Days after New Hampshire voters received a robocall with an artificially generated voice that resembled President Joe Biden’s, the Federal Communications Commission banned the use of AI-generated voices in robocalls.
It was a flashpoint. The 2024 United States election would be the first to unfold amid wide public access to AI generators, which let people create images, audio and video – some for nefarious purposes.
Institutions rushed to limit AI-enabled misdeeds.
Sixteen states enacted legislation around AI’s use in elections and campaigns; many of these states required disclaimers in synthetic media published close to an election. The Election Assistance Commission, a federal agency supporting election administrators, published an “AI toolkit” with tips election officials could use to communicate about elections in an age of fabricated information. States published their own pages to help voters identify AI-generated content.
Experts warned about AI’s potential to create deepfakes that made candidates appear to say or do things that they didn’t. The experts said AI’s influence could hurt the US both domestically – misleading voters, affecting their decision-making or deterring them from voting – and abroad, benefitting foreign adversaries.
But the anticipated avalanche of AI-driven misinformation never materialised. As Election Day came and went, viral misinformation played a starring role, misleading about vote counting, mail-in ballots and voting machines. However, this chicanery leaned largely on old, familiar techniques, including text-based social media claims and video or out-of-context images.
“The use of generative AI turned out not to be necessary to mislead voters,” said Paul Barrett, deputy director of the New York University Stern Center for Business and Human Rights. “This was not ‘the AI election.’”
Daniel Schiff, assistant professor of technology policy at Purdue University, said there was no “massive eleventh-hour campaign” that misled voters about polling places and affected turnout. “This kind of misinformation was smaller in scope and unlikely to have been the determinative factor in at least the presidential election,” he said.
The AI-generated claims that got the most traction supported existing narratives rather than fabricating new claims to fool people, experts said. For example, after former President Donald Trump and his vice presidential running mate, JD Vance, falsely claimed that Haitians were eating pets in Springfield, Ohio, AI images and memes depicting animal abuse flooded the internet.
Meanwhile, technology and public policy experts said, safeguards and legislation minimised AI’s potential to create harmful political speech.
Schiff said AI’s potential election harms sparked “urgent energy” focused on finding solutions.
“I believe the significant attention by public advocates, government actors, researchers, and the general public did matter,” Schiff said.
Meta, which owns Facebook, Instagram and Threads, required advertisers to disclose AI use in any advertisements about politics or social issues. TikTok applied a mechanism to automatically label some AI-generated content. OpenAI, the company behind ChatGPT and DALL-E, banned the use of its services for political campaigns and prevented users from generating images of real people.
Misinformation actors used traditional techniques
Siwei Lyu, computer science and engineering professor at the University at Buffalo and a digital media forensics expert, said AI’s power to influence the election also faded because there were other ways to gain this influence.
“In this election, AI’s impact may appear muted because traditional formats were still more effective, and on social network-based platforms like Instagram, accounts with large followings use AI less,” said Herbert Chang, assistant professor of quantitative social science at Dartmouth College. Chang co-wrote a study that found AI-generated images “generate less virality than traditional memes,” but memes created with AI also generate virality.
Prominent people with large followings easily spread messages without needing AI-generated media. Trump, for example, repeatedly falsely said in speeches, media interviews and on social media that illegal immigrants were being brought into the US to vote even though cases of noncitizens voting are extremely rare and citizenship is required for voting in federal elections. Polling showed Trump’s repeated claim paid off: More than half of Americans in October said they were concerned about noncitizens voting in the 2024 election.
PolitiFact’s fact-checks and stories about election-related misinformation singled out some images and videos that employed AI, but many pieces of viral media were what experts term “cheap fakes” — authentic content that had been deceptively edited without AI.
In other cases, politicians flipped the script — blaming or disparaging AI instead of using it. Trump, for example, falsely claimed that a montage of his gaffes that the Lincoln Project released was AI-generated, and he said a crowd of Harris supporters was AI-generated. After CNN published a report that North Carolina Lieutenant Governor Mark Robinson made offensive comments on a porn forum, Robinson claimed it was AI. An expert told Greensboro, North Carolina’s WFMY-TV what Robinson had claimed would be “nearly impossible”.
AI used to stoke ‘partisan animus’
Authorities discovered a New Orleans street magician created January’s fake Biden robocall, in which the president could be heard discouraging people from voting in New Hampshire’s primary. The magician said it took him only 20 minutes and $1 to create the fake audio.
The political consultant who hired the magician to make the call faces a $6m fine and 13 felony charges.
It was a standout moment partly because it wasn’t repeated.
AI did not drive the spread of two major misinformation narratives in the weeks leading up to Election Day – the fabricated pet-eating claims and falsehoods about the Federal Emergency Management Agency’s relief efforts following Hurricanes Milton and Helene, said Bruce Schneier, adjunct lecturer in public policy at the Harvard Kennedy School.
“We did witness the use of deepfakes to seemingly quite effectively stir partisan animus, helping to establish or cement certain misleading or false takes on candidates,” Daniel Schiff said.
He worked with Kaylyn Schiff, an assistant professor of political science at Purdue, and Christina Walker, a Purdue doctoral candidate, to create a database of political deepfakes.
The majority of the deepfake incidents were created as satire, the data showed. Behind that were deepfakes that intended to harm someone’s reputation. And the third most common deepfake was created for entertainment.
Deepfakes that criticized or misled people about candidates were “extensions of traditional US political narratives,” Daniel Schiff said, such as ones painting Harris as a communist or a clown, or Trump as a fascist or a criminal. Chang agreed with Daniel Schiff, saying generative AI “exacerbated existing political divides, not necessarily with the intent to mislead but through hyperbole”.
Major foreign influence operations relied on actors, not AI
Researchers warned in 2023 that AI could help foreign adversaries conduct influence operations faster and cheaper. The Foreign Malign Influence Center – which assesses foreign influence activities targeting the US – in late September said AI had not “revolutionised” those efforts.
To threaten the US elections, the centre said, foreign actors would have to overcome AI tools’ restrictions, evade detection and “strategically target and disseminate such content”.
Intelligence agencies – including the Office of the Director of National Intelligence, the FBI and the Cybersecurity and Infrastructure Security Agency – flagged foreign influence operations, but those efforts more often employed actors in staged videos. A video showed a woman who claimed Harris had struck and injured her in a hit-and-run car crash. The video’s narrative was “wholly fabricated”, but not AI. Analysts tied the video to a Russian network it dubbed Storm-1516, which used similar tactics in videos that sought to undermine election trust in Pennsylvania and Georgia.
Platform safeguards and state legislation likely helped curb ‘worst behavior’
Social media and AI platforms sought to make it harder to use their tools to spread harmful, political content, by adding watermarks, labels and fact-checks to claims.
Both Meta AI and OpenAI said their tools rejected hundreds of thousands of requests to generate AI images of Trump, Biden, Harris, Vance and Democratic vice presidential candidate Minnesota Governor Tim Walz. In a December 3 report about global elections in 2024, Meta’s president for global affairs, Nick Clegg, said, “Ratings on AI content related to elections, politics and social topics represented less than 1 percent of all fact-checked misinformation.”
Still, there were shortcomings.
The Washington Post found that, when prompted, ChatGPT still composed campaign messages targeting specific voters. PolitiFact also found that Meta AI easily produced images that could have supported the narrative that Haitians were eating pets.
Daniel Schiff said the platforms have a long road ahead as AI technology improves. But at least in 2024, the precautions they took and states’ legislative efforts appeared to have paid off.
“Strategies like deepfake detection, and public-awareness raising efforts, as well as straight-up bans, I think all mattered,” Schiff said.
