At least 554 deepfake videos and images circulated on social media during the first round of Brazil's 2026 general election campaign, even though the practice is banned by the Superior Electoral Court (TSE), the country's top electoral authority. The figure comes from VigIA, a project by the fact-checking agency Lupa that monitors AI use in the election together with Recod.ai, an artificial intelligence lab at the University of Campinas (Unicamp). Between August 16 and September 28, researchers confirmed 920 posts created or edited with AI, nearly one per hour. Deepfakes account for 60% of them.
A deepfake works like a digital mask: artificial intelligence wraps a real person's face and voice around an invented script, so they appear to say or do things that never happened. Among the most common cases are videos in which TV Globo journalists, such as César Tralli, William Bonner and Renata Lo Prete, "announce" opinion polls and results that do not exist. At least 30 pieces used the network's journalists. Celebrities like singer Ana Castela and TV host Luciano Huck also had their images manipulated to fake endorsements. "Journalism and journalists have a lot of credibility, and disinformers know it. The repetition of this strategy shows that it works to deceive voters," said Cristina Tardáguila, Lupa's founder.
Who is behind it
Most of the fake pieces, 525, came from ordinary users, many of them anonymous. Another 29 were posted by candidate or party profiles, which can trigger electoral sanctions ranging from fines to the loss of a candidacy or mandate. For a violation to stick, however, the TSE requires the content to be political advertising and realistic enough to fool voters.
"If we want to block deepfakes altogether because of the risk they pose to people's understanding of political events, we need to consider that most of this content comes from user accounts and anonymous avatars. Here the courts would need to work with the platforms to try to identify these people," said Beatriz Farrugia, the researcher in charge of the report.
The report points to a second large-scale violation. Electoral law allows AI in campaigns as long as the material is labeled, but only 371 of the 920 posts (about 40%) carried the disclosure. Of the 549 irregular publications, 75 were made by candidates or parties. For Tardáguila, the numbers show that AI "is already part of everyday political communication, not just isolated episodes."
Targets and platforms
President Lula, running for reelection, was the most depicted figure, appearing in 379 publications: 39% humorous or satirical and 34% attacking his candidacy. His runoff opponent, senator Flávio Bolsonaro, appeared in 190, with 40% seeking to favor him. Promoting candidacies was the most common purpose overall (414 cases, 45%), followed by attacks (about a quarter) and satire. On the platform map, Facebook (559 occurrences) and Instagram (365), both owned by Meta, concentrate most records, followed by X (96), TikTok (70) and WhatsApp (3). Since the same piece can circulate on more than one network, the figures do not add up to the total.
What we still don't know
The 554 deepfakes are a floor, not a ceiling. VigIA works with what arrives through reader tips and its own team's searches, and half of the collected universe (1,941 suspicious posts) is still awaiting review. Audio was left out for methodological reasons, and WhatsApp, where encryption blocks scanning, shows only three cases, which says more about how hard the app is to monitor than about real circulation there. The report also does not measure reach: it is impossible to know how many people saw each piece or changed their vote because of it. The incentives matter too. Platforms profit from engagement, which false and emotional content tends to generate, while detecting and removing deepfakes at scale is a cost, not a revenue stream.
In practice, voter self-defense remains manual. Be skeptical of videos in which well-known journalists "report" polls or results outside their networks' official channels, look for the AI-generated content label and check with fact-checking outlets. Lupa runs a WhatsApp channel for reporting suspicious content, and VigIA will keep monitoring social media through the October 25 runoff.