OpenAI Files TSE Compliance Plan While ChatGPT Still Ranks Candidates
OpenAI filed its Brazilian election compliance plan with the Superior Electoral Court under Portaria 463/2026 and Resolution 23.610/2019, detailing technical and policy barriers meant to stop ChatGPT from recommending votes, favouring candidates or spinning mass campaign messages. The plan also describes Codex-assisted handling of TSE and regional court orders, C2PA and SynthID labelling for generated images, a ban on paid political ads, blocks on non-consensual intimate deepfakes, and routing of factual election queries toward official sources and Associated Press live tallies. A consolidated execution report is due by December 19, 2026, with material risk changes to be notified within forty-eight hours.
Filed under Policy and dated September 22, 2026, this AI4Brazil briefing treats the filing-versus-live-behaviour gap as Brazilian electoral-AI news distinct from the Chamber’s delayed PL 2338 calendar. VIVA’s early-September test still drew an explicit preference between São Paulo gubernatorial figures Tarcísio de Freitas and Fernando Haddad, including score-style comparisons that electoral rules treat as ranking. OpenAI itself notes probabilistic models cannot guarantee perfect neutrality—leaving prosecutors, platforms and campaigns in a season where paper plans and chat transcripts can diverge.
Why it matters: Brazilian voters already meet chatbots before ballots. Compliance PDFs can set duties—but only if live refusals match the TSE’s ranking ban.
What it means in practice
Brazilian counsel and campaign digital leads should inventory public chatbot prompts that invite rankings; demand vendor escalation paths to TSE orders; assign an owner for weekly red-team scripts; run time-boxed audits across ChatGPT and rival assistants; and prefer disclosures that admit residual error rates. Place the plan beside chatbot ranking probes and the delayed AI framework vote.
Caveats come first. A filed plan is not proven enforcement; screenshots age quickly; and rivals face the same failure modes. AI4Brazil therefore presents OpenAI’s TSE package as directional policy context until December’s execution report and court follow-ups appear.
What to watch next: any TSE fines or remediation orders; December 19 reporting; and cross-platform accuracy studies. 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 being tested in production chats and remain imperfect. Organizations that benefit most will verify refusals themselves, keep humans on campaign content, and refuse to confuse a compliance annex with finished neutrality.