It sounds like a Black Mirror plot, but it was a real audit. Meta approved 13 out of 15 fake political ads, generated with artificial intelligence and aimed at Brazilian users, that the British organization Ekō submitted to the platform from an account based in the United States. The findings were released on Thursday, September 17, in a report by the newspaper Folha de S.Paulo, written by reporter Patrícia Campos Mello.
One thing needs to be clear: no Brazilian ever saw these ads. According to Folha, the pieces went through the standard ad approval process on Instagram and Facebook, without ever being published or shown to users. It was a stress test, and Meta's automated moderation waved through 87% of the ads that broke rules set by the TSE, Brazil's electoral court.
The research was conducted in August. Each ad carried a synthetic image created with chatbots such as ChatGPT, Grok and Gemini, modeled, according to Ekō, on conspiracy narratives already circulating on the Brazilian far right. None of the approved ads carried the AI-generated label required by a TSE resolution, and none was classified on the platform under "social issues, elections or politics", as electoral rules require.
What the fake ads said
The approved pieces included a fake picture of TSE president Kassio Nunes Marques announcing that the election date had changed; a synthetic image of Donald Trump with Senator Flávio Bolsonaro and the caption "vote Flávio", celebrating an official endorsement that, according to Ekō, never happened; an ad claiming polling places would be closed and the vote postponed to November; another urging people to "retake Brasília, as we did on January 8", a reference to the 2023 storming of government buildings in the capital; and an image stating that "PCC criminals are breaking the voting machines". Every piece was created by Ekō itself, with no connection to official campaigns.
Why does it matter? The TSE bars campaign advertising produced from abroad or bought through third-party accounts: in Brazil, only candidates and parties may buy internet ads directly from the platforms, which keeps the spending on the books. Ekō said the US-based account raises fears of foreign interference in the election. And of the two rejected ads, neither was blocked for incitement or disinformation: both were flagged only for "deceptive or misleading business practices".
"Meta learned nothing from its role in Bolsonaro's attempted coup. Four years on, its systems are still waving through ads inciting violence, hate speech and disinformation, weeks before Brazilians go to the polls again."
The statement comes from Vicky Wyatt, Ekō's campaign director. It is the group's second study on AI and Brazil's 2026 presidential race: in August, Ekō tested ChatGPT, Grok, Gemini and Meta AI and found that nearly a quarter of Meta AI's answers about the election were partly or entirely wrong. The TSE already considers it illegal for chatbots to recommend candidates, directly or indirectly.
Where to look: the full report is on Ekō's website, eko.org, in English. To keep an eye on real ads, Meta runs a free, open Ad Library where anyone can search ads classified under "social issues, elections or politics" by keyword or advertiser.