|
Getting your Trinity Audio player ready...
|
On the eve of the election, the growing use of artificial intelligence in campaign advertisements by candidates on the campaign trail is raising concerns about how AI-generated content could mislead voters.
In Massachusetts earlier this year, candidates for state office used AI-generated video, audio and photos in attack ads, drawing criticism that they were spreading misinformation in an effort to persuade voters.
In one instance, an unsuccessful Republican gubernatorial hopeful used AI to generate the voice of Gov. Maura Healey.
“I think the simplest way to look at the problem that AI presents is that it is incredibly convincing,” said Saurav Ghosh, director of federal campaign finance reform for the Campaign Legal Center, a nonpartisan legal organization that advocates for ethical campaigning.
Ghosh said he is concerned that AI “threatens” to mislead voters about important information, such as whether they are eligible to vote and when voting takes place, along with general trust in the electoral process.
“The ones that are satires or meant to be recognized as AI are one thing,” Ghosh said. “But I think the real concern is that some of the videos that aren’t meant to be recognized as AI-generated could really open up new possibilities for misleading voters and creating completely fraudulent, fabricated ideas — in terms of not only the impression of certain candidates who are seeking office, but also, more fundamentally, the framework for the election itself.”
A total of 220 ads were identified as being partially made with AI in federal campaigns across the country as part of a project from the Wesleyan Media Project at Wesleyan University. Campaigns spent more than $137 million on those ads, with only 27% including a disclaimer about AI-generated content.
“In campaign finance law, disclaimers are already required to inform voters who paid for an ad, which [is] considered important context for them to evaluate an ad when they’re looking at it,” Ghosh said. “This is basically a different kind of context, which is to say, not only do you need to know who paid for the ad, but also how was it made.”
The report also found Republicans more likely to employ AI for campaign ads than Democrats, with Republican candidates or groups sponsoring 81% of AI-generated ads compared to 12% among their Democratic counterparts.
Campaign attack ads featuring AI-generated content have also made their way into the Bay State ahead of November’s elections. Brian Shortsleeve, who was then a Republican candidate for governor, posted a campaign ad in January with an AI-generated clone of Gov. Maura Healey’s voice the day she launched her campaign.
“I’m Governor Maura Healey and I’m proud to announce my reelection campaign today,” the AI-generated voice says. “We have one of the highest electricity rates in the nation thanks to me for slapping on excessive fees to fund my climate agenda. We are the second most costly state to retire in; more than 5,000 employers have closed their doors, and we have sent more than 12,000 private-sector jobs to Republican states.”
Shortsleeve did not respond to requests for comment about the use of AI in his campaign ad.
Ghosh made a point to distinguish between the brazenness typically expected in campaign ads and those with AI deepfakes, arguing there needs to be a key legal distinction drawn between the two.
“Let’s say somebody comes out with a candidate ad that just blatantly lies about biographical facts, or voting record, or presents something completely false about an electoral opponent,” Ghosh said. “I think there’s an important distinction between that and putting out an AI-based deepfake of an electoral opponent that essentially shows them doing something they’ve never done or saying something they’ve never said.”
Cayce Myers, a public relations professor and director of graduate studies at Virginia Polytechnic Institute and State University, said he thinks whether using AI in a campaign ad could impact a politician’s credibility “depends on who the viewer is.”
“I think that if the person is receptive to what the message is, I don’t know that they’re going to be that upset that AI was used to give the message,” Myers said. “Oftentimes negative ads work as kind of that spark to make people go to the polls and to vote in a way that resonates with whatever the ad may do.”
AI as a tool for campaigns
Bruce Schneier, a fellow at the Berkman Klein Center for Internet & Society at Harvard University and a lecturer in public policy at the Harvard Kennedy School, said that AI “is a tool” that campaigns can use, begging the question why AI should be banned but no other tools.
“If you want to do good things, AI will help. If you want to do evil things, AI will help as well,” he said.
Myers explained how AI can be a “leveler” in political campaigns.
“AI is used in campaigns a lot because it’s a lower-cost manufacturer of content,” Myers said. “You can have good AI tools and somebody that has a subscription [can] create content using that.”
While running for state representative in the 1st Barnstable district, Brewster Democrat Steven Leibowitz told The Boston Globe he used ChatGPT to remove the background from an image of him and Sen. Jamie Eldridge, D-Marlboro, for an endorsement announcement, which led to both of their appearances being altered.
“Sometimes in a campaign, one is in a rush to do things,” Leibowitz wrote in an email after he was defeated in the Democratic primary on Sept. 1. “Personally, and I did not mention it at the time because I did not want it to be an excuse ever, as someone with a vision disability, I should have been more thorough.”
Using AI chatbots to reach voters
Deepfake images or audio are not the only potential problems.
“AI chatbots are very persuasive because they’re highly confirmatory to the viewpoint of the user,” Myers said. “We know that people seek out content in news that is reaffirming to their preexisting held beliefs, and so in that regard, AI plays into our own human nature to give us more and more of what we want, and that what we perceive to be as quote-unquote correct.”
“What that does, though, is it leaves out the content that may challenge your underlying assumption,” he added.
A 2025 study from the science journal Nature Portfolio randomly assigned participants to an AI chatbot instructed to favor either former Vice President Kamala Harris or President Donald Trump. After six minutes of interacting with the model, the pro-Harris chatbots persuaded likely Trump voters toward Harris four times more than the average effect of traditional political advertisements on voting behavior, compared to just roughly two times more for the pro-Trump chatbots that persuaded likely Harris voters toward Trump.
“A person may perceive themselves as extremely well informed because the AI is scraping the news content and distilling it to the user, but at the same time, it’s going to filter some of that news out,” Myers said.
A study published in August from the Brennan Center for Justice at New York University School of Law tested whether multiple chatbots, including ChatGPT, Gemini, Grok, Meta AI, Runway and Flux.2 would create fake election images and videos. The report found that “most of the time” the models generated images based on the content that was asked for directly.
Those images included what appears to be a noncitizen voting, someone collecting multiple ballots from a single person and someone stuffing multiple ballots into a mailbox, the researchers found. Many of the models were also reported to produce accurate copies of government insignia along with falsely altered documents.
AI chatbots could also be prone to “pink slime,” where partisan websites disguised as independent local news outlets feature politically driven talking points that shape chatbots’ responses about electoral races and voting information as they compile information from the internet.
“I think that phenomenon points to the larger problem,” Ghosh said. “If an AI LLM (large language model) or a chatbot powered by an LLM, if it is again only as good as the information that’s being fed into it, and there’s really no regulation or oversight of where that information comes from [or] how reliable it is, I think then that’s a distinct but equally concerning kind of misinformation.”
Lack of regulations for using AI in campaigns
As of now, there is no federal regulation for the use of AI in campaign ads ahead of November’s general election. Federal bills such as the AI Ads Act, which bans fraudulently misrepresenting political candidates or committees through AI-generated content, and the REAL Political Advertisements Act, would require disclaimers on ads using AI-generated images or videos have been introduced in Congress but have not advanced.
The Federal Election Commission voted in September 2024 to not open a rulemaking process for the use of AI in campaign ads. Instead, the commission opted to implement an interpretive rule explaining that fraudulent misrepresentation under federal guidelines is technology neutral, meaning allegations of fraud that involve AI-generated content will be ruled on a case-by-case basis.
On Beacon Hill, the Massachusetts House passed two bills in February aimed at addressing concerns that AI could spread misleading information about elections.
H.5093, “An Act to Protect Against Election Misinformation,” would ban the distribution of “materially deceptive audio or visual media,” including AI-generated media, within 90 days of an election.
H.5094, “An Act Enhancing Disclosure Requirements for Synthetic Media in Political Advertising,” would require disclosures of AI-generated media in political advertising “intended to influence voting for or against a candidate or ballot proposition in an election or primary.”
Both House bills were sent to the Senate Committee on Ways and Means in mid-February. The Senate approved a new draft of S.2631, its version of the election misinformation bill on Oct. 16, which was also referred to the committee.
The formal legislative session for both the House and Senate adjourned in July, but the Legislature can still negotiate and pass bills during informal sessions held until the two-year legislative session officially ends on Jan. 5, 2027. However, moving a measure not in conference during informal sessions requires unanimous consent in each branch.
