Skip to main content

AI in Brazil's 2026 Election: Cloned Voices, Synthetic Voters and Missing Labels

Manipulation Breakdowns · 15 min read · By D0

A Speech He Never Gave

On the night of October 4, Flávio Bolsonaro gave a speech. He had just finished first in Brazil’s presidential election and would face Lula in a runoff. g1 published the video.

Within days a second version circulated. Same footage. A different voice track. In it, he says: “[…] we will change this Brazil, making cuts in the economy, starting by taking away the pension that many retirees receive unnecessarily […]” The fake also had him cutting the 13th-month salary and Pé-de-Meia, a stipend that keeps poor students in school.

He said none of it. In the original, he celebrates the turnout and says voters showed up because they wanted change. Aos Fatos found the source footage by running a reverse search on the viral video. At points the words and the lips do not match. The mouth distorts when his face turns off-centre. By October 8 the fake had at least 200,000 views on X.

Nobody has identified who made it.

This is one item. Between August 16 and September 28, the Lupa–Unicamp VigIA monitor confirmed 920 AI-made election posts on social media. 554 were deepfakes. Only 371 carried the AI label Brazil’s electoral court requires. That is about 40 percent.

The runoff is on October 25. Here is what has been documented so far.

1. A Cloned Voice Over Real Footage

The two fakes of Flávio that Aos Fatos debunked around the vote do not invent a scene. They take a real clip and replace the audio.

The pension fake is one. Aos Fatos debunked another on voting day itself. It used a January video in which Flávio talked about visiting a Holocaust memorial. The new audio had him tell workers they must choose “fewer rights and a job, or all rights and unemployment.” It drew at least 140,000 interactions on X and 83,000 on Meta’s platforms.

Viewers trust the clip because the picture is real. You recognise the room, the face and the broadcaster’s framing. Only the audio is fake.

Both of these target Flávio. They push the same message: he will take things away from you. Neither fact-check identifies who made them.

2. Synthetic “Ordinary Voters”

Other AI content skips the candidate’s face and invents people instead.

Aos Fatos found at least 50 TikTok videos of AI-generated “voters” giving interviews or testimonials about Lula and Flávio. Together they drew 7.4 million views. In 18 percent of them, nothing told the viewer the people were synthetic. TikTok said it removed them after Aos Fatos flagged them.

The best-known persona is “Dona Maria”, an AI-generated middle-aged woman who criticises Lula and the PT. One of her videos passed 8 million views, according to Tech Policy Press. In late April, three parties in Lula’s coalition asked the TSE to suspend her profiles. TikTok and Instagram removed her main accounts in May. Backup accounts appeared.

A synthetic voter is not quoting a politician, so it does not look like a deepfake of anyone. It looks like a neighbour with an opinion. That is the point. Viewers check what a candidate says. They rarely check a stranger’s complaint.

3. Amplification Through Co-Authored Posts

On September 23, Aos Fatos published research by Instituto Democracia em Xeque. It mapped 255 Instagram profiles the researchers call AI content farms, posting far-right content. Between August 15 and September 15 they published or co-signed 1,204 posts. Those posts drew 58 million views.

The network ran on Instagram’s “collab” feature. One post appears on several accounts at once and pools their audiences. Co-authored posts produced 63 percent of all engagement. Only 155 of the 1,204 posts, 12.9 percent, carried any AI identification, and only 134 disclosed it themselves.

The researchers do not attribute control of the network to any politician or party. Seven politicians or candidates, and the PL’s national profile, co-authored posts with it.

The institute did not describe the farm’s accounts as bots. Reach came through collabs, a feature the platform built. An electoral court order blocked the network’s original account. A backup profile stayed online.

Each campaign has also accused the other of running a network. In July, Lula’s federation asked the court to act against 32 Instagram and TikTok profiles it described as anonymous, posting AI content against Lula. In a filing reported on October 8, Flávio’s campaign alleged a coordinated network of 1,525 pro-Lula Instagram profiles. Those are court filings, not findings. Independent research so far covers the right-wing network in more depth than anything on the left. That may reflect which networks have been studied, not which exist.

4. Skipping the Label

Brazil’s rules are strict on paper. The electoral court, the TSE, updated them on March 2, 2026. AI content is allowed if it is labelled. Deepfakes in campaign propaganda are banned. AI chatbots may not recommend candidates.

In practice, the label rule is broken often. About 60 percent of the posts VigIA confirmed had none. VigIA also found that most deepfakes, 525 of 554, came from ordinary or anonymous accounts. Only 29 came from candidate or party profiles.

The court narrowed its own ban on September 1. The case concerned an AI avatar of Jair Bolsonaro endorsing his son at the PL party convention. By 4 votes to 3, the TSE refused to remove it. It held that a deepfake is only unlawful when it serves as electoral propaganda, and a broadcast party convention did not qualify.

Since then, the court has still acted against AI videos. On September 22, TSE president Kassio Nunes Marques issued a preliminary order for Flávio to remove an AI video about Lula’s son. He wrote that a notice about digital manipulation does not lift the deepfake ban. On October 8, the court ordered three of Flávio’s posts on X removed. One was an AI video with no label, and the vote on it was unanimous. The two AI removals rest on different grounds. One video carried a notice and was still a banned deepfake. The other had no label at all.

5. Posting Into the Blackout

The 2026 rules added a blackout. No new AI content showing a candidate may be published or boosted from 72 hours before each round until 24 hours after. That holds even if it is labelled.

Lupa monitored the first-round blackout, from October 1 to October 5. It found 64 AI-manipulated election posts. 37 were deepfakes. Four came from official candidate accounts. When Lupa published on October 7, all but three were still online.

The runoff blackout begins around October 22.

6. Fake Electoral Court Ads

On voting day, NetLab at the Federal University of Rio de Janeiro monitored political ads. Paid election ads are banned from 48 hours before the vote to 24 hours after. NetLab still found 568 irregular ads running on Meta’s platforms.

Six Meta ads copied the TSE’s name and visual identity. They offered a “remote vote” by phone or from home. Some sent users to pages that asked for their CPF, the national taxpayer number, to “start” voting.

This one does not need AI at all. It needs a logo, a believable excuse and a form. The risk is a voter who stays home thinking they have voted, and a stolen ID number.

7. Political Ads With Hidden Payers

The Meta Ad Library shows who paid for each political ad. Several ad operations this year hid the payer anyway.

Agência Pública found at least 1,069 Meta ads mentioning Lula or Flávio that were paid in US dollars. They cost at least US$120,800 and ran without the political-ad label. They reached more than 82 million impressions before Meta removed them. Some payers named PagoLink Global, a payment platform registered in Delaware and not available in Brazil. The ads targeted both candidates: one clipped a Lula remark out of context, a clip Agência Pública attributes to Deputy Nikolas Ferreira (PL-MG), and another attacked Flávio’s business ties.

Lupa traced one network in detail. It had 68 Facebook pages and 38 new sites posing as news portals. Most of the pages posted for the first time between September 3 and 12. Every site carries a footer declaring AI use. The network ran 1,006 ads costing at least US$106,000. 600 attacked Flávio, 138 attacked Lula, and 122 promoted the candidate Augusto Cury. The payer on 710 of them is “Corporativo Lam, LLC”, a Delaware company with no named owner. Cury’s campaign denied any link.

A third case ran on X. Between September 18 and 21, at least 11 foreign accounts published paid posts promoting Flávio. The posts carried a “paid partnership” tag without naming who paid. One account was UHN Plus, a Florida outlet co-owned by Spanish journalist Javier Negre that describes itself as “powered by Polymarket”. Aos Fatos, confirming a Folha report, found that who paid for the posts remains unknown.

For most of these ads, no one has shown who commissioned them. All three cases show money entering Brazil’s campaign through channels that hide the payer.

Foreign Pressure and Foreign Money

Recent Decipon breakdowns covered Russian operations in Europe. In Brazil, the documented foreign activity so far is mostly open, and most of it comes from the United States.

Brazil’s intelligence agency, Abin, raised it twice. Reuters reported on September 30 that Abin told visiting French and German officials it had “confirmed the existence of interference from the US as well as Russia.” The account comes from a report of the meetings compiled by an attendee, seen by Reuters. Abin gave no details of any Russian activity. It described the US side as ranging “from lawfare to interference that is openly acknowledged.” Separately, Folha columnist Mônica Bergamo revealed a restricted Abin report sent to the Presidency, the TSE and the Justice Ministry in late August. It rated the risk of US interference as critical. It meant political, economic and diplomatic pressure. It did not claim any threat to the voting machines or the count.

Some of that pressure is on the record. On August 16, US Under Secretary of State Sarah Rogers said X’s new election filter showed censorship imposed by Brazil. The filter stops X from recommending candidates’ posts to people who do not follow them. X said it was complying with the TSE’s rules. On September 30, the Guardian reported a State Department grant line of about US$1 million titled “Countering Lawfare and Censorship in Brazil”. Brazilian outlets relaying the report say it would fund groups opposing the Supreme Court. No recipient has been named. SBT News, relaying the Guardian, reports that the State Department released funds for a list of 43 projects that includes it, hours before the fiscal year closed on September 30. In late September, three Democratic US senators, Jeanne Shaheen, Peter Welch and Tim Kaine, wrote to Secretary of State Marco Rubio. They said they had “seen US officials engage in a campaign of interference and destabilization in Brazil.”

On voting day, TSE president Nunes Marques said the court had received no internal alert about foreign interference and had found “no kind of intervention”.

Searches up to October 12 found no public report from researchers or platforms tying a covert foreign operation to this election. That includes Russian networks and Israeli influence-for-hire firms. The AI fakes in sections 1 to 5 spread inside Brazil. Where a maker has been identified, the maker is Brazilian. The payers in section 7 have not been identified. The closest case is the 11 foreign X accounts in section 7, which ran paid pro-Flávio posts with the payer hidden. Aos Fatos links them to a transnational network, but no one has shown who paid.

What the Influence Tactics Score Would Show

We scored two items with the Influence Tactics Protocol. The overall score runs from 0 to 100. Higher means stronger manipulation patterns. Scores from 26 to 50 count as moderate.

The pension deepfake: about 50 (moderate, at the top of the band).

  • Suspicious Timing: Very high. It appeared in the days after the first round, aimed at the runoff.
  • Authority Overload: High. It used the candidate’s own face and a real g1 broadcast.
  • Emotional Manipulation: High. It threatens pensions and a stipend for poor students.
  • Missing Information: High. No creator, no context, no label.

The fake TSE ads: about 39 (moderate).

  • Suspicious Timing: Very high. They ran on voting day.
  • Authority Overload: Very high. They copied the electoral court’s identity.
  • Urgent Action Demands: High. Rain, illness and mobility were offered as reasons to vote by phone now.

The ads score lower because they are not divisive and carry little emotion. They aim to steal data and keep voters home, not to damage a candidate. A low score on division does not make them less harmful.

What This Is Not

No public report measures how many people believed a fake or whether a vote moved, so there is no evidence that AI content changed the first-round result. Both candidates are targeted, and nothing shows that one side runs the fakes. No covert foreign operation has been publicly detailed so far.

What to Watch For Before October 25

  • Voting is in person only. There is no phone or online vote in Brazil. Any ad offering one is fraud. Do not type your CPF into it.
  • Check who paid. Political ads on Facebook and Instagram list a payer in the Meta Ad Library. No payer, a foreign location or an unknown company is a reason to distrust the ad.
  • Find the original clip. If a candidate says something shocking, search for the event. A g1 or TV broadcast of the same speech will show what was actually said.
  • Watch the mouth, not the face. Cloned audio over real video often drifts out of sync, especially when the speaker’s face turns away from the camera.
  • Treat a missing label as a warning. The electoral court’s rules require it. A campaign-style AI video with no label is already breaking them.
  • From October 22, new AI content about candidates is banned. If you see it during the blackout, treat it as a red flag.
  • Be careful with “it’s AI”. We found no reported case so far of a Brazilian candidate in this election dismissing a real recording as fake. It may be next, as we covered in The Liar’s Dividend. If it happens, ask for the original file, not a denial.

If a clip of a candidate arrives on WhatsApp in the last week, find the broadcast it came from before you forward it. If you can’t find one, don’t send it.

About the Sources

Most of the evidence here comes from Brazilian fact-checkers, university monitors and investigative outlets. Several of these organisations published findings that embarrass Meta, such as the dollar-paid ads and the ads that ran on voting day. We found no foreign-government sponsorship notice on the sites of the sources cited here. Where a figure comes from one source only, the article says whose figure it is.

This article is part of Decipon’s Manipulation Breakdowns series, examining specific influence operations through the Influence Tactics Protocol.

Sources: