Taylor Swift Endorses Kamala Harris — and Raises Fears of AI Election Misinformation

Taylor Swift’s much-anticipated endorsement of Vice President Kamala Harris also called out an AI-generated image of the pop superstar shared by former President Donald Trump. In doing so, she added her voice to a chorus of concerns over the technology and its potential for election-related misinformation.

The image reposted by Trump last month on Truth Social, his right-wing social media platform, shows Swift posing Uncle Sam-style and saying she “wants YOU” to vote for Trump. The former president added a message: “I accept!”

After Tuesday’s debate, Swift expressed alarm over the doctored image in her Instagram post endorsing Harris.

“Recently I was made aware that AI of ‘me’ falsely endorsing Donald Trump’s presidential run was posted to his site. It really conjured up my fears around AI and the dangers of spreading misinformation,” she wrote. “It brought me to the conclusion that I need to be very transparent about my actual plans for this election as a voter. The simplest way to combat misinformation is with the truth.”

As much as Swift’s fans were outraged by the photo — and some called on her to denounce it sooner — one AI expert says that a less partisan and provocative use of AI could pose a bigger threat to election integrity.

Lucas Hansen, a co-founder of the Berkeley-based AI knowledge nonprofit CivAI, said two kinds of AI content are at play in the election: inflammatory posts, like the one of Swift, as well as a post on social media site, X from Elon Musk that shows Harris donning a red suit with a communist hammer and sickle symbol, and other content that’s aimed less at angering voters and more at affecting their behavior.

In New Hampshire, for example, some voters received robocalls ahead of the March primary election using an assumed AI-generated voice resembling that of President Joe Biden. The calls told voters to “save their vote” for the November general election.

“That was discouraging people from turning out to the primary in New Hampshire, and that’s not inflammatory, but it can be substantially more effective,” Hansen said. “I think a lot of the highest impact things that we’ll see before and during the election are of that type.”

To show what AI is capable of, CivAI has been modeling other ways that it could be used to impact voting in the run-up to election day. Hansen said that could look like doctored news articles that resemble real publications publicizing a fake suspected shooting near a polling place or misinformation telling voters that they can avoid paying to park at a polling place by showing up to a certain location during a very tight time frame, “close to when polls close and not a legitimate place.”

It’s more likely that voters will be tricked by this kind of AI-generated content than by deepfakes of the candidates or celebrities like Swift, according to Hansen, who expects to see production of this kind of content ramp up — especially in swing states and targeting swing voters.

Another CivAI model shows how AI might use highly personalized content to target voters. One of its demos uses LinkedIn to find out a person’s top concerns and writes them an email about a fake, plausible-sounding tax or regulation supposedly proposed by a candidate on one of those issues.

Hansen tried it out on his sister, a freelance artist. The email to her said Biden planned to impose additional taxes on freelancers.

“That’s so targeted and so specific that a lot of people just wouldn’t think that’s possible to do right now at any sort of automated level,” Hansen said.

Jonathan Mehta Stein, the executive director of the nonpartisan watchdog group Common Cause, said that conspiracy theorists who believe election fraud is rampant could also use AI to mimic the sophisticated systems that elections officials run to detect duplicate voters.

“You can use AI to mimic that work, but with none of the controls and none of the safeguards” that elections officials use, he said. “You can create hugely erroneous lists of voters that you want to challenge.”

Inflammatory content is also still a concern, especially as high-performing AI tools become more accessible. About a month ago, X added a beta version of Grok 2.0 for its premium members, which is an AI image generation feature using a tool from Black Forest Labs.

“It’s the first extremely easily available image generation tool that’s high quality,” Hansen told KQED.

Though many people can detect such deepfakes, some are tougher to spot, he said — like one photo posted on X of Trump giving a thumbs up, surrounded by people wearing shirts that say “Walz’s for Trump.” It was captioned, “The *entire* Tim Walz Family just endorsed Donald J. TRUMP for President 2024!”

Although the Associated Press confirmed that some distant cousins of the Democratic vice presidential nominee did pose for a separate photo in Trump shirts, the image that showed Trump surrounded by supposed Walz relatives was AI-generated.

Hansen said that even as someone with a lot of experience detecting deepfake images, this one didn’t have any of the normal “tells,” like messed up hands or background blurring.

Though he said these images are often debunked and mostly sow division, he “expect[s] that we’ll see a large increase in the volume of those sorts of debates as a result of how accessible Grok is.”

Stein said the most significant risk could be how this content affects voters’ trust in politics.

“The danger is not that 10,000 voters will be deceived by a deepfake and change who they’re voting for,” though it’s possible, he said. “The danger is that 10 million voters who are awash in disinformation between now and Election Day throw up their hands and say, ‘Forget it, I’m not voting at all.’”

A bigger impact of the fake photos of Swift might be her response itself — since her post pointing people to vote.gov, the site has received more than 337,000 visits as of Wednesday afternoon, according to NPR.