São Paulo — The world's most popular AI chatbots are lying to Brazilian voters about their own election, and appear to be breaking the law while doing it, according to new research from corporate accountability group Ekō, published months before the country's October general election.
Examples cited in the report include Meta AI telling users that Jair Bolsonaro was never convicted of attempting a coup — months after Brazil's Supreme Court sentenced the former president to 27 years in prison for exactly that. Gemini provided a phone number for Brazil's electoral court that does not exist. ChatGPT named presidential candidates who had already dropped out of the race months earlier. And Grok pushed the unsupported far-right claim that Brazil's courts and government are coordinating to suppress free speech.
Researchers also found that chatbots answered identically-worded questions differently depending on the user's apparent politics — in some cases the chatbots offered voting recommendations entirely based on the user's presumed leanings. Asked whether to vote for Lula, Grok told the conservative profile “No, voting for Lula makes no sense based on the profile you described” — yet recommended Lula to the progressive profile.
This pattern could have legal consequences in Brazil. The country’s electoral authority, the TSE, has made it illegal for AI chatbots to directly or indirectly recommend, rank, or favor candidates. Three of the four chatbots tested appear to be breaking that law.
Vicky Wyatt, Ekō Campaign Director said: “Experts have warned for years that AI could supercharge the disinformation crisis - on the brink of this high stakes election the warning signs are flashing red. These companies need to put safeguards on their chatbots now — and Brazil's electoral authority needs to enforce the rules it already has on paper."
“Millions of Brazilians will turn to these chatbots looking for answers as they prepare to vote but they’re being fed false information and conspiracy theories that appear to break Brazilian electoral law.
Ekō researchers tested ChatGPT, Grok, Gemini and Meta AI, four of the most widely used AI chatbots in Brazil, which is one of the largest AI chatbot markets in the world. Researchers created three accounts for each chatbot posing as Brazilian voters with different political leanings — one Bolsonaro-aligned, one Lula-aligned, one with no stated views — and put the same 25 questions about the election to each. Every response was checked against Brazilian electoral law and public reporting, and rated for accuracy.
Meta AI was the most inaccurate of the four, followed by Grok and ChatGPT, with Gemini making the fewest errors — though every chatbot tested returned a meaningful share of false or misleading answers.
Breaking the results down further, by individual persona rather than chatbot overall, the least accurate result came from Meta AI when answering as a Lula-aligned user — answering 32% of the prompts with inaccuracies.
Brazilian lawmakers and electoral officials are already moving to tighten oversight of AI, wary of a repeat of 2022 — but this research suggests existing rules aren't being enforced. Ekō calls on AI companies to fix the errors identified in this research and build real safeguards, and on Brazil's electoral authority to act before votes are cast in October.
ENDS
For more information about a current Ekō campaign or to arrange an interview with a spokesperson please contact us at