TSE Tells Prosecutors AI Chatbots Show Extremely Responsible Behaviour
Brazil’s Superior Electoral Court told the Electoral Prosecutor’s Office that generative AI chatbots have generally shown “extremely responsible” behaviour by warning users they cannot give electoral recommendations, Folha de S.Paulo reported on September 24, 2026. The note, sent earlier in September by the court’s disinformation advisory unit, answered a probe opened after civil-society reports alleged ranking and favouritism that could breach TSE rules barring chatbots from ranking or suggesting candidates and parties.
Filed under Policy and dated September 28, 2026, this AI4Brazil briefing treats the TSE reply as Brazilian election-AI oversight news distinct from yesterday’s Sedigi AI-companion child-risk note. The court said it contacted X, OpenAI and Google about reported failures; companies stressed continuous testing and framed issues as point failures, while TSE argued the harder problem is users paraphrasing bans with softer verbs such as “rank” or “prioritise,” and suggested clearer technical notes plus compliance-plan evidence for prosecutors.
Why it matters: Brazilian voters will meet chatbots throughout the 2026 cycle. Transparent court–prosecutor dialogue can tighten safeguards—but only if age gates, audit logs and human complaint paths become enforceable product requirements before runoff season.
What it means in practice
Brazilian product, platform and counsel leads should inventory ranking-like prompts in election modes; demand named owners for TSE compliance-plan updates; assign an owner for prosecutor document requests; run time-boxed red-teams on paraphrase jailbreaks; and prefer designs that keep humans on escalation when political advice is requested. Connect the reply to OpenAI’s TSE compliance plan and the electoral prosecutors’ chatbot probe.
Caveats come first. A reassuring court note is not a dismissal of the probe; Anthropic, DeepSeek and Perplexity were outside some compliance-plan thresholds; and journalistic jailbreaks can still force rankings. AI4Brazil therefore presents the TSE language as directional electoral context until formal decisions conclude.
What to watch next: whether prosecutors seek court orders; how many platforms file simplified plans; and whether TSE publishes didactic examples of banned prompts. Readers can continue on the AI4Brazil homepage, or browse the Newsroom for additional briefings.
Bottom line: treat this update as orientation, not instruction. Brazilian election AI rules are still maturing through institutional letters and remain early. Organizations that benefit most will treat compliance as engineering, keep humans on political escalations, and refuse to confuse a polite TSE adjective with finished enforcement.