The AI Election, Revisited: Deepfakes in 2024

The AI Election

The 2024 election cycle’s encounter with generative AI — including the New Hampshire Biden robocall, deepfakes in the Slovak and Indonesian elections, and AI-generated content in the US, India, and EU. Post-election analyses found AI’s impact was smaller than feared but that the threat is structural.

The Predicted Apocalypse

Going into 2024, the predictions were dire. Analysts, journalists, and policymakers warned that generative AI could be used to create convincing deepfakes of candidates, to spread disinformation at scale, to suppress voter turnout through impersonation, and to undermine trust in the electoral process. The phrase “AI election” entered the vocabulary, suggesting that 2024 would be the year when AI fundamentally changed how elections were fought and won.

The concerns were not abstract. They were based on real incidents. In September 2023, two days before Slovakia’s parliamentary election, an audio recording was posted to Facebook allegedly featuring Michal Šimečka, the leader of Progressive Slovakia, and journalist Monika Tódová discussing vote-rigging. The recording was quickly denounced as a deepfake, but not before it had spread widely on social media. Robert Fico’s SMER-SD party won the election with 22.94% of the vote, and Fico returned as Prime Minister. Whether the deepfake swung the election is debated — the Times of London argued it did, while the Harvard Misinformation Review concluded it did not — but the incident was widely cited as a warning of what was to come.

In April 2023, the Republican National Committee had released a fully AI-generated attack ad against President Biden, within hours of his re-election announcement — the first fully AI-generated US political attack ad. The ad used AI-generated images to depict a dystopian future under a second Biden term. It was a signal that AI had arrived as a tool of American political campaigning.

In November 2023, Meta announced that it would require political advertisers to disclose AI-generated or digitally altered content, beginning in 2024. The announcement was a recognition, by one of the world’s largest social media platforms, that AI-generated political content was a real and growing problem.

These incidents, and others, created a climate of concern. The question was not whether AI would be used in the 2024 elections — it clearly would be — but how much damage it would do.

The New Hampshire Biden Robocall

The most-documented AI incident of the 2024 cycle was the New Hampshire Biden robocall. On January 21, 2024, two days before the New Hampshire Democratic primary, voters in the state began receiving robocalls that appeared to be from President Biden. The calls used an AI-generated clone of Biden’s voice, and they told Democrats not to vote in the primary — suggesting that voting in the primary would help Republicans, and that Democrats should save their vote for November.

The calls were quickly identified as a deepfake. The voice was not Biden’s; it was a clone, created using AI voice-cloning technology. The message was designed to suppress Democratic turnout in the primary, and it was a clear attempt at voter suppression.

The investigation that followed revealed the people behind the calls. On February 23-26, 2024, Paul David Carpenter, a street magician from New Orleans, publicly identified himself as the creator of the Biden voice clone (he said he had been paid about $150 to create the audio). Carpenter had been hired by Steven Kramer, a political consultant who was working for the presidential campaign of Dean Phillips, a Democratic congressman from Minnesota who was challenging Biden for the Democratic nomination. Kramer admitted to commissioning the calls.

The regulatory response was swift. On February 8, 2024 — less than three weeks after the calls — the Federal Communications Commission (FCC) unanimously adopted a Declaratory Ruling that AI-generated voices in robocalls are “artificial or prerecorded voices” under the Telephone Consumer Protection Act (TCPA) — the 1991 statute restricting telemarketing — making them illegal without prior consent. This was a significant ruling: it meant that the existing legal framework for regulating robocalls could be applied to AI-generated voices, without the need for new legislation.

On May 23, 2024, Kramer was indicted by the State of New Hampshire on felony voter-suppression charges and misdemeanour impersonation-of-a-candidate charges. On the same day, the FCC proposed a $6 million fine against Kramer for the illegal robocall and spoofing campaign. (The FCC had issued a Notice of Apparent Liability on February 22, 2024; the May 23 action was the formal proposal alongside the indictment.) On September 26, 2024, the FCC formally adopted the $6 million fine.

The Kramer case tracks the full regulatory arc of an AI election incident: the incident itself, the investigation, the regulatory response, the criminal charges, and the civil penalty. It is the most complete case study of how the system responded to an AI-related election threat. And the response, in this case, was effective. The calls were identified quickly, the perpetrators were identified, and the regulatory and legal systems responded with appropriate speed.

But there is a coda. On June 13, 2025, a New Hampshire jury found Kramer not guilty of the voter-suppression charges. Kramer had argued that the calls were a “warning” about AI risks, not an attempt to suppress votes. The civil fine stood, but the criminal acquittal was a reminder that the legal system does not always agree on how to handle these cases.

Slovakia, Indonesia, India, and the EU

The New Hampshire robocall was the most-documented incident, but it was not the only one. AI was used in elections around the world in 2024, in ways that ranged from malicious to merely novel.

In Slovakia, the September 2023 deepfake audio of Michal Šimečka was the most-discussed pre-2024 incident. It was a clear example of AI being used to spread disinformation in the final days of an election campaign, when there is little time to debunk the false information before voters go to the polls. Whether it swung the election is debated, but it demonstrated the potential for AI to be used as a tool of last-minute electoral manipulation.

In Indonesia, the February 2024 presidential election featured a different kind of AI use. Prabowo Subianto, a former general with a controversial human rights record, used AI-generated “gemoy” cartoon avatars — cute, chubby, cartoon versions of Prabowo — to rebrand himself. The avatars were shared widely on social media, and they helped to soften Prabowo’s image, transforming him from a “feared former general” to a “cute Gen Z icon.” Prabowo won the election. The use of AI in this case was not malicious in the same way as the Slovak deepfake — it was a creative campaign tool — but it illustrated how AI could be used to manipulate political images and personas.

In India, the April-June 2024 general election featured widespread use of AI-generated content. Deepfakes of Bollywood stars Aamir Khan and Ranveer Singh, criticising Prime Minister Narendra Modi, circulated widely on social media (the stars had not made the statements attributed to them). Both the ruling Bharatiya Janata Party (BJP) and the opposition Congress Party used AI-generated content. Modi himself used AI-generated Hindi translations of his speeches to reach non-Hindi-speaking audiences. The Indian election was a reminder that AI is a tool that can be used by all sides of a political contest, and that its use is not limited to one party or one ideology.

In the European Parliament elections in June 2024, the European Commission documented “known information interference operations” involving AI technologies. The Alan Turing Institute’s Centre for Emerging Technology and Security (CETAS) published a widely-cited retrospective finding that only 27 pieces of AI-generated content went viral during the summer 2024 European elections. This was a strikingly low number, given the concerns that had been raised before the elections. It suggested that, while AI-generated content was present, it did not achieve the viral reach that had been feared.

The “Liar’s Dividend”

One of the most-discussed concepts going into the 2024 elections was the “liar’s dividend.” The term was coined by Robert Chesney, a law professor at the University of Texas, and Danielle Citron, a law professor at Boston University, in a 2019 paper published in the California Law Review (see B83). The idea is simple but powerful: once deepfakes become common, it becomes easier for politicians to dismiss real evidence of their own wrongdoing as “fake.” If everything could be a deepfake, then nothing is reliably real, and the very concept of evidence is undermined.

The liar’s dividend was a particular concern in the 2024 election because of the figure at the centre of American politics: Donald Trump. Trump had a long history of dismissing unflattering evidence as “fake news,” and the rise of deepfakes gave this strategy new credibility. If audio or video of Trump saying something damaging emerged, his supporters could plausibly argue that it was a deepfake, even if it was real.

The Brennan Center for Justice (the NYU law institute), in a report titled “Deepfakes, Elections, and Shrinking the Liar’s Dividend,” argued that the liar’s dividend was arguably a more pervasive harm than direct deepfakes. The harm is not just that false content is created, but that the existence of false content makes it harder to hold people accountable for real content. This is a subtle but important point. The damage done by deepfakes is not just the false information they spread; it is the erosion of the shared reality on which democratic accountability depends.

The liar’s dividend was not just a theoretical concern in 2024. It was invoked in practice. When the Access Hollywood tape — in which Trump was recorded making lewd comments about women — resurfaced during the campaign, some of Trump’s supporters suggested that the tape could be a deepfake. It was not — the tape was real, and had been verified years earlier — but the suggestion that it could be a deepfake was a sign of how the concept had entered the political discourse.

What the Post-Election Analyses Found

After the November 2024 elections, a series of retrospectives examined what had actually happened. The findings were remarkably consistent, and they were somewhat reassuring.

The Alan Turing Institute’s CETAS found only 27 pieces of AI-generated content that went viral during the summer 2024 European elections. The Harvard Ash Center published a retrospective titled “The apocalypse that wasn’t,” arguing that the predicted AI-driven electoral disaster had not materialised. The Knight First Amendment Institute at Columbia published “Don’t Panic (Yet): Assessing the Evidence and Discourse Around Generative AI and Elections,” which reached a similar conclusion. The Washington Post published an article on November 9, 2024 — four days after the US election — with the headline “AI didn’t sway the election, but it deepened the partisan divide.”

The broad consensus of these retrospectives was that AI did not demonstrably swing any major election in 2024. The predicted apocalypse — mass-scale AI manipulation swinging election outcomes — did not happen. The AI-generated content that did circulate was, for the most part, identified, debunked, and contained. The existing defensive infrastructure — fact-checkers, platform policies, journalistic scrutiny — proved more effective than many had feared.

But the retrospectives also cautioned against complacency. The Brennan Center warned that “effects are likely to be greater in the future.” The technology is improving rapidly, and the defensive infrastructure is not keeping pace. The 2024 elections may have been a reprieve, not a resolution.

The retrospectives also noted that the harm done by AI in 2024 was more diffuse and harder to measure than the predicted outcomes. AI may not have swung any elections, but it did contribute to the deepening of partisan divides, to the erosion of trust in media and institutions, and to the normalisation of synthetic content. These are not the dramatic harms that were predicted, but they are real harms, and they are likely to accumulate over time.

The Washington Post’s framing — that AI “deepened the partisan divide” even as it did not sway the election — captured the nuanced reality. AI did not destroy democracy in 2024. But it did make the information environment worse, in ways that are hard to measure and hard to reverse.

The Regulatory Response

The regulatory response to AI in elections was rapid, if uneven.

At the federal level, the FCC’s February 8, 2024 Declaratory Ruling on AI voices in robocalls was the most significant action. The ruling made it clear that existing law — the Telephone Consumer Protection Act — applied to AI-generated voices, and it provided a basis for enforcement. The $6 million fine against Steve Kramer was the first major enforcement action under this ruling.

At the state level, 20 US states had enacted election-related deepfake laws by 2024. The most prominent was California’s AB 2655, the “Defending Democracy from Deepfake Deception Act of 2024,” signed by Governor Gavin Newsom on September 17, 2024. AB 2655 required platforms to remove or label deepfakes related to elections, and it provided for civil penalties.

But the state laws faced legal challenges. On August 5, 2025, a federal judge struck down California’s AB 2655 on First Amendment grounds. X Corp, the company formerly known as Twitter, had sued to block the law, arguing that it violated the right to free speech. The court agreed, holding that the law’s restrictions on deepfakes were too broad and too vague. The ruling was a significant setback for state-level efforts to regulate election deepfakes, and it raised questions about whether any state law could pass constitutional muster.

At the international level, the European Union’s AI Act includes provisions on deepfakes. Article 50(4) of the AI Act requires providers of AI systems that generate synthetic content to label it as such, and the deepfake-labeling obligations come into legal force on August 2, 2026 (see B86). The EU approach is more comprehensive than the US approach — it applies to all AI-generated content, not just election-related content — but it is also more controversial, with critics arguing that it is too broad and too burdensome.

The platform policies were also significant. Meta (Facebook and Instagram) required political advertisers to disclose AI-generated or digitally altered content. Google and YouTube required similar disclosures. TikTok maintained a near-total ban on political advertising. X (formerly Twitter) had less consistent synthetic-media policy enforcement, and its owner, Elon Musk, was himself a source of AI-related controversy — in July 2024, he reposted a parody Kamala Harris campaign ad that used an AI-cloned voice.

Lessons for 2026-2028

The 2024 elections provided several lessons for the elections to come.

First, the predicted apocalypse did not happen, but the underlying risks remain. AI-generated content did not swing any major election in 2024, but the technology is improving, and the defensive infrastructure is not keeping pace. The 2026 midterms and the 2028 presidential election in the United States, and elections in other countries, will face more capable AI systems and more sophisticated AI-generated content. The fact that 2024 was not a disaster does not mean that future elections will be safe.

Second, the liar’s dividend is a real and growing problem. The ability of politicians to dismiss real evidence as “fake” is a threat to democratic accountability, and it is getting worse as deepfakes become more common. Addressing the liar’s dividend will require not just technical solutions — like provenance standards and content authentication — but also cultural and institutional changes that rebuild trust in shared facts.

Third, the regulatory response is uneven and contested. The FCC’s robocall ruling was effective, but it was limited to a specific use case. The state deepfake laws are facing constitutional challenges, and their future is uncertain. The EU’s AI Act is more comprehensive, but it is also more controversial. There is no consensus on the right regulatory approach, and the debate is likely to continue for years.

Fourth, the platform policies matter, but they are inconsistent. Meta, Google, and TikTok have implemented disclosure requirements and content policies, but X has been less consistent, and the platforms’ enforcement is uneven. The role of platforms in addressing AI-generated election content is still being worked out, and it is likely to remain a subject of debate.

Fifth, the global picture is diverse. The 2024 elections saw AI used in different ways in different countries — from the malicious deepfake in Slovakia, to the creative campaign tool in Indonesia, to the widespread use by all sides in India. There is no one-size-fits-all approach to AI in elections, and the responses will need to be tailored to the specific political and technological context of each country.

The India AI Impact Summit, scheduled for February 2026, will be the first global AI summit in the Global South, and it will be an opportunity to broaden the conversation about AI and elections beyond the major Western powers. The 2026 US midterms will be the next major test of the American electoral system’s ability to handle AI-generated content. And the 2028 presidential election will be the next presidential test.

The 2024 elections were not the AI apocalypse that had been predicted. But they were not a clean bill of health either. They were a warning, and a reprieve. The warning is that the risks are real, and they are growing. The reprieve is that the systems we have — fact-checkers, platform policies, regulatory frameworks, journalistic scrutiny — can still, just barely, keep up. Whether they will be able to keep up in the future is an open question, and one that the next elections will answer.