On September 1, 2026, Brazil's Superior Electoral Court (TSE) voted 4-3 against fining presidential candidate Flávio Bolsonaro over an AI-generated video of his father, Jair Bolsonaro, shown at the Liberal Party (PL) convention on July 25. By 5-2, the court also set the test for an illicit electoral deepfake: synthetic content realistic enough to pass as an authentic recording. Obviously artificial material such as memes, caricatures and animations falls outside it. The ruling landed five weeks before the October 4 election, which makes it the working rulebook for the first AI-saturated presidential campaign in Brazil.
The case for the strictest reading
The strongest argument for a harsher outcome deserves a fair statement. Brazil's rules already ban deepfakes in campaigning outright. Resolution 23.732 of February 2024 amended Resolution 23.610/2019 to prohibit synthetic audio or video that creates, replaces or alters the image or voice of a living, deceased or fictional person, and to require clear labelling of AI-altered material. Jair Bolsonaro, whose likeness was reproduced, is a polarising figure whose supporters may well have taken a convincing digital double at face value. Regulators who see a bright-line ban being tested on day one could reasonably argue that leniency invites imitation. Dissenting justices Ricardo Villas Bôas Cueva, Floriano de Azevedo Marques and Estela Aranha took that side on the fine.
What the majority actually decided
The 4-3 outcome turned on context more than technology. According to Congresso em Foco's report, the majority of Kassio Nunes Marques (the rapporteur), André Mendonça, Dias Toffoli and Antonio Carlos Ferreira found that the video was shown at a party convention, not in electoral advertising, so the irregularity alleged by the PT-led federation did not arise. That is a narrow holding about venue and timing. It does not bless AI resurrection of political figures in campaign material.
The broader 5-2 vote is the more consequential one. The court defined deepfakes as content produced or manipulated by AI to create, reproduce or alter a person's image, voice or statement with enough realism to appear as an authentic recording. A video that announces its own artificiality, in a setting where nobody is deceived, is not what the ban was written to catch.
Why the realism test is good policy
A deepfake rule that reaches every AI-touched pixel regulates a technology; a rule that reaches deception regulates a harm. The realism criterion keeps the target on the second. Without it, a satirical animation of a candidate, a stylised meme or an obviously synthetic jingle could be treated as a prohibited deepfake simply because a model generated it. In a country where political humour is a staple of online debate, that would hand incumbents and well-funded litigants a tool to suppress mockery with a labelling technicality.
The test also fits the evidence on how courts have struggled. A July 2026 Consultor Jurídico analysis describes three competing interpretations that emerged from regional electoral courts in the 2024 municipal elections: total prohibition, conditional permission, and context-dependent limits that tolerate satire. It also notes a drift in recent TSE decisions toward treating any alteration of digital content for electoral purposes as sufficient to establish irregularity. The September ruling pulls back from that drift. Whether it holds depends on how lower courts apply "realistic enough" in practice, a standard that will need case law to harden.
What stays prohibited
Nothing here loosens the core rules. Deepfakes remain banned in electoral propaganda, and AI-generated material must carry explicit, prominent disclosure naming the technology used. Resolution 23.755, approved on March 2, 2026, added a blackout: new synthetic content using the image, voice or statements of candidates or public figures may not be published, republished or boosted between 72 hours before voting ends and 24 hours after it closes, even if labelled. Penalties for violations can reach cancellation of a candidate's registration or revocation of a mandate.
The blackout is the most defensible of these tools, because it targets timing. A fabricated clip released on the eve of voting leaves no time for debunking, correction or a court order, and the harm is irreversible once ballots are cast. A narrow, time-limited, content-neutral restriction is proportionate in a way that an open-ended ban on creative AI use is not. Its weakness is enforcement: it depends on platforms removing content quickly, and on distinguishing new synthetic content from re-circulated older material.
The privacy gap the ruling does not touch
Deepfake rules address only the output. EFF, Access Now and Data Privacy Brasil argued in August that electoral integrity also depends on data protection. Microtargeting, synthetic personas posing as spontaneous supporters, and "synthetic voters" used to test messages all rely on personal data, and the groups urged closer coordination between the data protection authority (ANPD) and the TSE, plus impact assessments for voter profiling. Those recommendations are about the upstream machinery that makes any deepfake persuasive, and they are less likely to chill speech than content-removal powers.
What to watch
Three things will show whether the framework works. First, whether courts treat "realistic enough to pass as authentic" consistently, or whether judges fall back on the strictest reading when a case is politically charged. Second, whether platforms can comply with the 72-hour window without over-removing lawful commentary. Third, whether the party-convention exception is exploited: if venue can shield an AI resurrection, campaigns will route similar content through it. A regime that distinguishes deception from expression, and spends its heaviest tools on the final days before a vote, is the proportionate design. The test now is whether the TSE applies it with the same discipline it showed on September 1.