States rush to combat AI threat to elections

By ZACHARY ROTH

New Hampshire Bulletin

Published: 04-03-2024 6:00 PM

This year’s presidential election will be the first since generative AI — a form of artificial intelligence that can create new content, including images, audio, and video — became widely available. That’s raising fears that millions of voters could be deceived by a barrage of political deepfakes.

But, while Congress has done little to address the issue, states are moving aggressively to respond — though questions remain about how effective any new measures to combat AI-created disinformation will be.

The New Hampshire House passed a bill Thursday to require that all political advertisements in the state disclose to viewers if they used “synthetic media,” defined as any image, audio recording, or video recording of an individual’s appearance, action, or speech, that is realistic but fake.

That bill, House Bill 1596, would mandate that disclosure for all advertisements within 90 days of the election. The disclosure would need to be visible on a static or print advertisement, appear for the whole duration of a video advertisement, and be spoken clearly at 2-minute intervals for an audio advertisement.

The bill would apply to people, corporations, and political committees. News organizations that replay AI generated clips as part of a news report would be exempt as long as the outlet clearly identified the media as false and AI-driven. The bill also includes exemptions for satire and parody efforts. 

The problem is global. Last year, a fake, AI-generated audio recording of a conversation between a liberal Slovakian politician and a journalist, in which they discussed how to rig the country’s upcoming election, offered a warning to democracies around the world.

Here in the United States, the urgency of the AI threat was driven home in February, when, in the days before the New Hampshire primary, thousands of voters in the state received a robocall with an AI-generated voice impersonating President Joe Biden, urging them not to vote. A Democratic operative working for a rival candidate has admitted to commissioning the calls. 

In response to the call, the Federal Communications Commission issued a ruling restricting robocalls that contain AI-generated voices.

Article continues after...

Yesterday's Most Read Articles

Not all have been in favor of legislative efforts. In New Hampshire, half of the members on the House Election Law Committee opposed the bill. Critics said it would have a “chilling effect on freedom of expression” and could cause candidates or campaigns to self-censor. After its passage in the full House Thursday, the bill heads next to the Senate. 

Some conservative groups even appear to be using AI tools to assist with mass voter registration challenges — raising concerns that the technology could be harnessed to help existing voter suppression schemes. 

“Instead of voters looking to trusted sources of information about elections, including their state or county board of elections, AI-generated content can grab the voters’ attention,” said Megan Bellamy, vice president for law and policy at the Voting Rights Lab, an advocacy group that tracks election-related state legislation. “And this can lead to chaos and confusion leading up to and even after Election Day.”

Disinformation worries

The AI threat has emerged at a time when democracy advocates already are deeply concerned about the potential for “ordinary” online disinformation to confuse voters, and when allies of former president Donald Trump appear to be having success in fighting off efforts to curb disinformation.

But states are responding to the AI threat. Since the start of last year, 101 bills addressing AI and election disinformation have been introduced, according to a March 26 analysis by the Voting Rights Lab.

On March 27, Oregon became the latest state — after Wisconsin, New Mexico, Indiana, and Utah — to enact a law on AI-generated election disinformation. Florida and Idaho lawmakers have passed their own measures, which are currently on the desks of those states’ governors.

Arizona, Georgia, Iowa, and Hawaii, meanwhile, have all passed at least one bill — in the case of Arizona, two — through one chamber.

As that list of states makes clear, red, blue, and purple states all have devoted attention to the issue.

In New Hampshire, Rep. Angela Brennan, a Bow Democrat, praised the House vote on Thursday.

“I am grateful that, in this first election year where artificial intelligence is accessible at everyone’s fingertips, the House was able to come together to pass HB 1596 on a strong bipartisan voice vote,” she said in a statement. “The dissemination of disinformation is detrimental to our democracy and this body is sending a strong message that Granite State voters deserve better than dirty tricks.” 

States urged to act

Meanwhile, a new report on how to combat the AI threat to elections, drawing on input from four Democratic secretaries of state, was released March 25 by the NewDEAL Forum, a progressive advocacy group. 

“(G)enerative AI has the ability to drastically increase the spread of election mis- and disinformation and cause confusion among voters,” the report warned. “For instance, ‘deepfakes’ (AI-generated images, voices, or videos) could be used to portray a candidate saying or doing things that never happened.”

The NewDEAL Forum report urges states to take several steps to respond to the threat, including requiring that certain kinds of AI-generated campaign material be clearly labeled; conducting role-playing exercises to help anticipate the problems that AI could cause; creating rapid-response systems for communicating with voters and the media, in order to knock down AI-generated disinformation; and educating the public ahead of time. 

Secretaries of State Steve Simon of Minnesota, Jocelyn Benson of Michigan, Maggie Toulouse Oliver of New Mexico, and Adrian Fontes of Arizona provided input for the report. All four are actively working to prepare their states on the issue.

Loopholes seen

Despite the flurry of activity by lawmakers, officials, and outside experts, several of the measures examined in the Voting Rights Lab analysis appear to have weaknesses or loopholes that may raise questions about their ability to effectively protect voters from AI.

Most of the bills require that creators add a disclaimer to any AI-generated content, noting the use of AI, as the NewDEAL Forum report recommends.

But the new Wisconsin law, for instance, requires the disclaimer only for content created by campaigns, meaning deepfakes produced by outside groups but intended to influence an election — hardly an unlikely scenario — would be unaffected. 

In addition, the measure is limited to content produced by generative AI, even though experts say other types of synthetic content that don’t use AI, like Photoshop and CGI — sometimes referred to as “cheap fakes” — can be just as effective at fooling viewers or listeners, and can be more easily produced. 

For that reason, the NewDEAL Forum report recommends that state laws cover all synthetic content, not just that which use AI.

New Hampshire’s proposed law does that: The legislation applies to any image that has been “created or intentionally manipulated,” and is not limited to generative AI. 

The Wisconsin, Utah, and Indiana laws also contain no criminal penalties — violations are punishable by a $1,000 fine — raising questions about whether they will work as a deterrent.

The Arizona and Florida bills do include criminal penalties. But Arizona’s two bills apply only to digital impersonation of a candidate, meaning plenty of other forms of AI-generated deception — impersonating a news anchor reporting a story, for instance — would remain legal. 

And one of the Arizona bills, as well as New Mexico’s law, applied only in the 90 days before an election, even though AI-generated content that appears before that window could potentially still affect the vote. 

Experts say the shortcomings exist in large part because, since the threat is so new, states don’t yet have a clear sense of exactly what form it will take.

“The legislative bodies are trying to figure out the best approach, and they’re working off of examples that they’ve already seen,” said Bellamy, pointing to the examples of the Slovakian audio and the Biden robocalls.

“They’re just not sure what direction this is coming from, but feeling the need to do something.”

“I think that we will see the solutions evolve,” Bellamy added. “The danger of that is that AI-generated content and what it can do is also likely to evolve at the same time. So hopefully we can keep up.”