Deepfakes, distrust and disinformation: Welcome to the AI election

This article is part of a series, Bots and Ballots: How artificial intelligence is reshaping elections worldwide.

Callum Hood has the power to undermine any election with a few keystrokes from his Boston apartment.

Hood, a British researcher, fired up some of the latest artificial intelligence tools made by OpenAI and Midjourney, another AI startup. Within seconds of him typing in a few prompts — “create a realistic photo of voter ballots in a dumpster”; “a photo of long lines of voters waiting outside a polling station in the rain”; “a photo of Joe Biden sick in the hospital” — the AI models spat out reams of realistic images. 

Political operatives and foreign governments using these tools — which can generate lifelike images seemingly indistinguishable from real-world photos — hope to sow doubt across the more than 50 elections planned worldwide in 2024. Hood used the U.S. vote in November as just one example. But such tactics, he added, could be replicated across any election — from the European Union to India to Mexico.

“You can generate as many of those images as you want, very, very quickly,” Hood said during a Zoom call.

Hood is head of research at the Center for Countering Digital Hate, a nonprofit that successfully created scores of examples of harmful AI-generated election disinformation content to warn about the technology’s potential to undermine democracy.  

Listen to the experts

“What is the usefulness of a technology that can produce such realistic images, and how does it outweigh the potential harm?” Hood asked after producing even more AI-powered images of alleged spoiled ballots with a few clicks. “It really is unclear.”

No one denies the potential for harm.

Damaging so-called deepfake — the term for misleading, AI-generated content — audio clips of British Labour Party Leader Keir Starmer and his Slovak counterpart, opposition head Michal Šimečka, spread like wildfire on social media before being debunked by fact-checkers. False robocalls, allegedly from U.S. President Joe Biden and also created using AI tools, similarly flooded the recent New Hampshire Democratic primary to urge people not to vote. Those, too, were quickly outed as fakes.

Easy access to AI tools like ChatGPT and its competitors risks a tidal wave of politically motivated falsehoods flooding social media feeds in ways that seemed unimaginable even a few years ago. In an era of entrenched partisan politics and growing skepticism about what is published online, AI has the potential to make this year’s motherlode of election cycles significantly more difficult to predict.

But what’s still unclear, based on POLITICO’s discussions with more than 30 politicians, policymakers, tech executives and outside experts from many of the countries holding elections in 2024, is what actual impact AI will have when more than 2 billion people head to the polls from New Delhi to Berlin and Washington. 

Despite the latest AI tools’ apparent wizardry, almost all deepfakes are still easily — and quickly — debunked, including those produced by Russia and China, which already use the technology in global influence campaigns. Many people’s political views are hard to budge, and experts warn that AI-generated content won’t change most voters’ party affiliations — no matter how compelling the made-up photo, video or audio clips appear.

In multiple elections so far this year, including in countries like Pakistan and Indonesia, where generative AI has been rolled out at scale, there’s little — if any — evidence the technology unfairly skewed outcomes toward one politician over another. With so many social media posts published daily worldwide, the ability for AI-powered falsehoods — even lifelike ones like those created by Hood — to cut through is an uphill battle. 

Publicly, lawmakers, tech executives and outside groups monitoring elections urge caution when dealing with a technology developing faster than it can be controlled. But for many, the impact of AI-generated disinformation — already months into this year’s global election cycle — remains more theoretical than hard fact.

“The trends, they may change, they may change dramatically,” Nick Clegg, president of global affairs at Meta, said when discussing how AI-powered falsehoods had affected elections worldwide in 2024. “But, so far, [the trends] have not suggested something wildly out of order.”

The threat is real; the evidence, less so

Clegg’s words are of little comfort to Cara Hunter.

Only weeks before the Northern Irish legislative election in 2022, the then-24-year-old local politician received a WhatsApp message from someone she didn’t know. The man quickly asked if she was the woman in an explicit video — a 40-second deepfake clip of Hunter allegedly performing graphic sexual acts. 

Within days, the AI-generated video had gone viral, and the Northern Irish politician was bombarded with increasingly sexual and violent social media messages from men around the world.

Listen to the experts

“It was a campaign to undermine me politically,” said Hunter, who narrowly won her Northern Ireland Assembly seat by just a few votes, despite the deepfake pornography. “It has left a tarnished perception of me that I can’t control. I’ll have to pay for the repercussions of this for the rest of my life.”

Hunter isn’t alone in being targeted by an AI-powered political campaign. At least three Western national leaders have been shown lifelike copies of themselves in AI companies’ briefings to highlight how realistic such digital forgeries have become, according to five officials with knowledge of those discussions who were granted anonymity to speak candidly.

In Moldova, President Maia Sandu has been repeatedly targeted with AI-generated deepfakes to ridicule both her personally and her pro-Western government. The country’s national security authorities accused the Kremlin of the attacks, a new tactic of a yearslong interference campaign. 

Russian-linked groups also created another forged video — this time, of an AI-generated Tom Cruise criticizing the upcoming Paris Olympics via a spoofed Netflix documentary. Moscow has targeted the upcoming global sporting event with sophisticated influence campaigns after the International Olympic Committee banned Russian athletes’ participation. They’ll now be admitted as neutral athletes. 

Countries’ politicians have also tapped into the latest AI advances for domestic political gain. In Poland, the political party of current Prime Minister Donald Tusk posted a deepfake audio clip of his opponent on social media during the country’s recent election campaign. Meanwhile, in the U.S., onetime Republican presidential candidate and current Florida Gov. Ron DeSantis used a similar AI-generated image of former President Donald Trump to attack his opponent, with little success. 

In response, more than 20 of the world’s leading tech companies — including TikTok, Meta and OpenAI — pledged during a set-piece event at the recent Munich Security Conference to combat underhanded uses of AI tech during this year’s global election cycle. The European Commission started an early-stage probe into the most advanced AI tools as part of the 27-country bloc’s new social media rules. The U.S. Congress, too, held hearings about AI’s potential harm, including those associated with elections.

The companies’ voluntary commitments include efforts to stop bad actors, including foreign governments and domestic political campaigns, from creating harmful deepfakes. They will also share industry-wide techniques — like automatically labeling AI-generated images and videos — to let the public know, instantly, when they’re confronted with such doctored content. The corporate response comes as almost all governments have little, if any, technical expertise to tackle the ever-changing threat of election-related AI.

“People are constantly coming up with new ways to try [to] break the systems,” Natasha Crampton, who leads Microsoft’s efforts to make AI systems more responsible, said. The tech giant’s commercial links with both U.S.-based OpenAI and Mistral, a French competitor, have made it a key player in combating AI-generated disinformation worldwide.

Please panic responsibly

If anyone has an interest in connecting this cavalcade of AI-powered falsehoods to harmful outcomes at the polls, it’s Josh Lawson.

As an executive at Meta, the parent company for Facebook, Instagram and WhatsApp, Lawson once led the firm’s response to global policy elections. The tech giant’s platforms remain critical for how both legitimate information and disinformation reach voters. Now, at the Aspen Institute, Lawson oversees the think tank’s AI Elections Initiative, a project to curb the emerging technology’s negative influence on the 2024 election cycle and beyond.

Yet, despite dozens of meetings with election officials, civil society groups and tech companies — including a recent event in New York that included former presidential candidate Hillary Clinton and Nobel Peace Prize winner Maria Ressa — Lawson has yet to find concrete evidence of AI-generated disinformation directly changing voters’ habits, especially during nationwide elections. It may be out there, he concedes, but no one has yet demonstrated evidence of AI-powered trickery upending a nationwide vote.

“It's really frustrating,” he said. 

That frustration — and resignation — was borne out in POLITICO’s interviews with national security officials, tech company executives and outside watchdog groups.

These individuals, many of whom were granted the condition of anonymity to speak about their ongoing confidential work, acknowledged that AI offered a faster and cheaper way to create disinformation, both during and outside of election periods. But the technology — as of early 2024 — was more an additional instrument, as opposed to a genuine new threat, to help spread harmful election-related falsehoods.

“We’re looking, we really are,” a senior official at a leading AI company, who was granted the condition of anonymity to describe internal deliberations, said. “But we just haven’t seen a mass impact of the AI covert content that’s out there.”

The misleading and the mundane

In part, that’s because such AI-generated content often isn’t very convincing. 

A recent Russian-backed attempt to peddle a deepfake social media video, allegedly highlighting differences between Ukrainian President Volodymyr Zelenskyy and one of his top generals, made basic linguistic errors, while the audio lip-synced badly with the video, based on analysis from Microsoft shared with POLITICO. Recent dubious images of Trump allegedly surrounded by Black supporters were debunked within hours on social media — before they could be shared widely.

“That type of misinformation isn’t successful in the wild,” Felix Simon, a researcher at the University of Oxford who has tracked the inability of harmful AI-generated content to break through with the public, said. People’s wariness of what they see online, coupled with most AI-generated content garnering little, if any, attention in the sea of daily social media posts, has limited the impact of such content, he added.

Listen to the experts

For Amber Sinha, an India-based AI expert at the Mozilla Foundation, a nonprofit connected to the tech company behind the Firefox browser, people’s legitimate fixation on so-called generative AI — the most advanced uses of the tech powering deep fakes — missed the most likely way for AI to affect this year’s global election cycle.

In India, a country of 1.4 billion people where voting in its general elections starts on April 19, political parties and outside consultants are already using more mundane AI techniques, like so-called machine learning tools, to mine voters’ personal data to better target them with online political advertising.  

Such tactics, according to Sinha, have allowed these operatives to crunch reams of often sensitive personal information like people’s socioeconomic status to pinpoint what messages would most break through with would-be supporters. Currently, those efforts are more widespread — and more effective — than the small, but growing, stable of AI-generated images created to mislead people.

“It’s not as shiny as generative AI,” Sinha said. “But these boring uses of AI continue to play an important role.”