A test built for a specific video
On September 1, 2026, Brazil's Superior Electoral Court (TSE) voted 5-2 to fix a legal definition of what counts as an illegal deepfake under Article 9º-C, § 1º of Resolution nº 23.610/2019, as amended by Resolution nº 23.755 of March 2, 2026. Per the TSE's own summary, a deepfake now requires synthetic content "produced or manipulated through artificial intelligence or equivalent technology, with a degree of realism or verisimilitude," that creates, reproduces, or alters the image, voice, or manifestation of a real, deceased, or fictional person — and the ban only bites when that content qualifies as electoral propaganda (TSE). Ministers Villas Bôas Cueva and Floriano de Azevedo Marques dissented.
The case that forced the question was almost tailor-made for it: an AI-generated video of the imprisoned former president Jair Bolsonaro endorsing his son Flávio's candidacy, screened at the Liberal Party's national convention on July 25, 2026. The opposition Brasil da Esperança federation (PT, PCdoB, PV) sought fines and removal. Rapporteur Kássio Nunes Marques argued the video was shown at a closed party event merely streamed online, not campaign advertising — a framing that let the majority sidestep ruling the underlying content itself illegal.
The steelman for a broader ban
Critics of a narrow test have a real point. Justice Estela Aranha and civil-society groups including EFF, Access Now, and Data Privacy Brasil have argued that Brazil's 2026 elections face genuinely new manipulation risks — "AI avatars and synthetic characters that simulate real voters," bot-amplified synthetic content, and microtargeted disinformation that a narrow, case-by-case realism test may be too slow or too permissive to catch before it spreads (EFF, Access Now, Data Privacy Brasil). A bright-line rule — any AI manipulation of a candidate's likeness is banned, full stop — is easier to enforce at platform scale and doesn't ask an electoral judge to adjudicate, video by video, whether a given deepfake was "realistic enough." Given Brazil's history with WhatsApp-driven disinformation in the 2018 and 2022 cycles, caution is not paranoia.
But that case-by-case cost is also the point. A blanket ban on any AI-altered depiction of a candidate — regardless of whether anyone could mistake it for real — criminalizes satire, campaign parody, obviously stylized ads, and the disclosed, labeled synthetic content that Resolution 23.755 already requires campaigns to flag explicitly. The TSE's own architecture already assumes AI use is normal and disclosable: the resolution mandates that campaigns label AI-generated material "in an explicit, prominent and accessible manner," naming the specific technology, and imposes an absolute 72-hour pre-election and 24-hour post-election blackout on new synthetic candidate content regardless of labeling, with fines of roughly R$5,000 to R$30,000 (~$970–$5,800) plus possible registration cancellation for violations (The Rio Times). Layering a strict-liability deepfake ban with no realism or intent threshold on top of that disclosure regime would punish transparency rather than deception — exactly backwards.
Why the three-part test is the right calibration
The test the TSE adopted — AI creation or alteration, a realism threshold capable of deceiving a viewer, and characterization as electoral propaganda aimed at influencing the vote — tracks the actual harm the resolution exists to prevent: a voter mistaking fabricated content for real evidence of a candidate's words or actions. It excludes what Justice André Mendonça, whose framework the majority substantially adopted, described as "memes, caricatures, animations and other obviously fictional content produced with artificial-intelligence tools." That line matters because Brazil already has one of the world's more aggressive AI-election frameworks — mandatory labeling, a hard pre-vote blackout, criminal liability for sexualized deepfakes — layered on top of a general deepfake ban. A test with no deception threshold would have made that entire apparatus redundant with a much blunter instrument: anything AI-touched near a candidate is illegal per se.
The narrower test also gives the TSE room to actually enforce the blackout and labeling rules aggressively — the parts of the framework that scale to millions of pieces of content — without also litigating, video by video, whether every satirical AI clip clears a bar it was never meant to have to clear. That is a more sustainable division of labor than asking judges to police the entire universe of AI-touched political content under a single undifferentiated ban.
What to watch
The unresolved question is how consistently lower electoral courts apply "realistic enough to deceive" once the marginal cases — an AI voice clone with a disclosure label buried in small text, a photorealistic image posted without context — start arriving ahead of October's vote. The TSE's own August 14 ruling ordering removal of a different, undisclosed photorealistic Flávio Bolsonaro video suggests the realism threshold has real teeth when disclosure is genuinely absent. That is the test working as intended: catching deception, not AI use as such.
The Court also defined that the prohibition on the use of deepfake in elections requires that the content be characterized as electoral propaganda — TSE, September 1, 2026