Brazil's TSE rules deepfake ban applies only to electoral propaganda

Brazil's electoral court, the TSE, ruled on 1 September 2026 that a deepfake ban applies only when the material qualifies as electoral propaganda, assessed on content, language, audience and context. Whether something is synthetic is therefore not enough: judges must test qualification, message and missing AI labelling separately.

The Tribunal Superior Eleitoral (TSE) set down a legal definition of deepfake: synthetic or manipulated content with sufficient realism that creates, reproduces or alters the image, voice or utterance of a living, deceased or fictional person. Equally important is what the court ruled alongside this: the ban reaches such material only when it qualifies as electoral propaganda, and that qualification depends on the message, the language used, the addressees and the context in which it was circulated.

On 8 October 2026 the TSE applied that line. It unanimously ordered the removal of three posts in which AI had been used. A simulation of Lula that was not identified as AI-made breached, according to the court, the transparency duty in Article 9-B of Resolution 23.610/2019. The TSE describes that duty as requiring an express, clear and accessible identification of political or electoral material created or edited with AI. Strikingly: according to the court the material had a caricatured, fanciful appearance, but that did not release the maker from the labelling duty.

Why does the TSE rule that not every deepfake falls under the electoral ban?

Because the ban hangs on a legal category, not on a technique. The TSE separates three questions that must not be conflated: is the material synthetic or manipulated, does the message fall within the statutory category of electoral propaganda, and has the mandatory AI identification been complied with? In our analysis, the TSE decisions indicate that a removal order may follow where material qualifies as electoral propaganda and the applicable mandatory AI-identification duty has not been met; the precise consequence remains case-specific.

That distinction shifts the centre of gravity. A detector that establishes that a video is synthetic answers only the first question. The second question is legal in nature and calls for an assessment of the content, language, addressees and circumstances. The labeling question is a separate compliance inquiry under the applicable rule.

How do other jurisdictions handle the reliability of possibly AI-made evidence?

Our analysis is that the Polish and EU material discussed here places AI-related evidence within existing evidence frameworks, while the Irish practice directions add explicit verification duties for material used in the proceedings to which they apply. A peer-reviewed study in the Journal of Modern Science concludes that Polish and EU rules have no separate regime for AI-generated evidence: such material is admitted and weighed under free assessment of evidence, with authentication possible through experts among other routes. The study points out emphatically that labelling under the EU AI Act (Regulation (EU) 2024/1689) is not necessarily complied with by makers of forged material, and that the right to contest evidence effectively remains a central safeguard.

The Commonwealth Lawyers Association describes the same risk from two sides: deepfakes can lead to the false acceptance of forged material and to the false rejection of genuine material. The authors advocate layered authentication through provenance, integrity, independent corroboration, forensic examination and genuine contestability, with the judge remaining responsible for admission, weight and procedural fairness. The Courts Service of Ireland translated this into practice directions for the High Court and Court of Appeal, which came into effect on 1 September 2026. Within the proceedings to which they apply, material created with generative AI must be independently verified under human control. The directions also warn that fabricated, manipulated or deepfake material must not be used or presented to the court, requiring particular care with the authenticity of exhibits, images and recordings. Non-compliance can lead to costs decisions, rejection of pleadings or disciplinary referral.

A workable sequence for a deepfake claim

• Separate the question of whether the material is synthetic from the question of what it legally means.
• Preserve the original file, the metadata and the provenance chain before anything is shared or processed.
• Compare the content against independent sources rather than relying on the visual impression.
• Record whether the mandatory AI identification was present, separate from the authenticity question.
• Give the opposing party a real chance to contest both the material and the authentication method used.

What does this mean for executives, lawyers, electoral bodies and newsrooms?

In our analysis, because the TSE ties the ban to qualification as electoral propaganda rather than solely to the technique, lawyers and enforcers should not rely on a detector result alone; they should document content, audience, language and context separately and assess any labelling defect under the applicable rule. Because a caricatured or fanciful image, according to the TSE, does not exempt from the transparency duty, campaign teams that publish AI material without clear identification risk removal even without deceptive intent; in our view it is prudent to set up a fixed labelling rule and a final check before publication. Because the Irish direction can attach costs and other procedural consequences to non-compliance, decision-makers would, in our assessment, be wise to update their internal protocol for independent verification of AI material rather than wait for a national equivalent. For the broader line, see our topic hub on AI governance and decision-making.

This fits earlier developments in which judges did not accept AI material at face value, such as a case about an AI video of a victim in a court case, and the growing verification duty for AI citations on the signing lawyer. What matters is what labelling does and does not do: an AI label can disclose that AI was used, but it does not by itself establish that the underlying material is authentic, just as an AI watermark as a weak provenance signal remains. The TSE line moreover applies within Brazilian electoral law and is not globally binding; its value is that it shows how a supreme court translates an abstract norm into qualification, message and labelling, each tested separately.

Sources: The article draws on two decisions of Brazil's Tribunal Superior Eleitoral, a study in the Journal of Modern Science, an analysis by the Commonwealth Lawyers Association and a practice direction from the Courts Service of Ireland.