Brazil presidential election dogged by alleged bias by Meta AI tools

Meta AI is a chatbot integrated into WhatsApp. More broadly, Meta uses automated systems across WhatsApp, Facebook and Instagram to generate content, apply safety rules, recommend posts and moderate political material. These tools matter because they can shape what voters see and what campaigns can publish.

During Brazil’s election, users asked Meta AI to make campaign posters. Images showed it refusing Lula-related requests while producing an image for “Flávio”. Lula’s campaign also submitted examples of the system assigning Bolsonaro’s number, 22, to fictional politicians. It allegedly translated “Flávio futuro presidiário” as “Flávio future president”.

Meta said its systems blocked prompts using “Lula” and “Bolsonaro”, but not “Luiz” or “Flávio”. It said Lula’s advertising account was suspended for one hour by mistake. The dispute shows how automated tools can influence political communication while remaining difficult for outsiders to inspect.

Lula’s campaign said Meta treated the two presidential campaigns unequally. It formally accused the company of suspending Lula’s campaign advertising account, suppressing posts and blocking AI-generated content. Brazil’s electoral prosecutors also said Meta had not fully followed court orders requiring equal treatment.

The campaign submitted screenshots to the electoral court. They allegedly showed Meta AI assigning Bolsonaro’s electoral number, 22, to generic fictional politicians. Another example allegedly translated “Flávio futuro presidiário”, meaning “Flávio future prisoner”, as “Flávio future president”. Images also showed the chatbot refusing to create posters for Lula while producing one for “Flávio”.

Meta denied favoring either candidate. It said Lula’s account was suspended for one hour by mistake. It also said prompts containing “Lula” and “Bolsonaro” were blocked, while “Luiz” and “Flávio” were not. United Nations experts expressed concern about possible unequal treatment.

AI-generated and manipulated election content was widespread before Brazil’s first voting round. Fact-checking organisation Lupa identified 920 pieces, mainly on Facebook and Instagram. The volume matters because voters may encounter persuasive political material without knowing whether it is authentic, altered or computer-generated.

The investigation found that 60 per cent of the identified content was unlabelled. Another investigation found 26 Meta advertising accounts that appeared to bypass restrictions on foreign-funded political advertising by posting AI-generated videos. One video depicted Bolsonaro and Argentina’s president Javier Milei freeing Brazil from communists.

The article describes this election as Brazil’s first since large language models became widely popular. Meta’s independent oversight board said the company’s election safeguards had failed. Experts therefore viewed Brazil as an early indication of how AI could affect political campaigns across Latin America and beyond.

Presidential elections can turn on what voters see, trust and share. If a platform suppresses one campaign’s posts, blocks its advertising or lets another campaign bypass restrictions, candidates do not compete on equal terms. AI-generated misinformation adds another risk because realistic images, videos or text can spread quickly and confuse voters.

Brazil’s case offers concrete examples. Lupa found 920 pieces of AI-generated or manipulated election content before the first round, with 60 per cent unlabelled. Another investigation found 26 Meta advertising accounts apparently sidestepping restrictions on foreign-funded political advertising. Meta AI also appeared to produce different responses about Lula and Bolsonaro.

The article does not prove that these incidents changed the result. However, Lula’s campaign credited an intensive WhatsApp operation with helping its comeback, while polls showed the gap narrowing. Meta’s oversight board said election safeguards had failed, highlighting the need for fairer, clearer systems.

Meta owns WhatsApp, Instagram and Facebook, so its rules and automated systems influence political advertising, posts and generated content. The company decides how to apply those systems and says it did not favor either candidate. It acknowledged a one-hour mistake involving Lula’s advertising account.

Brazil’s electoral courts issue orders requiring equal treatment, while electoral prosecutors investigate whether platforms follow them. Lula’s campaign formally accused Meta and submitted screenshots to the electoral court. The article also describes prosecutors saying Meta had not fully complied with court orders. Bolsonaro’s campaign is the other side of the election contest, but the article does not describe it filing a complaint.

These roles create an oversight chain. Meta operates the technology, campaigns challenge decisions, and public authorities seek to enforce electoral fairness. United Nations experts and Meta’s independent oversight board add scrutiny. Yet opaque systems make it difficult to determine whether repeated discrepancies are deliberate bias or technical failure.

When an AI platform refuses a political question, the clearest alternative named in the article is an official electoral website. These sites can provide authoritative election information without asking a chatbot to summarize candidates or interpret disputed claims. That matters because chatbot answers may be incomplete, inconsistent or overly cautious.

InternetLab found that some of eight major chatbots refused genuine questions about candidates and directed users to official electoral websites. Yet the same systems sometimes agreed to draft misleading claims about election integrity for campaign use. This contrast shows why voters should prefer primary electoral information over generated answers.

The article also identifies electoral courts and prosecutors as public institutions involved in checking campaign treatment. Voters can therefore consult official electoral information and court-related announcements, while treating campaign statements as partisan claims that require comparison. The article does not name specific websites or provide a complete source list.

Large language models, or LLMs, generate text by using learned patterns to predict a response from a prompt. Image systems similarly turn instructions into visual content through trained statistical models. Separate safeguards can block names, topics or requested outputs. These systems matter politically because their answers can shape campaign messages and public understanding.

The article gives several examples. Meta said its systems blocked prompts containing “Lula” and “Bolsonaro”, but not “Luiz” or “Flávio”. The chatbot refused some genuine candidate-related requests, yet agreed to draft misleading claims about election integrity. It also allegedly produced different outputs involving Bolsonaro’s number and Flávio’s name.

Outsiders cannot easily inspect the training data, internal rules, model updates or decision process. David Nemer said Meta’s systems were opaque, making deliberate bias impossible to determine. He added that repeated outcomes are difficult to dismiss as mere technical failure. The article therefore calls for clearer safeguards and defined goals.