Brazil's Election AI Rules: What Chatbot Providers Must Control

What Brazil's 2026 election rules require

Resolution 23.755, issued by Brazil's Tribunal Superior Eleitoral on 2 March 2026, amended the country's election-advertising rules. Article 28 now prohibits AI application providers from ranking, recommending, suggesting or prioritising candidates, campaigns, political parties, federations or coalitions. It also prohibits automated opinions, voting recommendations and direct or indirect political favour or disfavour, regardless of a user's request.

The same resolution requires explicit labelling of synthetic multimedia used in election advertising. New AI-generated or altered content using a candidate's or public figure's image, voice or manifestation cannot be published, republished or promoted from 72 hours before until 24 hours after voting ends. Separate provisions require providers to offer a dedicated reporting route for candidates and political groups.

Why a generic refusal policy is insufficient

The rules create a product boundary, not merely a moderation statement. A global assistant may answer factual questions about registration, voting dates and candidate policies while refusing to rank candidates or recommend a vote. That distinction has to hold across languages, model versions, search tools, voice interfaces and follow-up prompts.

Testing published by campaign group Eko on 18 August examined ChatGPT, Grok, Gemini and Meta AI using three voter personas and 25 election questions per system. Eko reported recommendations and factual errors across the tested services. Its findings are advocacy research rather than a regulatory ruling, but they illustrate the operational risk: a compliant policy can still fail when context, retrieval or response routing changes.

Build a jurisdiction-aware political control stack

Providers and businesses offering public-facing assistants should combine several controls:

• Jurisdiction detection: identify when Brazilian election rules apply without collecting more personal or location data than necessary.
• Political-domain routing: distinguish neutral civic information from candidate ranking, persuasion, synthetic media and other restricted requests.
• Current evidence: retrieve election facts from approved, dated sources and expose the source and update time to reviewers or users where appropriate.
• Consistent refusals: test direct prompts, coded language, comparisons, voice requests and multi-turn attempts across every supported model and channel.
• Escalation and records: preserve privacy-conscious evidence, triage reports quickly and link every remediation to an owner, version and retest date.

Make the compliance plan measurable

Article 125-B requires internet application providers to prepare a compliance plan covering relevant duties, measurable periodic indicators, implementation deadlines, expected results and a trajectory for reaching them. That structure is useful beyond Brazil. A defensible plan should report restricted-request detection, successful bypasses, factual-error rates on approved civic questions, time to correct stale data, complaint response time and regression-test coverage by model and language.

Executives should ask for evidence at the service level, not a single global average. A system can pass in Portuguese text and fail in voice, or comply on its default model while an experimental route recommends a candidate. Release gates should therefore bind legal rules to exact models, retrieval sources, interfaces and jurisdictions.

What business leaders should do next

Inventory every public assistant that can answer political or civic questions, including embedded vendor chatbots, customer-service agents and search experiences. For each one, identify the provider, model, location logic, retrieval sources, refusal policy, reporting route and accountable owner. Then run a Brazil-specific evaluation before 4 October and after every material model or prompt change.

ELYMENT AI's guides to legal AI governance, incident reporting and human approval show how to connect rules to operational evidence. ELYMENT AI can help teams translate jurisdictional obligations into test cases, release gates and accountable response workflows without turning legitimate civic information into a blanket ban.

This article is general business analysis, not legal advice. Organisations operating in Brazil should obtain advice on how the resolution applies to their specific service.

Sources

• Tribunal Superior Eleitoral: Resolution 23.755 (2 March 2026) - Official resolution covering political recommendation restrictions for AI providers, synthetic-content rules, reporting routes and measurable compliance plans.
• Tribunal Superior Eleitoral: 2026 election-advertising explainer (10 July 2026) - Official explanation of the 2026 election-advertising timetable and the court's updated rules for online and AI-generated content.
• Reuters: Brazil races to rein in AI before the election (1 September 2026) - Independent reporting on early enforcement activity, chatbot testing and the practical challenge of applying Brazil's rules before the October vote.
• Eko: AI chatbot testing for Brazil's election (18 August 2026) - Advocacy research describing its multi-persona test design and reported recommendation and factual-accuracy failures across four major chatbots.

Continue learning

• California SB 574: What Legal AI Governance Should Include
• AI Agent Incident Reporting: The Metrics Businesses Need
• Where AI Agents Need Human Approval

Frequently asked questions