Brazil’s AI Election Rules Face Their First Big Test Ahead of October Vote

Brazil heads into its October 4, 2026 first-round election with one of the more detailed electoral AI frameworks in place. The hard part is no longer simply writing rules. It is applying them quickly and consistently when inexpensive synthetic videos, images and chatbot answers can reach voters before their source, context or legal status is established. 12

What the TSE’s AI rules require

Brazil’s Superior Electoral Court (TSE) has set several limits for candidates, parties, platforms and AI-enabled campaign material:

• Prominent disclosure: Electoral content created or significantly altered with AI or equivalent technology must carry an explicit, prominent and accessible notice. The obligation applies across text, audio, video and images.
• Ban on electoral deepfakes: The court prohibits deepfakes in electoral propaganda. On September 1, the TSE defined a deepfake, by a 5–2 vote, as synthetic content produced or manipulated with AI or equivalent technology that has realism or verisimilitude and creates, reproduces or alters the image, voice or expression of a living, deceased or fictional person. The court also held that the prohibition applies when the material is electoral propaganda.
• No AI voting recommendations: Chatbots and virtual assistants may not recommend votes, favor candidates or parties, rank them, or prioritize political information in a way that benefits an electoral actor—even when a user asks for that guidance. 14
• Election-period blackout: The TSE prohibits the publication, republication and paid boosting of new AI-generated or AI-altered content involving the image, voice or likeness of candidates or public figures from 72 hours before voting through 24 hours after polls close. 12

The framework distinguishes between AI use that is disclosed and permissible and synthetic material that crosses into prohibited electoral propaganda. That distinction is crucial: a rule against “AI” broadly would sweep up ordinary editing and creative campaign production, while a narrow rule can leave difficult cases at the boundary.

The Flávio Bolsonaro avatar case exposed the boundary problem

The highest-profile test involved Flávio Bolsonaro’s campaign launch, where an AI-generated version of former President Jair Bolsonaro appeared. Jair Bolsonaro was barred from the election and under house arrest, making the simulated appearance politically consequential. 3

Flávio Bolsonaro’s lawyers said the video did not solicit votes and that the technology was not concealed. 8 Opponents argued that the clip breached rules against deepfakes. 7

The TSE’s subsequent ruling clarified that not every AI depiction automatically constitutes a prohibited deepfake. The material must meet the court’s realism-based definition and qualify as electoral propaganda. That approach leaves important fact-specific questions for future cases: Was the material clearly identified as synthetic? Was it shown in an internal party setting or distributed as campaign advertising? Did it seek electoral support? And could a reasonable viewer take it as authentic?

Those questions illustrate the regulator’s core challenge. Disclosure can help voters recognize synthetic material, but it does not necessarily neutralize the persuasive force of a realistic likeness, especially when a clip is edited, reposted or stripped of its original label.

Why synthetic campaign media can damage election integrity

AI can support routine political communication, including stylized images and clearly fictional visuals. The higher-risk uses are those that make invented events, endorsements, statements or apparent evidence look real. Reuters reported that Brazilian authorities were concerned about AI-generated content spreading through voters’ social media feeds ahead of the election. 2

The most consequential forms can include:

• fake footage of a candidate speaking or acting;
• invented endorsements or voter testimonials;
• fabricated polling claims or momentum narratives; and
• synthetic portrayals designed to resemble documentary evidence.

The risk is not merely that a false clip exists. It is that it can create an impression of first-hand proof, broad public support or a sudden electoral shift. A correction may arrive later, reach fewer people, or fail to restore confidence in what voters have already seen. That is the logic behind both mandatory labels and the tighter restrictions around election day. 12

Chatbots are an enforcement problem, not just a content problem

The TSE’s rule against AI candidate recommendations is unusually direct: it applies even when a voter explicitly asks a chatbot whom to support. 14 But the rule’s practical reach is uncertain because general-purpose AI systems are operated globally and respond dynamically to individual prompts.

Research published by corporate accountability group Ekō found that ChatGPT, Grok and Gemini appeared to recommend candidates in Brazil, while nearly a quarter of Meta AI responses tested were partially or entirely incorrect. The findings do not by themselves establish how any particular provider will comply in every interaction. They do show the gap between a formal electoral prohibition and the real-world behavior of widely used systems.

For regulators, that creates several difficult tasks at once: identifying an unlawful answer, preserving the interaction as evidence, determining which service or platform is responsible, obtaining a timely remedy, and ensuring that a corrected system behavior is applied at scale. A Brazilian rule can set obligations for election activity in Brazil, but it cannot by itself guarantee that every globally deployed chatbot will provide compliant answers to every voter query.

The unresolved test: speed and consistent enforcement

The 2026 rules give the TSE clearer tools than a general ban on misinformation: disclosure requirements, a defined deepfake standard, restrictions on recommendation systems and a time-based blackout around voting.

Yet their effectiveness will depend on rapid enforcement, platform cooperation and factual judgment in cases that do not fit neatly into “real” or “fake.” The Bolsonaro-avatar dispute showed that realism, consent, labeling, distribution and electoral purpose can all matter. As AI-generated political media becomes easier to make and harder to trace, Brazil’s election will test whether a detailed rulebook can keep pace with the technology it seeks to govern.

Available reporting does not provide sufficiently reliable underlying data to state a definitive total for AI-related TSE cases, specific removal orders, or national figures for Brazilians’ AI use and willingness to consult chatbots about candidates. Those claims require the original court data or survey methodology before they can be treated as established facts.