AI Regulation Tracker / Court order, Brazil
A synthetic voice is still the speaker: Brazil tests the AI avatar workaround
If your organisation has ever asked whether a generated likeness could carry a message the real person is not permitted to deliver, a Brazilian court order signed on 28 July 2026 is the cleanest statement yet of why that plan does not work.
What did the court actually order?
One thing. In the dispositive paragraph the Justice writes that, under arts. 21 and 243 of the Internal Rules of the Supreme Federal Court, he orders the defence to submit a statement within 48 hours clarifying whether authorisation was given for the use of the image and voice of Jair Messias Bolsonaro in the AI-produced video used to publicise his son's campaign launch. Counsel are to be served electronically and the Prosecutor General's Office notified.
There is no finding that the video broke any rule, and no revocation of anything. Press coverage in several languages has framed this as a deepfake ruling. It is not one. It is a court asking a question with a clock on it, in a docket where the answer changes which body of law is in play.
Why does the answer matter so much?
Because the two possible answers point in opposite legal directions, and the order says so. The recitals record that the person in the proceeding is subject to a judicial prohibition, in the court's own capitalised wording, on the "DIVULGACAO DE MANIFESTOS POLITICOS-ELEITORAIS, INCLUSIVE POR TERCEIROS, INDEPENDENTEMENTE DO MEIO UTILIZADO". Our rendering: dissemination of political-electoral manifestos, including through third parties, whatever the medium. His political rights are suspended under art. 15, III of the Federal Constitution.
Read the order against that. If he agreed to the video, the order states this would constitute a new and serious transgression of the judicial decision, liable to reversion of the humanitarian house arrest to the closed regime. If he did not agree, the order states the conduct may amount to what is known as a deep fake, a practice expressly prohibited by electoral legislation, and identifies the third parties as the candidate Flavio Nantes Bolsonaro and the Liberal Party. Note the word "may". Both branches are conditional.
That is the portable part. The instrument is a phrase in a Brazilian criminal-execution docket, but the reasoning is what any regulator will reach for: a restriction attaches to a person's communications, and routing those communications through a generated version of that person does not create a new and unrestricted speaker.
Why does a disclosure label not settle it?
The order reproduces the video's opening words: "Isso que voces estao vendo aqui, nao sou eu. Isso e uma simulacao da minha imagem e da minha voz feita com inteligencia artificial." Our rendering: what you are seeing here is not me; this is a simulation of my image and my voice made with artificial intelligence. The video announced itself as synthetic.
Now read the rule the Justice quotes. Art. 9-C of Resolution TSE 23.610/2019 bans fabricated or manipulated content in electoral propaganda. Paragraph 1 goes further:
Sec. 1. It is prohibited to use, in order to harm or to favour a candidacy, synthetic content in audio or video format, or a combination of both, that has been digitally generated or manipulated, even with authorisation, to create, replace or alter the image or voice of a living, deceased or fictitious person (deep fake).Art. 9-C, para. 1, Res.-TSE 23.610/2019. Operative Portuguese: "ainda que mediante autorizacao". English rendering ours.
Four words carry the weight. "Ainda que mediante autorizacao" means the consent of the person depicted is not a defence. The prohibition is written as a rule about the electoral process rather than about the individual's personal rights, so the individual cannot waive it away.
Disclosure sits in a separate provision. Art. 9-B of the same resolution requires whoever is responsible for the propaganda to state explicitly, prominently and accessibly that content was fabricated or manipulated and which technology was used. A label discharges art. 9-B. Nothing in the text of art. 9-C, paragraph 1 says a label discharges that paragraph. We are describing the structure of the rule, not asserting how it applies to this video, which is what the court has not decided.
How does this compare with the disclosure-first regimes?
Most jurisdictions writing AI-election rules chose a labelling duty, on the theory that an informed voter is a protected voter. Brazil kept a labelling duty and layered an outright prohibition on top of it. The comparison below is not about severity. It is about whether consent and disclosure are answers.
| Instrument | Core mechanism | Does the subject's consent cure it? | Does a disclosure label cure it? |
|---|---|---|---|
| Brazil, art. 9-C para. 1, Res.-TSE 23.610/2019 | Prohibition on synthetic audio or video of a real person used to harm or favour a candidacy | No. The text reads "even with authorisation". | The text contains no disclosure exception. |
| Brazil, art. 9-B, Res.-TSE 23.610/2019 | Duty to state explicitly and prominently that content is fabricated or manipulated, and which technology was used | Not the question the article asks | Yes. Disclosure is the obligation. |
| South Korea, art. 82-8, Public Official Election Act | Ban on realistic AI-generated campaign material inside a 90-day pre-election window; labelling required outside it | Not the operative test inside the window | Outside the window, a clear label is the duty |
| Louisiana, Act 806 (HB 459), effective 1 August 2026 | Disclosure requirement for AI-created or AI-manipulated candidate likeness in electioneering communications | Not the operative test | Yes. Clear disclosure is the compliance path. |
The rulebook this order reaches for is covered in Brazil's electoral court codifying its deepfake ban and AI labelling duty for 2026. The two disclosure-first models in the table are South Korea's 90-day ban and labelling duty and Louisiana's Act 806.
What should a compliance or communications team do with this?
If you advise anyone subject to a restriction on their own communications, this is now a live question in your file.
- Treat a synthetic likeness as the depicted person's channel, not as a separate asset. A settlement gag clause, a regulatory bar, a court order: all are drafted around the person, and a generated version of that person is where the restriction gets tested.
- Stop relying on a signed likeness release as the end of the analysis. Brazil's rule is written so that the subject's authorisation does not cure it, because the interest protected is the integrity of the vote rather than the individual's image rights. Any rule protecting a process rather than a person behaves the same way.
- Separate your labelling control from your permission control. They answer different provisions, and a team that has built only the labelling control has built half a system.
- If you produce or host political content in Brazil for the 2026 elections, note that paragraph 2 of art. 9-C carries cassation of the registration or of the mandate.
- Log who approved a generated likeness and on what basis. The first question this order asks is a records question. A vendor who cannot answer it in 48 hours has a problem independent of the merits.
What did the court rely on, and what is a different case?
The order cites two decisions of the Superior Electoral Court. One, published on 8 April 2026 (AgR-AREspe 060036660), holds that using the image of a political figure without agreement amounts to unlawful propaganda. The other, published on 20 May 2026 (AgR-REspe 060020163), restored a fine of R$15,000 against a 2024 candidate for mayor of Fortaleza who posted a TikTok video in which internationally known public figures appeared with artificially manipulated speech simulating support for his candidacy. Its headnote states that the prohibition in art. 9-C, paragraph 1 is objective in nature, so proof that voters were actually misled is not required.
That fine belongs to the Fortaleza case. No penalty has been imposed in EP 169 over this video. The order also notes Resolution TSE 23.757 of 2 March 2026, whose art. 6, paragraph 4 treats AI-generated synthetic content used in breach of electoral rules as misuse of the means of communication and, depending on circumstances, as abuse of political and economic power.
What this order does not do
It bans nothing. It does not decide that the video was a deepfake, that anyone breached a court order, or that any party is liable. It does not revoke the house arrest; it says reversion to the closed regime is a possible consequence in one branch. It is not a decision of the full court or of the electoral courts, and it creates no rule binding anyone outside EP 169. Where the press has reported a finding, we have not, because the text does not contain one.
Frequently asked questions
Did Brazil's Supreme Federal Court rule that the video was an illegal deepfake?
No. The 28 July 2026 despacho in EP 169 decides nothing about the video. It orders the defence of Jair Messias Bolsonaro to state, within 48 hours, whether the use of his image and voice in the AI-produced video was authorised. The order sets out two possible readings and says the answer is needed to determine which applies.
What does art. 9-C, paragraph 1 of Resolution TSE 23.610/2019 actually prohibit?
In the official Portuguese text it prohibits the use, to harm or to favour a candidacy, of synthetic audio or video content that has been digitally generated or manipulated, in the words of the rule "ainda que mediante autorizacao", meaning even with authorisation, to create, replace or alter the image or voice of a living, deceased or fictitious person. The provision was inserted by Resolution 23.732/2024.
Does labelling a video as AI-generated take it outside the Brazilian prohibition?
Labelling and prohibition are two separate provisions. Art. 9-B imposes a duty to state explicitly and prominently that content was fabricated or manipulated and which technology was used. Art. 9-C, paragraph 1 is a flat prohibition on synthetic audio or video of a real person used to harm or favour a candidacy, and its text contains no exception for disclosure. Meeting the labelling duty in art. 9-B does not, on the face of the text, answer the prohibition in art. 9-C.
What is the penalty exposure under art. 9-C of the Brazilian resolution?
Art. 9-C, paragraph 2 states that breach of the caput and of paragraph 1 constitutes abuse of political power and misuse of the means of social communication, carrying cassation of the registration or of the mandate. Separately, in a different case published on 20 May 2026, a 2024 Fortaleza mayoral candidate had a fine of R$15,000 restored for an AI-manipulated campaign video.
Last verified: July 28, 2026