Defined, Not Defeated: Brazil’s Deepfake Problem
When delegates gathered at the Liberal Party's convention in São Paulo on July 25, a familiar face urged them to "open arms" for the presidential campaign of Senator Flávio Bolsonaro. The face belonged to his father, former President Jair Bolsonaro. The fatherly gesture, however, was false. Bolsonaro Sr. never recorded it; he is serving a 27-year sentence for plotting a coup to overturn the 2022 election results. The Liberal Party aired a video generated by artificial intelligence.
On September 1, Brazil's Superior Electoral Court (TSE) ruled on that video: it let the video stand, gave Brazil its first legal definition of a deepfake, and left judges to apply that definition case by case. That decision reveals how the country plans to handle synthetic media before the October 4 general election. For now, its plan rests on judges reviewing complaints one at a time and on a voluntary detection tool built with Google. Neither can keep pace with content that spreads in hours.
The court split twice by a 4-3 vote. Justices declined to fine Flávio Bolsonaro or remove the video, reasoning that the Liberal Party convention served delegates rather than voters and that the clip never explicitly asked for votes. In a separate 5-2 vote, the court set Brazil's first legal definition of a deepfake: "synthetic content produced or manipulated via artificial intelligence or equivalent technology," realistic enough "to create, reproduce or alter the image, voice or speech of any living, deceased or fictional person." The TSE had already banned deepfakes in campaigns in 2024, with sanctions ranging from removal and fines to the cancellation of a candidacy, but it never said what counted as one. Did any realistic AI creation qualify, or only content meant to mislead? Nobody knew. That gap matters far beyond the courtroom, because a candidate cannot obey a rule no one can define, and a voter scrolling past a convincing clip of a politician has no way of knowing whether the law protects them from it. The ban applies only when content qualifies as "electoral propaganda," and streaming a party convention online does not automatically turn a statement of support into illegal campaigning. Campaigns now have a map of where synthetic content can travel without triggering the rule. The Liberal Party's own video shows the route: played for delegates, streamed online, and left standing by the court.
The word "fictional" matters because of a character called "Dona Maria," an AI-generated elderly Black woman who criticizes President Luiz Inácio Lula da Silva. A 37-year-old Uber driver created her using Google Gemini. At least 12 of her videos have passed one million views, and her profiles have reached an estimated 100 million people, close to half of Brazil. In April, three parties in Lula's coalition sued to suspend the accounts for early campaigning and undisclosed AI use; the case still sits before Justice Estela Aranha. Dona Maria impersonates no one. Under the new definition, she can still count as a deepfake. That shift matters because the most influential synthetic voices in this election may not belong to any real person at all.
The court also left a harder question unanswered: does a label make a deepfake acceptable? The Bolsonaro video carried a clear warning that AI produced it, and the defense argued that a clearly labeled video cannot deceive anyone. Justice Aranha disagreed and called for the removal of deepfakes from elections, labeled or not. Her position points to a limit of disclosure: a label tells viewers that a video is synthetic, yet it cannot undo the emotional pull of watching an imprisoned leader promote his son's campaign. The court has not decided whether disclosure is enough.
Online platforms remain the biggest open question. The ruling assigns no new duties to the companies that host and spread this content, even though their algorithms decide how far it travels. So far, platforms have offered tools rather than commitments. In August, Google and the TSE launched Likeness ID, a YouTube feature that alerts registered candidates when a video appears to use their face. Candidates must verify their identity, review each match, and ask YouTube to remove the video under the company's own privacy policies. Google Brasil President Fábio Coelho called the tool especially important for women, describing them as: "one of the groups most affected by attacks on the image during the elections." It helps, but it covers one platform and only candidates, and it does nothing about a synthetic character like Dona Maria. Fernando Neisser argues that the TSE "missed the opportunity to bring these platforms to the table." The Senate passed an AI framework in late 2024 that would require visible watermarks on synthetic content, but the Chamber of Deputies has yet to approve it.
Judges are therefore writing Brazil's rules case by case, weeks before the election. The approach has real strengths. Courts move faster than legislatures, and the TSE produced a usable definition in a single ruling while Congress has debated for two years. However, courts act only after the harm. Each ruling becomes a precedent the next judge must interpret, and each question leaves loopholes open, from labels to platforms to even fictional characters.. Other democracies heading into elections should watch the Brazilian case closely. Defining a deepfake was the easy part. Enforcing the definition is the hard one. Brazil will find out on October 4 whether it can.