ChatGPT, Grok, and Google’s AI search tool illegally recommended, according to Brazilian electoral law, candidates for Brazil’s 2026 presidential vote. It also gave numeric scores for each person’s policy plan. JOTA, a Brazilian news outlet, said this may break a guideline that stops AI systems from backing or promoting specific candidates.
The checks were done on Aug. 18 and Aug. 19. Each model got the same prompts on both days. The prompt asked the systems to judge proposals from five top figures. It covered public safety, health care, schools, and the economy. It also asked for a final overall view.
Results differed by tool. With a review that leaned on the policy texts, ChatGPT ranked Ronaldo Caiado of PSD first. It put President Luiz Inacio Lula da Silva next, then Romeu Zema, then Flavio Bolsonaro, and last was Renan Santos.
Google’s AI Mode came out with a different order. It listed Zema first, then Caiado, then Flavio Bolsonaro, then Lula, and ended with Santos.
Grok also used a full list. It placed Caiado first, Flavio Bolsonaro second, Zema third, Santos fourth, and Lula fifth.
ChatGPT gave Caiado the top score on the economic plans. Zema came next. Then Santos. After that, Lula and Flavio Bolsonaro.
The Superior Electoral Court in Brazil, the TSE, passed Resolution 23.755/2026. Under this rule, AI tools cannot “rank, recommend, suggest or prioritize” candidates, parties, or coalitions. The same limits also cover any direct or indirect electoral preference. The systems are not allowed to push voters toward specific choices.
Legal experts told JOTA that the chatbots answers broke the rule. Rogério Mehanna said this is a serious issue. He pointed out that many users see AI as neutral and fair. Because of that, people may treat the output as reliable. Guilherme Barcelos, a constitutional law specialist, said the AI cannot sort candidates or give them point values.
Google said it could not recreate JOTA’s findings. Google said its products are meant to stay neutral. It also said its AI features in search still produce answers that reflect the user’s question and what is available on the open web.
OpenAI said ChatGPT was trained to not steer people toward any candidate. It also said it will not comply with requests that ask users who they should vote for. In a statement, OpenAI said it is still improving its systems. It added that it will work in a practical way with election groups worldwide. The company said it keeps a close back and forth with the TSE. It said this is to help protect Brazil’s democratic process.
The report said Grok’s parent company, X, did not reply to JOTA’s request for comment before publication.
JOTA also quoted experts who said this did not prove a hidden political plan. Filipe Medon, a law professor at Fundação Getulio Vargas in Rio de Janeiro, said the safer guess was that the protections meant to stop disallowed answers did not work as intended. He said platforms have to block direct votes advice and also plan for the many ways people will try to get around the rules.
A different report from the corporate accountability group Ekō said several AI chat tools gave wrong election details. It also said the systems gave advice about voting, either on purpose or by implication. The researchers looked at 300 replies. Those answers came from Meta AI, ChatGPT, Grok, and Gemini. They were tested using 25 questions. The questions were asked through three fake Brazilian voter types. The testing ran from June 24 to July 1.
Meta AI showed the worst results. In its 75 replies, 24% were judged partly or fully wrong. ChatGPT and Grok both had a 12% error rate. Gemini was lower, at 8%. The report said that this clashed with a TSE rule.
