and images purposefully created using AI tools to mislead – pose a huge risk to US voters ahead of November's general election, not only by introducing false content into the race but also by undermining public trust.
Gingrich said he wasn't fooled by the Biden deepfake but was concerned it may have depressed voter turnout. The message was sent to about 5,000 New Hampshire voters just days before the state's primary.
“That may be a bad thing for people who don't really know what's going on in the Democratic Party,” said Gingrich, who is chairman of the Burlington Democratic Committee in Burlington, New Hampshire.
“If they really didn't think they should vote on something and Joe Biden had told them not to, they might not have shown up to vote.”
Online groups are vulnerable
The Biden call isn't the only deepfake so far this election cycle: Before Florida Governor Ron DeSantis dropped his presidential bid, his campaign shared a video that included an AI-generated image of Donald Trump embracing immunologist Anthony Fauci, two figures who have clashed publicly during the COVID-19 pandemic.
And in September, another robocall was sent to 300 likely South Carolina Republican primary voters, this time listening to an AI-generated voice impersonating Sen. Lindsey Graham asking who they were voting for.
The practice of altering or falsifying content, especially for political gain, has been around since the dawn of American politics. Even the nation's first president, George Washington, had to deal with a series of “fake letters” that appeared to call into question the cause of American independence.
But AI tools are now advanced enough to mimic humans quickly and cheaply, increasing the risk of disinformation.
A study published earlier this year by researchers at George Washington University predicted that by mid-2024, everyday “AI attacks” will become so commonplace they could threaten November's general election.
Neil Johnson, the study's lead author, told Al Jazeera that the biggest risk comes not from recent obviously fake robocalls with eyebrow-raising messages, but from more convincing deepfakes.
“It's going to be nuanced imagery, altered imagery — not outright disinformation, because disinformation catches the attention of disinformation checkers,” Johnson said.
The study found that online communities are interconnected in ways that allow bad actors to pump large amounts of manipulated media directly into the mainstream.
Communities in battleground states may be especially vulnerable, as may parenting groups on platforms like Facebook.
“The parenting community is going to have a big role to play,” Johnson said, citing how misinformation about vaccines has spread rapidly during the pandemic.
“Suddenly, we [disinformation] “There are many things that are not false or untrue, but are exaggerations of the truth.”
Declining public trust
But deepfakes aren't just aimed at voters: Larry Norden, senior director of the Election and Government Program at the Brennan Center for Justice, works with election officials to help them spot fake content.
For example, a bad actor could use AI tools to tell election workers to close polling places early by manipulating the voice of their boss or sending messages purporting to come from that boss's account, Norden said.
He is instructing poll workers to review the messages they receive to protect themselves.
Norden emphasized that bad actors can create misleading content even without AI: “The great thing about AI is that it makes it easier to create at scale,” he said.
Last year, Norden created a deepfake video of himself to illustrate AI's capabilities during a presentation on the risks posed by AI.
“It took no time at all,” Norden said, explaining that all it took was entering previous TV interviews into the app.
His avatar wasn't perfect—his face was a bit blurry, his voice a bit choppy—but Norden noted that AI tools are improving rapidly: “I think since we recorded it, the technology has gotten more sophisticated and it's becoming harder and harder to tell.”
It's not just the technology that's the problem: as deepfakes become more commonplace, people will become more aware of them and will be more skeptical of the content they view.
This erodes public trust and makes voters more likely to reject truthful information, and politicians may exploit that skepticism for their own purposes.
Legal scholars call this phenomenon the “liar's dividend,” and concerns about deepfakes could make it easier for legitimate audio and video subjects to claim that their recordings are fake.
Norden pointed to Access Hollywood audio released before the 2016 election, in which then-candidate Trump can be heard saying about interacting with women, “You can do anything. Grab a girl by the pussy.”
The tape was thought to be authentic and would hurt Trump's approval rating among female voters. But if a similar recording were leaked today, Norden said, a candidate would easily call it fake. “It's probably easier for the public to reject something like that than it was a few years ago.”
“One of the problems the U.S. has right now is a lack of trust, and this may only make things worse,” Norden added.
What can be done about deepfakes?
Deepfakes have become a growing concern in U.S. elections, but there are relatively few federal laws regulating their use: the Federal Election Commission (FEC) has not yet restricted deepfakes in elections, and bills in Congress remain stalled.
States are rushing to fill the void: Twenty states have enacted laws restricting deepfakes in elections, according to a legislative tracker published by consumer advocacy group Public Citizen.
Additionally, Hawaii, Louisiana and New Hampshire have passed bills that are awaiting their governors' signatures.
Norden said it wasn't a surprise to see states acting before Congress. “It proves true again that states act first because the states are supposed to be the laboratories of democracy. We all know it's really hard to get anything passed in Congress,” he said.
Voters and political groups are also taking action. After receiving a fake phone call from Biden in New Hampshire, Gingrich joined a lawsuit led by the League of Women Voters to demand that the fraud be held accountable.
The caller turned out to be political consultant Steve Cramer, who claimed he was trying to draw attention to the need for regulation of AI in politics. Cramer also admitted to impersonating Senator Graham as the man behind the South Carolina robocalls.
Cramer came forward after NBC News revealed that he had hired a magician to generate a deepfake of Biden's voice using publicly available software.
The deepfake took less than 20 minutes to create and cost just $1, according to the lawsuit.
But Cramer told CBS News that his efforts have garnered “$5 million worth of attention” and that he hopes AI regulation “will take hold, or at least start to bring benefits.”
“My purpose was to make a difference,” he said.
Possibility of applying existing legislation
But Kramer's case shows that existing laws can be used to curb deepfakes.
For example, the Federal Communications Commission (FCC) ruled earlier this year (PDF) that voice imitation software falls under the Telephone Consumer Protection Act of 1991 and is therefore illegal in most cases.
The commission ultimately recommended Kramer be fined $6 million for the illegal robocalls.
The New Hampshire Department of Justice has also charged Kramer with voter suppression and candidate impersonation, which carry a maximum sentence of seven years in prison. Kramer has pleaded not guilty. He did not respond to Al Jazeera's request for comment.
Norden says it's important to note that none of the laws Kramer allegedly violated are specific to deepfakes. “The criminal charges against him have nothing to do with AI,” he says. “Those laws exist independently of the technology used.”
However, these laws are not as easy to apply to bad actors who cannot be identified or who are located outside the United States.
“They're seeing China and Russia already experimenting with these tools, according to intelligence, and they expect they'll be used,” Norden said. “So in that sense, I don't think you can solve this problem through legislation.”
Both Norden and Johnson believe the lack of regulation makes it even more important for voters to educate themselves about deepfakes and how to spot accurate information.
Gingrich, meanwhile, said he knows manipulative deepfakes will become more prevalent, and he feels voters need to educate themselves about the risks.
Her message to voters? “Make sure you know you can vote.”
