The scene repeats itself in the dark of a bedroom, a thumb sliding across the phone screen until it stops on a familiar face: a prime-time news anchor announcing, in his usual voice and cadence, the results of an election poll that never existed. During Brazil's first-round campaign, videos like this circulated by the hundreds on social media, even after the Superior Electoral Court (TSE), the body that oversees elections in the country, banned this kind of content.
At least 554 deepfake videos and images were posted on social networks between August 16 and September 28, according to a report by the VigIA project, run by the fact-checking agency Lupa and released on Thursday (1st), as reported by Agência Brasil. Deepfakes, pieces that use artificial intelligence to make a real person say or do things they never said or did, accounted for 60% of the 920 posts with confirmed AI use identified by the project in the period, an average of almost one piece of content per hour.
Electoral rules ban deepfakes and set penalties that can go as far as stripping a candidate of their candidacy or mandate, as long as the piece is classified as political propaganda and is realistic enough to deceive voters. Even so, 29 of the mapped deepfakes were spread by profiles belonging to candidates or parties. Most, however, came from ordinary and often anonymous accounts: 525 of the fake posts originated from this kind of profile, which, according to researcher Beatriz Farrugia, who authored the report, poses a concrete challenge to enforcing the law. "Here the courts would need to work with the platforms to try to identify these people," she told Agência Brasil.
Journalists and celebrities used as bait
The faces chosen by the fabricators are not random. According to Lupa, at least 30 of the identified deepfakes involved journalists from TV Globo, Brazil's largest broadcaster, such as César Tralli, William Bonner and Renata Lo Prete, shown "reporting" fake polls or invented results of the presidential race, including a supposed defeat of President Luiz Inácio Lula da Silva (PT) by Flávio Bolsonaro (PL), a senator and son of former president Jair Bolsonaro. Celebrities such as singer Ana Castela and TV host Luciano Huck also had their images manipulated to simulate endorsements of candidates.
"Journalism and journalists carry a lot of credibility, and disinformation agents know it. The repetition of this strategy shows that it works to deceive voters," said Cristina Tardáguila, founder of Lupa and coordinator of the Lupa Observatory.
Lula was the public figure most depicted in AI-generated pieces, with 379 appearances, of which 39% were humor or satire and 34% attacked his candidacy. Flávio Bolsonaro appears in 190 posts, 40% of them favorable to him and 28% humorous. Overall, promoting candidacies was the most frequent purpose, present in 414 of the 920 cases (45%), while a quarter of the content carried attacks. Facebook concentrated most occurrences (559), followed by Instagram (365), X (96) and TikTok (70).
The report also points to another recurring violation: the law allows the use of AI in campaigns as long as the piece is labeled as such, but only 371 of the 920 posts, about 40%, carried the mandatory label. Of the 549 posts deemed irregular for lacking identification, 75 were made by candidates or parties. In total, VigIA collected 1,941 suspicious electoral contents in the period and analyzed 972 of them, using three detection tools from the Artificial Intelligence Laboratory at Unicamp (Recod.ai), human review and tips sent via WhatsApp.
Back on the phone screen in the dark bedroom, the anchor's video keeps playing, identical to a real newscast segment. It is in that second of hesitation, between believing and doubting, that the VigIA report locates the true battleground of this election.