Brazil is defining what a deepfake is, five weeks before it votes

Brazil votes on October 4, and its electoral court spent Tuesday working through a question that many regulators are still struggling to answer: what exactly counts as a deepfake?

The Tribunal Superior Eleitoral was expected to rule on broader restrictions for AI-generated campaign material, Reuters reported.

The proposed rules go beyond simply banning deceptive deepfakes. They would require AI-generated campaign material to be labelled, prohibit systems that recommend candidates, and introduce a blackout period during which new synthetic political content could not be published or promoted immediately before and after voting.

The definition matters because Brazil has already encountered a case that sits awkwardly between disclosure and deception.

In July, Jair Bolsonaro, who is under house arrest and barred from public communication, appeared at his son’s campaign launch as an AI-generated avatar that explicitly identified itself as a simulation.

That disclosure is central to the legal argument. Brazil’s existing electoral rules target fabricated content that can mislead voters, but an AI avatar that openly tells viewers it is an AI representation does not fit neatly into that definition.

Lula’s coalition petitioned the TSE, arguing that the video still violated rules governing synthetic content used to influence voters.

Flávio Bolsonaro’s campaign responded that a clearly labelled simulation should not be treated as a deepfake, an argument that is difficult to dismiss if the definition is based primarily on whether viewers are being deceived.

The court itself has shown some disagreement over where that line should sit. TSE president Kássio Nunes Marques has suggested that AI-generated campaign material can be permissible when it does not cause harm, while Supreme Court justice Alexandre de Moraes has warned that the same technology could violate Bolsonaro’s house-arrest restrictions and contribute to electoral disinformation.

Brazil already had rules covering AI before this election. The TSE introduced requirements ahead of the 2024 municipal elections, including disclosure of synthetic content, which helps explain why the Bolsonaro avatar was presented as a simulation rather than as an attempt to impersonate him without disclosure.

Tuesday’s ruling is expected to provide a clearer definition of what is and is not permitted, although the timing is difficult.

With five weeks remaining before a national election, campaigns, platforms and regulators will have very little time to adjust to whatever framework the court establishes.

The problem Brazil is trying to address exists far beyond its borders. Disclosure-based rules, including those being developed under the EU’s AI Act, are built around the assumption that telling people when content has been generated or manipulated by AI can reduce its ability to mislead them.

The Bolsonaro case puts that assumption under pressure because the content can be completely transparent about its synthetic nature and still be politically persuasive.

There is also a question that the existing rules were not really designed to answer. A politician who is legally prohibited from speaking publicly but appears through an AI-generated representation has not physically addressed the crowd, yet the political message can reach voters almost as effectively as if the politician had been standing there in person.

The proposed restrictions on recommendation systems introduce a separate problem because they move beyond the content itself and into the way platforms distribute political material.

Prohibiting algorithms from recommending particular candidates would require regulators to examine ordinary platform mechanics, which is considerably more difficult to monitor than whether an individual video carries an AI disclosure.

Brazil has more experience with this kind of intervention than many countries. Its courts have been willing to take a more active role in regulating online platforms and political speech, and the TSE had already established AI-specific electoral rules before the current campaign began.

Enforcement is likely to be more difficult than defining the rules. A blackout on synthetic political content around election day is useful only if someone can identify synthetic material quickly enough to stop it from spreading, which becomes particularly difficult when content moves through private messaging groups rather than public social networks.

The blackout provision is also the least tested part of the proposal. Treating AI-generated political content as a separate category that requires a quiet period around an election is a significant step beyond existing European rules, which generally focus on transparency and the circumstances in which synthetic content can be misleading.

European regulators therefore have a practical reason to watch what happens in Brazil.

The EU’s own transparency requirements rely on some of the same assumptions about disclosure, while European elections have yet to produce a comparable case involving a politician who is prohibited from communicating but can still reach voters through an AI-generated representation.

The scale of Brazil’s electorate makes enforcement harder still. More than 150 million people are eligible to vote, and messaging apps are deeply embedded in political communication, much of it taking place in private conversations and groups that regulators cannot realistically monitor in the same way as public posts.

Brazil is therefore trying to establish the boundaries while the campaign is already under way, and the political actors affected by those boundaries have a direct interest in how they are drawn.

Leaving the rules undefined would create its own problems, but establishing them only weeks before voters go to the polls leaves very little room for regulators, campaigns and platforms to learn how they work.

The election takes place on October 4, leaving the TSE with roughly five weeks for whatever rules it approved on Tuesday to be tested in practice.