Brazil Battles Rising AI Misinformation From Chatbots at the Ballot Box

Brazil’s Superior Electoral Court has banned AI chatbots from offering voting advice as authorities fight biased algorithmic influence ahead of the 2026 elections.

Brazil’s electoral rules barred AI chatbots from offering users voting tips in their responses. However, the six months leading up to Brazil’s presidential election saw chatbots breaking this rule with impunity by answering user enquiries on the best candidate, among others. Significantly, this follows the head of Brazil’s electoral court (TSE), Justice Carmen Lucia’s warning in June about the possibility of chatbots contributing to AI misinformation regarding the country’s October vote. Also, the court had imposed new regulations in March to regulate how chatbots operate during the 2026 election cycle, as well as make platforms liable for false content.

Furthermore, the TSE has prohibited AI tools from providing recommendations, rankings, or opinions about candidates and political parties, even upon receiving user prompts. According to tests conducted by Agence France-Presse (AFP) weeks after this rule was announced, three leading chatbots continued to rank political candidates. They provided varied responses that were noticeably biased in favor of particular candidates on the basis of experience, age, and political ideology, among other variables. This is viewed as a concerning trend towards technology being capable of influencing voting in a highly polarized and hyper-connected Brazil by referring to incorrect information.

Moreover, all platforms refused to take responsibility for their chatbots’ responses, with OpenAI stating that ChatGPT is “trained not to favor candidates” and is continuously refining its models. On the other hand, Google pointed to the role of user prompts shaping Gemini’s responses, which do not necessarily reflect the company’s views. CIO Bulletin views this standoff between the TSE and AI chatbots as reflective of the latter’s replies being susceptible to errors since these are generated by probabilities based on training data.