Brazil Study Finds 1,423 AI Election Posts

BRAZIL · ELECTIONS

A Brazilian fact-checking tracker has counted 1,423 AI election posts since mid-August. More than half carried no warning label.

The Brazil election 2026 runoff on Sunday, 25 October, follows a campaign flooded with AI election posts. The fact-checking agency Lupa now counts 1,423 of them, and 716 involve deepfakes.

A deepfake is a fake image, video or voice that shows a real person doing or saying something they never did. The latest Brazil election odds put conservative senator Flávio Bolsonaro near 83% against President Lula.

What the Lupa Tracker Found

VigIA is a monitoring project run by Lupa with the Recod.AI laboratory at the University of Campinas (Unicamp). Its live dashboard covers 16 August to 22 October and was updated on Friday, 23 October.

Lupa’s database lists 1,486 analysed posts, of which 1,423 were confirmed as AI-made or AI-altered. Another 32 showed no AI and 31 were inconclusive.

Lupa sorts the 1,423 posts into three types. There are 707 synthetic posts, 520 deepfakes and 196 deepfakes mixed with other synthetic content.

Images made up 751 posts and videos 667. Instagram carried 794, Facebook 690, X 105, TikTok 72, WhatsApp 3 and YouTube 3, and a post can appear on several.

Who Was Behind the Posts

Ordinary users made 560 of the 1,423 posts (39.4%), and official candidate accounts made 548 (38.5%). Anonymous, unknown or avatar accounts made 300 (21.1%), and political parties made 8.

Candidate accounts have grown as a source. Lupa’s first-round report of 1 October counted 226 posts from candidate profiles among 920 posts.

Lupa founder Cristina Tardáguila said in that report that campaigns know they can be punished, so misuse is left to militants. Researcher Beatriz Farrugia told Agência Brasil that courts would need platforms’ help to identify anonymous accounts.

Exame, a Brazilian business magazine, published an analysis on Saturday, 24 October, that rests on the earlier count. It quoted deepfake researcher Bruno Sartori saying the content was less harmful than feared, mostly memes.

Who Was Shown and How

President Luiz Inácio Lula da Silva, of the left-wing Workers’ Party (PT), appears in 509 posts, the most of any public figure. Flávio Bolsonaro, a conservative senator of the Liberal Party (PL), appears in 247.

Flávio’s father, former president Jair Bolsonaro, appears in 120 posts, ahead of Supreme Court justice Alexandre de Moraes with 56. Jailed ex-banker Daniel Vorcaro appears in 51 and congressman Nikolas Ferreira in 48.

By purpose, 768 posts (54.0%) promote a candidacy and 360 (25.3%) attack someone. Humor or satire accounts for 264 (18.6%) and neutral information for 31.

By 1 October Lupa had found 30 or more deepfakes of TV Globo anchors, including William Bonner and Renata Lo Prete. They appeared to read out fake poll numbers, including a supposed Lula defeat to Flávio.

The Labeling Rule and the Gap

Brazil’s electoral court, the TSE, requires AI content in campaigns to carry an explicit, prominent label. Lupa’s database shows 768 of the 1,423 AI posts (54.0%) without a label and 655 (46.0%) with one.

Among the unlabeled posts, 386 came from ordinary users and 188 from anonymous accounts. Candidate accounts made 185 and parties 5.

The rule sits in article 9-B of Resolution 23.610, as amended by Resolution 23.755 of 2 March 2026, the TSE says. Deepfakes that favour or harm a candidacy are banned, with a fine of BRL 5,000 to BRL 30,000 (about US$990 to US$5,950).

Platforms must remove such content, and the court can cancel a candidacy or mandate for abuse of power. A separate ban covers the final days and applies to Sunday’s runoff.

From 72 hours before the vote to 24 hours after, new AI content altering a candidate’s voice or image is banned. Before the first round Lupa found 64 such posts between 1 and 5 October, only 4 from candidates.

How Platforms and the Court Responded

Lupa reported on 15 October that 37 posts by candidates and parties were flagged in the first 11 days of the campaign. Of those, 29 were still online on Tuesday, 13 October, and the rest had been removed.

Lupa sent those cases to the TSE on 11 September, and they were forwarded to the electoral prosecutor-general’s office on 16 September. The TSE told Lupa that complaints should go straight to that office.

Meta said it is committed to protecting the electoral process and that Lupa’s sample does not reflect the scale of its work. TikTok said it removed 348,963 videos in Brazil between 1 June and 6 September for civic-integrity, election, misinformation and deceptive-AI violations.

X did not respond to Lupa before the first-round report. Poder360 counted 14 of 75 pre-campaign court actions between the Lula and Flávio camps (19%) as AI disputes.

What It Means for US Readers

A clip that looks like a Globo bulletin with a poll result may be fake, as Lupa found 30 or more. US readers and investors who follow the vote on social feeds should wait for official results.

Meta, X, TikTok and Google are US companies, and the TSE rules require platforms to act on illegal content in Brazil. Related coverage includes a court order on false posts, a pulled pro-Lula AI video and a guide to verifying AI-generated content.

More context is in the Brazil news hub and the Brazil guide for newcomers. Nothing in the tracker suggests a direct US sanction, tariff or legal step.

What Prediction Markets Say

On Polymarket, a US prediction market, bets gave Flávio Bolsonaro 83% and Lula 17.5% (24 October, 1:07 p.m. ET). About US$195 million has been traded across the Brazil presidential election markets.

These prices are bets, not polls. They do not measure whether any AI post changed a vote.

Why we show this: prediction markets turn real-money bets into a live probability that moves within minutes of the news, which is why investors, campaigns and newsrooms in the United States now follow them closely. We show them next to polls and official results, never instead of them.

What Is Not Known

The dashboard is live, so its totals change as Lupa reviews posts. The effect of the posts on votes is not measured.

Lupa’s own breakdowns use different categories at different dates, so the 1 October report and the dashboard are not strictly comparable. Exame’s analysis uses the earlier count.

The exact hour when the ban on new AI content began was not stated in the reports reviewed. The TSE’s full resolution text was not opened, so the rule is quoted from its news page.

What Comes Next

Brazilians vote in the runoff on Sunday, 25 October, and Lupa says it will keep monitoring AI content through the second round. Its final VigIA report is listed for November.

The tracker does not show that any post decided the first round, and it does not mean the result will be disputed. Courts can still order removals and fines before and after the vote.

Frequently Asked Questions

What is a deepfake?

A deepfake is AI-made or AI-altered media showing a real person doing or saying something they never did. Lupa lists 520 posts as deepfakes and 196 more as deepfakes mixed with other synthetic content.

How many AI election posts has Lupa found?

Lupa’s live dashboard shows 1,423 AI-made or AI-altered posts between 16 August and 22 October. Its first-round report of 1 October had counted 920.

Who made the AI posts?

Ordinary users made 560 posts (39.4%) and official candidate accounts 548 (38.5%). Anonymous or avatar accounts made 300 (21.1%).

Is AI content legal in Brazilian election campaigns?

AI content is allowed if clearly labeled, but deepfakes that favour or harm a candidacy are banned. New AI content altering a candidate’s image or voice is also banned from 72 hours before the vote to 24 hours after.

Does this mean the vote is unsafe?

No. The tracker counts misleading posts, not votes, and it says nothing about how ballots are counted. It measures the information environment around the vote.

Sources: Lupa VigIA first-round report, 1 October (earlier count); Lupa, 7 October; Lupa, 15 October; Lupa VigIA live dashboard (Raio-X do VigIA), 16 August to 22 October; Electoral court (TSE), AI rules; Agência Brasil via Acessa, 2 October; A Crítica de Campo Grande, 2 October; Exame, 24 October; Poder360, 16 June; Exchange rates, open.er-api.com (all accessed 24 October 2026; conversions use BRL 5.04 per US$1 on 24 October 2026).