On April 7, 2026, Brazil's Chamber of Deputies opened consideration of PL 212/26, a bill by Deputy Rubens Pereira Júnior (PT-MA) that would make the creation and distribution of AI-generated electoral deepfakes a crime. Anyone who produces or circulates synthetic audio or video to defame candidates, parties, or public officials and sway voters would face two to six years' imprisonment plus fines — penalties raised by one-third if the conduct occurs between candidate registration and election day. The bill amends three statutes at once: the 1965 Electoral Code, the 2014 Marco Civil da Internet, and the 2018 LGPD data-protection law.
The timing is not accidental. Brazil votes in October 2026, and as Tech Policy Press notes, this will be the first national contest in which generative-AI tools are "cheap, widely accessible and used at a massive scale."
The strongest case for the bill
The problem PL 212/26 targets is real and measurable. The Digital Forensic Research Lab documented 78 confirmed or alleged deepfake cases during Brazil's 2024 municipal elections between May and October — 65 of them aimed at mayoral candidates, spread across 67 cities, and distributed mostly through WhatsApp. Nearly 80% clustered in a handful of malicious categories: false crime accusations, fabricated media appearances, and non-consensual pornographic fakes of five female candidates. These are not satire. They are targeted reputational attacks engineered to deceive, and the harm lands fastest in the final days of a campaign when there is no time to rebut a convincing fake.
Proponents can fairly argue that administrative remedies move too slowly for that window, and that the deterrent weight of criminal law is what serious electoral fraud has always carried. Pereira Júnior's own justification is that existing rules are "insufficient to guarantee a rapid response." On the merits of the threat, he is not wrong.
But the courts are already on this
Where the bill overreaches is in assuming Brazil starts from a blank slate. It does not. In February 2024 the Superior Electoral Court (TSE) issued Resolution 23.732, whose Article 9º-C already prohibits the use of deepfakes "to harm or favor a candidacy," treating violations as abuse of political power that can cost a candidate their registration or mandate. Article 9º-B requires explicit, prominent disclosure of any AI-generated electoral content. The TSE carried that framework into the 2026 cycle and tightened it further.
So PL 212/26 does not fill a regulatory vacuum — it stacks a prison sentence on top of an administrative regime that is functioning. That distinction matters for proportionality. Cassation of a mandate is a sanction calibrated to the electoral context; incarceration for two to six years is the machinery of criminal law, with all the chilling reach that implies.
Vagueness is the speech problem
The bill's drafters deserve credit for building in carve-outs: good-faith sharing without knowledge of falsity is not punishable, and satire, parody, and identified academic research are exempt so long as the material is labeled and not meant to deceive. Those exceptions are exactly the right instinct. The difficulty is that they sit on top of terms the law does not crisply define.
What separates a punishable "deepfake intended to defame and influence voters" from a sharp-edged political meme, an exaggerated campaign ad, or an AI-assisted edit of a genuine gaffe? Tech Policy Press warns that the TSE itself has not settled whether content must intend to deceive or whether any realistic AI creation qualifies — and legal experts fear that ambiguity could drive "excessive litigation against legitimate expression." Criminal penalties amplify that risk enormously. A satirist confident of an acquittal still faces investigation, prosecution, and legal costs; the rational response is to self-censor near an election, which is precisely when robust political speech matters most. The exemptions protect you only after you have proven you qualify for them.
A proportionate path
None of this argues for doing nothing. It argues for matching the tool to the harm. The clearest abuses — non-consensual sexual deepfakes, fabricated criminal accusations, voice-cloned "confessions" — are already crimes under defamation, fraud, and Brazil's recent AI-and-image-abuse statutes, and they should be prosecuted as such. For the electoral sphere specifically, the TSE's rapid administrative remedy plus mandatory labeling addresses the speed problem without criminalizing a category of expression defined this loosely.
The platform side of PL 212/26 is more defensible: a 24-hour takedown obligation after notice from the Electoral Justice or the affected party, backed by fines of up to 1% of Brazilian revenue, is a transparency-and-notice mechanism rather than a speech crime. Even there, lawmakers should ensure the notice runs through judicial or electoral-authority channels rather than letting any "affected party" trigger removals, to avoid weaponized takedowns.
Brazil is right to take electoral deepfakes seriously. But a democracy defends itself best by punishing fraud precisely, not by hanging a six-year sentence over a definition its own courts admit they have not pinned down.