AI chatbots assigned grades to the presidential candidates' campaign proposals and built rankings among them, in breach of a rule from Brazil's TSE, the Superior Electoral Court, that bans this kind of automated answer. The finding comes from a report by Folha de S.Paulo published on Wednesday (23), in the final stretch of the presidential campaign.
The norm is Resolution 23.755/2026, which changed Brazil's electoral propaganda rules, according to the specialized site desinformante. It bars AI systems from ranking, recommending or prioritizing candidates and from voicing opinions or any form of political favoritism in automated responses. Think of the poll worker at a voting station: they can tell you where to read each candidate's government plan, but never whom to vote for. When a chatbot hands out grades, it stops describing and starts judging, using criteria the user cannot see.
The pattern goes back to April
This is not the first time the problem has been documented. Between April 6 and 10, the Observatório IA nas Eleições, a project by Aláfia Lab and Data Privacy Brasil, ran the same 14 prompts through five assistants (ChatGPT, Gemini, Grok, DeepSeek and Meta AI). All five identified and profiled pre-candidates, and four recommended names for different voter profiles, with rankings of best proposals on the economy, public safety and education. In several cases, the report says, answers came with subjective evaluations, adjectives and opaque criteria. The research exists, explains Matheus Soares, content coordinator at Aláfia Lab, because people now ask assistants about politics directly instead of relying only on search engines.
"When a chatbot ranks or recommends pre-candidates, it stops merely providing information and starts influencing political perceptions. This is an important governance challenge so that these tools operate in compliance with electoral rules," said Carla Rodrigues, coordinator of Platforms and Digital Markets at Data Privacy Brasil.
The pattern continued. On August 21, JOTA Info reported that ChatGPT, Grok and Google's AI assistant were still ranking best candidates, and Tribuna do Planalto noted that Ronaldo Caiado topped the presidential lists on ChatGPT and Grok. In July, a complaint reported by the organization Conectas took the case to Brazil's Electoral Public Prosecutor's Office, arguing that ChatGPT, Gemini, Grok and DeepSeek can influence the freedom to choose a vote. And on September 4, Folha reported that a prosecutor's office is investigating big tech conduct and has asked the TSE for information about meetings with AI platforms.
On the corporate side, the commercial incentive points away from the rule. ChatGPT, Gemini, Grok, DeepSeek and Meta AI compete to be the user's default answer box, and a closed answer, with a grade and a winner, feels more helpful than returning a list of links to check. The April report concluded that the companies had yet to implement safeguards capable of meeting the TSE's determinations, and this week's coverage suggests the behavior persists in the official campaign.
What we still don't know, and what to do
It remains unclear whether the TSE or the Electoral Public Prosecutor's Office will impose penalties before the first round, or how many voters actually ask a chatbot for voting advice: there is no public measure of the reach of these answers. Nor is it known whether the companies adjusted their models after the complaints. The practical advice for voters is to use the assistant as a starting point: ask for proposals by theme, demand the sources, and check the government plans filed with the electoral court. Then repeat the question in a second chatbot. If each one returns a different ranking, the grade stops looking like a fact and shows what it always was: a choice made by the model.