We’ve been warned about fraudulent AI fakes. We have memes.
First came the election results, then came the meme party.
Following the Democratic wave of support on Election Day, content creators used artificial intelligence to flood the Internet with fake videos showing candidates and celebrities in various celebrations.
One AI-generated skit, featuring cast members from the hit TV show “The Office,” showed the boss adapting to a progressive society, wearing a Pride flag sweater and calling for an emergency Diversity Day. He got over 16 million views in one post on X.
The video was emblematic of the kinds of AI fakes that circulate on social media during midseason. Many of the most popular videos circulating online have been vying for laughs and sometimes ridicule – an absurd mix of bizarre events that apparently never happened.
The results come as a minor relief to researchers tracking the spread of AI fakes, as well as a reckoning.
Experts have warned before every US election since at least 2020 that artificial intelligence replicas of real people – “deepfakes” – and other deceptive content generated by the technology could subvert democracy, misleading voters about scandals that never happened or lewd fake comments.
Some ads published this year featured realistic versions of AI candidates. But far more of the videos circulating online were parodies, jokes and memes that were too absurd to be believed.
“The case that everyone fears is a highly inflammatory or incriminating deepfake that goes viral on the eve of an election and really moves the needle on how people voted,” said Henry Ajder, an AI expert and founder of Latent Space Advisory, an AI consulting firm. “It hasn’t taken yet.”
The memes have spread widely online, depicting the real candidates in bizarre situations, such as one series mocking President Trump as a small business owner with Vice President JD Vance as his son and employee. Creators opposing the bid of Senator-elect Abdul El-Sayed posted a barrage of videos disparaging the candidate, often using racist stereotypes.
Hundreds of AI-generated ad campaigns also flooded the Internet and radio broadcasts, including some that were deceptive and not disclosed as AI-generated, according to the agency’s report Wesleyan Media Projectresearch group at Wesleyan University. The vast majority of ads were clearly fake and depicted impossible scenes.
Some were more reality-based, including an ad featuring an AI version of Vikki Goodwin, then a Democratic gubernatorial candidate in Texas, allegedly supporting a “death tax.”
“There are a lot of people who don’t know me yet who might think it’s me,” she told The New York Times in September. The campaign behind the ad said at the time that it was an “obvious parody made for entertainment”.
Ms. Goodwin ended up winning her race by more than six percentage points.
One frequent target of attack ads, Sen.-elect James Talarico, a Texas Democrat, won his election handily. Although there is plenty of material from Mr. Talaric to generate AI fakes or audio clones, experts have found no evidence of his deceptive deepfake representing real footage. Instead, the candidate appeared in dozens of AI-generated dancing and singing videos that were viewed millions of times online.
These AI fakes can still have political power, rallying supporters and reviving criticisms against candidates that can stick in the minds of voters.
“This is something political campaigns have been doing forever – it’s nothing new,” Mr Ajder said. “The means of production are new, the possibility of scaling, the approach – the viscerality is different.”
The memes correspond to content posted during the primary elections in the United States and in recent general elections in other countries. Report from Canadian researchers at McGill University and the University of Montreal found that AI-generated content made up about 6 percent of election-related images on X during the country’s federal election last year. However, more than 90 percent of the views went to more innocuous content, such as memes, jokes and false portrayals of the candidates as superheroes or supervillains.
Experts have pointed to several past cases where AI-generated fakes appeared to mislead voters. In 2023, a sound clip generated by artificial intelligence was created credited with a possible reversal of the course of elections in Slovakia. In 2024, an AI replica of President Joseph R. Biden’s voice was used to call voters in New Hampshire, advising them to skip the primary.
In this year, the political group created a phone line for the AI version Nithya Raman, a Los Angeles mayoral candidate who eventually won her race. On the phone line, he appeared to support tent camps.
Such fakes tend to target partisans who are more likely to believe them, rather than appeal to swing voters who tend to be more skeptical. They may spread more easily in group chats or private groups, the researchers said.
And online videos with a more comedic edge have racked up millions of views.
“The threat of deepfakes has not gone away,” said Shaina Raza, a scientist at the Vector Institute of Artificial Intelligence who has studied AI risks. “It’s taking different forms than we expected.”