In Brazil’s 2026 election, a voter may ask a chatbot which candidate best matches her priorities. Under Resolution No. 23.755/2026, which amended Brazil’s standing electoral-propaganda rules, the Superior Electoral Court, known as the TSE, prohibits AI providers from ranking, recommending, or prioritizing candidates, parties, or campaigns, even at a user’s request. Automated responses may not favor a political actor directly or indirectly.
This marks a significant shift. Most election rules concentrate on deepfakes, fabricated claims, and undisclosed synthetic material. Brazil has moved upstream to the moment when AI converts political information into judgment. A chatbot can influence an election even when every fact in its answer is true.
The weak link beyond deepfakes
Democratic safeguards were largely designed for political influence that could be seen. Speeches can be challenged and advertisements archived. Even a viral claim or deepfake leaves an object that different people can examine together.
Stay ahead of the geopolitical week.
MD Briefing delivers expert analysis across five global fronts — the Indo-Pacific, energy, geoeconomics, European security, and the Middle East — every Monday morning. Free.
A chatbot interaction takes place in private and adapts to the user’s wording. Its answer may change across languages or conversation histories. A candidate may never know how an AI system describes her to voters, and there is no single message for regulators to archive.
The influence may lie in selection rather than fabrication. A model decides which candidates deserve mention and whose program appears first; it also determines which sources count as credible. This is custodial bias at the electoral layer: the system shapes who remains politically visible before making an explicit recommendation. A smaller party may disappear because its digital footprint is thin. Truthful sentences can still produce a distorted political map.
Evidence from the Netherlands illustrates the problem. Before the 2025 general election, the Dutch Data Protection Authority found that two parties appeared as the top recommendation in more than 56 percent of chatbot responses. Ahead of the 2026 municipal elections, it found that chatbots frequently ignored local parties. Such distortion requires no secret partisan instruction.
Why Brazil moved first
Brazil reached this rule through experience, not abstract foresight. Its institutions remember what happened when private digital communication moved faster than electoral oversight.
The 2018 presidential election made WhatsApp a central political infrastructure. Messages moved through family and community networks, acquiring credibility from the person who forwarded them. Research identified WhatsApp as a primary channel for politically motivated false information traveling through encrypted, socially trusted spaces.
By 2022, the target had widened from political opponents to the legitimacy of electronic voting and the electoral authority itself. The refusal to accept the result culminated on 8 January 2023, when supporters of former president Jair Bolsonaro attacked Congress, the Supreme Court, and the presidential palace. For Brazil’s electoral institutions, digital disinformation had become part of a chain leading from private persuasion to institutional violence.
The TSE converted that experience into institutional memory. Its counter-disinformation work began in 2017 and later became permanent. In 2024, the court required disclosure of AI-manipulated campaign material and prohibited political deepfakes. In 2026, it extended its attention to AI-generated electoral advice. The new rules will become particularly consequential from 16 August, when electoral propaganda is permitted to begin.
This learning curve helps explain why Brazil moved more quickly than many wealthier democracies. Recent political trauma meets an unusually powerful institution: the TSE serves as both court and election administrator, turning lessons from one cycle into rules for the next.
A sharper exposure across the Global South
Brazil’s experiment matters because the conditions that increase conversational AI’s power recur across emerging democracies. Adoption is outpacing institutional capacity to test political behavior, while literacy remains uneven. Where citizens rely on trusted intermediaries to interpret complicated choices, a personalized machine can acquire authority without a democratic mandate.
Unequal data intensifies the risk. National parties leave thick digital trails; local candidates may rely on fragmented sources. Guardrails can also behave differently in dialects or poorly documented contexts.
This produces a double asymmetry. Voters most exposed to authoritative AI answers may have the least capacity to verify them, while countries most in need of safeguards often have the least leverage over global AI providers. A platform can update its system from abroad, yet the consequences appear inside a local election it may only partly understand.
Southeast Asia shows why the warning is immediate. The Philippines entered its 2025 elections with rules covering social media, AI, and internet campaigning, but their center of gravity remained campaign content and digital platforms. Indonesia’s 2024 election featured synthetic political personas and interactive tools across roughly 205 million registered voters. During Thailand’s February 2026 general election, concern over manipulated material led TikTok to launch an election information center with the Election Commission, Cofact, and Thai PBS Verify. These responses still concentrated on visible campaign content, platform integrity, and deepfakes rather than the private electoral advice produced by general-purpose chatbots.
Exposure is uneven. Large democracies face the problem at scale. Timor-Leste, preparing for a presidential election in 2027, represents a smaller and less documented edge of the same risk. A study of its media sector identified election periods as fertile ground for misinformation and disinformation. In a multilingual environment with thinner political records and sharp gaps between Dili and rural areas, chatbot omissions may be difficult to detect. Direct evidence of electoral-chatbot influence remains limited, making Timor-Leste a prospective warning rather than an established case.
The next vulnerability may emerge outside formal campaigns. Voters can ask global AI systems to interpret crowded candidate fields and local histories. Safeguards built around nationally visible politicians may perform far less reliably for a municipal candidate in the Philippines, a district contest in Indonesia, or a political figure speaking in Tetum.
Brazil and Europe are regulating at different speeds.
The European Union provides the closest comparison. Under Annex III of the AI Act, systems intended to influence voting behavior are high-risk. Voter-advice applications may operate under requirements for risk management, documentation, and human oversight.
The contrast sharpened in June 2026, when the EU’s Digital Omnibus extended the Annex III deadline from August 2026 to December 2027. Brazil was preparing to enforce a narrow behavioral restriction as Europe delayed its broader high-risk regime by 16 months.
The distinction also lies in intended purpose. A dedicated voter-advice application clearly falls within the European category. A general-purpose chatbot may not, even when users turn it into an electoral adviser. Europe regulates a defined risk category; Brazil intervenes when political information becomes personalized advice. They are moving at different speeds and targeting different moments of influence.
Brazil’s model is a pioneering Global South governance experiment whose enforcement deserves close observation, not an automatically superior solution.
A necessary experiment, with unresolved costs
Early testing already shows how large the compliance-readiness gap may be. In April 2026, the Observatório IA nas Eleições submitted 14 identical prompts to ChatGPT, Gemini, Grok, DeepSeek, and Meta AI. All five were profiled pre-candidates and produced thematic rankings; four recommended candidates to simulated voter profiles. The report identified the outputs as apparent non-compliance with the TSE’s new restrictions, while stressing that the tests took place before the official campaign period and therefore did not establish completed electoral infractions. Because the official campaign period had not begun, it did not present them as completed electoral infractions. Even so, they revealed how much current model behavior must change before the rule can bite.
The rule may cause chatbots to refuse legitimate political questions too broadly. A system trying to avoid liability could decline to compare programs or help a voter navigate official information. People able to consult many sources would retain that capacity, while those needing accessible explanations could lose a useful tool.
Compliance may favor global providers able to build election-specific safeguards, while local AI companies face heavier costs. Auditing private conversations raises questions about data protection and expression. Indirect political advantage will be difficult to test because bias may appear as omission or ordering rather than endorsement.
Brazil’s 2026 election will therefore be a stress test. The test goes beyond whether chatbots refuse to name a preferred candidate. Regulators must determine whether official electoral information remains accessible and consistent across languages. They must also examine whether ostensibly factual answers quietly erase smaller political actors.
Other Global South democracies should adapt the principle rather than copy the rule. Election authorities can separate factual information from personalized recommendations and ground procedural answers in official sources. Independent tests should examine which political actors remain visible across repeated prompts and local languages. Providers should document material model changes during election periods. Provenance must extend beyond AI-generated images to the sources, model version, and ranking criteria shaping a political answer without demanding private chain-of-thought reasoning.
Brazil learned during the social-media era that private digital intermediaries can reshape public democracy before institutions can see the damage. Its 2026 rule turns that experience into anticipatory governance. For Southeast Asian democracies, the warning has arrived before the next electoral cycle rather than after it.
The next threat to electoral integrity may arrive without a viral lie or an obvious deepfake. It may sound like a calm, personalized, and seemingly neutral answer to a voter who simply asks, “Whom should I trust?”
