What the TSE actually decided
Brazil's Superior Electoral Court (TSE) issued two rulings on September 1, 2026, in Representação 0601315-97.2026.6.00.0000, a case brought by the Federação Brasil da Esperança (the PT-PCdoB-PV coalition) over an AI-generated video shown at the Liberal Party's national convention on July 25. The video recreated the image and voice of former president Jair Bolsonaro — sentenced in November 2025 to 27 years in prison for leading a coup attempt — endorsing his son, Sen. Flávio Bolsonaro, for president.
By 5-2, the justices adopted a binding definition: a deepfake is "synthetic content produced or manipulated via artificial intelligence or equivalent technology" realistic enough to "create, reproduce or alter the image, voice or manifestation" of a living, deceased, or fictional person. Justices Ricardo Villas Bôas Cueva and Floriano de Azevedo Marques dissented, arguing realism shouldn't be a threshold at all. Separately, by 4-3, the court declined to fine or sanction Flávio Bolsonaro. Rapporteur Kassio Nunes Marques, joined by André Mendonça, Dias Toffoli, and Antonio Carlos Ferreira, held that a party convention is an internal deliberative event, not public electoral propaganda — so the deepfake prohibition, which under TSE's Resolução 23.732/2024 (amending Resolução 23.610/2019) applies only to content already characterized as campaign propaganda, never attached.
The steelman for a stricter line
The dissent's concern deserves to be taken seriously. Jair Bolsonaro is legally barred from public office and, as an incarcerated felon, from the retail politicking that normally lets a political patriarch visibly bless a successor. A synthetic version of him doing exactly that — first shown at a convention, then inevitably clipped and recirculated across WhatsApp and social media — restores exactly the influence disqualification was meant to remove, regardless of the room it was first played in. Cueva's objection that deepfake status shouldn't hinge on "degree of realism" is also defensible: a crude but effective fake that most voters still believe is real does just as much damage as a photorealistic one, and tying the legal category to production quality invites litigation over pixels instead of harm. Add that a CNT/MDA poll found 84% of Brazilians want some restriction on AI-generated campaign content — half of them favoring an outright ban — and the pressure on the court to draw a hard line is obviously real, not manufactured.
Why the narrower ruling is still the better one
But the TSE's actual framework — set out in Article 9º-B and 9º-C of Resolução 23.732/2024 — doesn't ban synthetic political content; it requires disclosure. Anyone using AI-generated or manipulated audio, images, or video in a campaign must say so "explicitly, prominently and accessibly," with format-specific rules (a spoken disclaimer at the start of audio, a watermark plus audio note on images, both for video). Only deepfakes disseminating "notoriously false" facts, or content that isn't labeled at all, cross into the outright prohibition — one that carries real teeth: cassation of registration or mandate and a criminal referral under Article 323 of the Electoral Code. That is a proportionate architecture. It treats synthetic media as a labeling problem first and an existential threat only when it's used to deceive, which is the correct ordering of priorities for a technology that is otherwise just a new production tool for the same political speech campaigns have always run.
Seen against that framework, gating the deepfake prohibition on "electoral propaganda" status isn't a loophole — it's the same distinction Brazilian and most democratic electoral law already draws between internal party business and public campaigning, applied consistently to AI content instead of carved out for it. A blanket rule that any realistic AI depiction of a candidate's family member is automatically punishable, independent of audience or context, would sweep in campaign biography films, satire, and ordinary internal party media alongside genuinely deceptive material — exactly the kind of content-based overreach that chills legitimate expression under the guise of election integrity.
The real gap
The court's own logic exposes where enforcement should actually tighten: a video's status shouldn't freeze at the moment of first display. If a convention clip is deliberately built for downstream recirculation as an implicit endorsement, the disclosure requirement in Article 9º-B should travel with the content when parties or supporters repost it publicly — that's a labeling-compliance question, not a definitional one, and TSE already has the tool. The court's newly signed cooperation agreement with ElevenLabs to block unauthorized voice cloning points the same direction: build technical guardrails and disclosure infrastructure now, ahead of the October 4 first round, rather than expanding what counts as "propaganda" after the fact. That's the harder, less headline-friendly work — and the one actually suited to a technology that isn't going away.