Deepfakes Are Dominating Global Elections, and Regulation Is Playing Catch-Up

This round of Brazil’s electoral battle between incumbent President Luiz Inácio Lula da Silva and Flávio Bolsonaro ended at the ballot box on October 25, 2026, when Lula secured an unprecedented fourth term. One of the campaign’s most pivotal lessons, however, wasn’t one of left versus right, but of a country’s ability to keep pace with technology. In this case, the election spotlighted the ability for a government to decide whether the content a certain technology creates is “fake.”

Former Brazilian President Jair Bolsonaro is under house arrest. He is forbidden from making public statements, even on social media. So then how did he endorse his son’s presidential bid during his party’s convention in July to a massive crowd? An AI avatar. After the 46-second video disclosed its AI origins, the elder Bolsonaro’s avatar slammed the nature of his arrest and coronated his son, Flávio, as his chosen successor. When a coalition led by Lula’s Workers’ Party challenged the video in court, Flávio’s camp claimed no wrongdoing due to its AI disclosure. Moreover, the real Jair Bolsonaro, through his representative, denied authorizing the video. Just over a month later, Brazil’s electoral court ruled that the country’s deepfake ban did not apply, because the video was shown at a closed party event and was not used in campaign advertising. In the same decision, the court also narrowed what content the ban covers.

So, what makes AI-generated content a deepfake? And how should democracies regulate it? Brazil isn’t alone as it confronts these questions. AI-generated content is becoming a fixture of modern political campaigns, used for everything from satire to misleading depictions of candidates. As its use expands, regulators everywhere face three politically contentious debates: (1) what counts as a deepfake, (2) where legitimate free expression ends and intentional deception begins, and (3) whether any legal rule can be enforced at the speed of AI’s rapid evolution.

The Pervasiveness of AI-Generated Content in Elections

Several other major elections this year have highlighted just how quickly the usage of AI content has spread:

• In Colombia, which held its presidential election in May, nearly all of the candidates turned to AI-generated videos as part of their campaigns. For example, Pacto Histórico candidate Iván Cepeda released a video featuring an animated refrigerator to support the government’s minimum wage increase.
• Israel, which held its general parliamentary election in late October, has seen an explosion of AI-generated content, with an estimated 1 in 8 election-related social media posts altered by AI. Some of those posts have come directly from Israel’s most prominent politicians, including Prime Minister Benjamin Netanyahu, Finance Minister Bezalel Smotrich, National Security Minister Itamar Ben-Gvir, and Yashar Party leader Gadi Eisenkot.
• In Spain, which has its general election on November 29, the People’s Party reportedly circulated an AI-generated video impersonating Prime Minister Pedro Sánchez’s wife and other Socialist Party figures admitting to misconduct.
• Ahead of the U.S. midterm elections on November 3, an estimated $137 million has been spent on over 220 videos that use AI-generated content. About 70 percent of the videos do not contain a disclosure of AI use.

In each of the countries highlighted so far, content has been created and disseminated not only by fringe or anonymous accounts, but by political leaders and parties who in many cases have the backing of funding and campaign infrastructure.

The Challenges of Regulation

The recent use of AI in elections has revealed three key challenges in regulation: defining what constitutes a political deepfake, distinguishing between political expression and deception, and enforcing restrictions at the speed and scale of AI-generated content.

What Constitutes a Deepfake?

Brazil’s Superior Electoral Court (TSE) confronted the challenge of deepfakes prior to the country’s 2024 municipal elections. The court’s resolution issued an outright ban on deepfakes for political campaigns and a labeling requirement for AI content. It defined a deepfake then as “synthetic content” in audio and/or video mediums “which has been digitally generated or manipulated, even with authorization, to create, replace, or alter the image or voice of a living, deceased, or fictitious person.” The prohibition applied even with an individual’s consent.

By design, the ban was broad. However, the court narrowed it this year. In the Bolsonaro case, the TSE decided, in a 5–2 vote, to limit the prohibition of deepfakes in electoral campaigns to those with “a degree of realism or verisimilitude,” and only when it qualified as “electoral propaganda.” The court’s president, Kassio Nunes Marques, believed the latter qualification was enough to resolve the case. Because the video targeted party members at a closed convention and was not an “explicit request for votes,” the deepfake ban did not apply. There was dissent on both conditions. TSE Minister Ricardo Villas Bôas Cueva argued the video favored Bolsonaro’s candidacy (and that classification as “propaganda” was the wrong test), while another TSE Minister, Floriano de Azevedo Marques, stated the video was indeed a deepfake, asserting that the definition of a deepfake did not need to include “realism” as an attribute, subscribing to a stricter reading of the 2024 rule. TSE president Nunes Marques said the court needed to redefine what a deepfake was, owing to the speed of the medium’s evolution.

Brazil’s electoral court has since ordered the removal of AI content generated by both camps, including three videos from Bolsonaro, an AI video posted by Workers’ Party congressman Lindbergh Farias, and a pro-Lula video shared by the head of the Attorney General’s office.

Proponents of AI content regulation might have admired Brazil’s 2024 rule as ambitiously broad and strict, especially since the definition of a banned deepfake was fairly simple: any “synthetic” likeness that could help or hurt a political candidate. Yet the rule’s breadth narrowed under the pressures of the 2026 presidential contest, its most significant test to date. What Brazil now has is a rule that is more precise but significantly harder to apply. The court now needs to decide, on a case-by-case basis and often after the content’s dissemination, whether it counts as “electoral propaganda,” and only then weigh competing interpretations of “realism.” Whatever its merits in protecting expression, Brazil’s rule emerged from its stress test with considerably less clarity.

Brazil’s experience suggests a cautionary lesson for other democracies. Deepfake bans are easy to write in principle, but the ban’s application to high-stakes elections or leading national candidates may yield political and legal caution that could detract from its ultimate goal.

Political Expression vs. Deception

The question that looms larger than simply whether AI was used is whether or not the AI was used to deceive.

Brazil’s deepfake rules cover only campaign advertising, which leaves out the vast amounts of AI-generated content coming from ordinary citizens or anonymous accounts. Supporters of both Lula and Bolsonaro deployed AI content depicting the candidates and other political figures. Some portrayed the candidates dancing, while others simulated realistic news reports—two types of content with vastly divergent effects. While other Brazilian laws do make it illegal to knowingly distribute misinformation, potential loopholes and the burden of proof around intentionality make them harder to apply to the mass spread of deepfakes.

Israel, by contrast, opted for a more permissive approach. In July 2026, Israel began requiring that AI-generated or manipulated content in election campaigning be labeled. Specifically, the rules cover content showing “a person, place, event, document, or object that could appear to have been authentically documented when it was not.”

While such requirements increase transparency, they also raise questions about what labels are meant to catch. Critics of expansive labeling requirements argue that this approach does not address the actual problem, because labels are not required exclusively for deceptive content, but instead for synthetic content generally. For example, Israel’s rule depends on whether content could be perceived as “real,” not whether it was meant to deceive—meaning a labeled deception complies.

At the same time, even when labels are required, questions are raised about whether disclosing AI use can mitigate the effect that intentionally deceptive content has on voters. Labels are often difficult to notice or easily overlooked, limiting their ability to counteract misleading information. In July, Israeli Justice Noam Sohlberg, who chairs Israel’s Central Elections Committee, ordered Finance Minister Bezalel Smotrich to remove an AI clip of opposition leader Gadi Eisenkot shaking hands with Palestinian Authority President Mahmoud Abbas, even though Smotrich disclosed AI usage, because Smotrich’s Facebook audience would be unlikely to see the disclaimer given the site’s display. Yet in October, the same judge refused to order the removal of another AI video posted by Smotrich depicting legal officials being dragged by police, saying the video was to be “judged by the voters, not by election law.” In this case, Sohlberg pushed back on the Yashar Party’s legal claim that the video threatened law enforcement and was election interference.

The same problem presented itself in Spain. Spain’s government approved a bill in 2025 requiring companies pay up to $38 million in fines if they did not label AI-generated content. An AI video from the People’s Party (PP) disparaging Socialist Party figures included an AI label, but it was small and faint. Even so, that faint label may help fend off potential legal challenges from the individuals depicted in the video. Arguably, the recent Spanish and Israeli examples carry the risk of misleading voters and could therefore be considered “politically deceptive.” This raises a broader question about the discretionary power in the hands of regulators and judges to determine what constitutes “political expression” versus “political deception.” When these determinations depend on subjective interpretation, there is a risk of influencing which political messages are allowed to circulate versus which are not, potentially allowing some deceptive content to remain online while restricting expression in other cases.

None of these approaches cleanly distinguish free expression from intentional deception. Brazil’s ban (with its current nuances) targets campaigns, which does not reach a significant portion of AI-generated content. In Israel, the same judge sought to remove one inflammatory video, while letting another stand. And in Spain, a label too faint to notice may still offer legal protections for the party posting it. The question of whether AI content violates rules comes down to a case-by-case reading by judges. And sometimes, the judges will leave the judgment to voters themselves, reflecting a principled reluctance to let the state decide which political messages may circulate. That deference still has its own risk, however, as synthetic content grows harder for many to distinguish from authentic content.

Enforcing Regulation at the Speed of AI

Even when governments have established restrictions on political deepfakes, enforcement poses its own challenge. Generative AI tools have reduced the financial and technical barriers to producing and distributing synthetic content, allowing deepfakes to spread much faster than regulators can respond to them. Between 2023 and 2025, online deepfakes were estimated to have increased from 500,000 to 8 million.

Authorities must decide whether AI-generated content violates the rules before ordering its removal, a process that often takes place after the content has already circulated. In the final days and weeks before an election, this delay could be especially problematic, as AI-generated content can reach a substantial audience before enforcement mechanisms are utilized. The video generated by Spain’s PP exemplified this challenge, as it generated over 2 million views across social media platforms in less than a day, before legal complaints were filed. And in Israel, Sohlberg’s decisions on both of Smotrich’s AI videos came weeks after the videos’ initial posting.

What Do These Issues Mean for the United States?

With more elections coming up and AI tools becoming increasingly sophisticated and accessible, the challenges posed by deepfakes will only become more pressing. The examples offered by Brazil, Israel, Spain, and other countries that have attempted to regulate synthetic content demonstrate that effective legal mechanisms are not as simple as banning content or requiring disclosure. Instead, regulators must consider an approach that effectively distinguishes legitimate expression from deceptive content, while simultaneously addressing this content in a timely manner before it can spread misinformation to voters.

With the midterms approaching on November 3 and a presidential election two years away, the United States has no federal law specifically addressing deepfakes. In 2024, the Federal Election Commission opted against creating a separate rule governing AI-generated content in elections, concluding its existing fraud provisions already applied. The Fraudulent Artificial Intelligence Regulations (FAIR) Elections Act, proposed in June 2026, would prohibit the distribution of false AI-generated content in election media but has not been enacted.

In the absence of a comprehensive federal regulation, the current deepfake regulatory landscape is a patchwork of state-based laws, with different restrictions and consequences. California, Maryland, Minnesota, and Texas all maintain bans on political deepfakes for varying lengths of time prior to an election, and 27 other states require disclosures instead. However, the inherently borderless nature of AI-generated media demonstrates that a patchwork approach will be ineffective, as this content can spread across state lines regardless of where it originated.

The experiences of Brazil, Israel, and Spain demonstrate that the challenge of regulating AI-generated political content is not just prohibiting deepfakes or requiring their disclosure, but determining what content poses the risk of crossing from political expression to deception. Ultimately, the effectiveness of these varying approaches will not depend solely on what they explicitly prohibit, but on how effectively they can be enforced before Election Day.

A’ndre Gonawela is a fellow with the Strategic Technologies Program at the Center for Strategic and International Studies (CSIS) in Washington, D.C. Kuhu Badgi is program manager and research associate for the Strategic Technologies Program at CSIS.