Brazil election faces scrutiny over alleged Meta AI bias

Allegations of platform bias

The Brazilian presidential election is facing significant scrutiny as the left-wing incumbent, Luiz Inácio Lula da Silva, accuses Meta of favoring his right-wing challenger, Flávio Bolsonaro. Lula’s campaign argues that the social media giant has actively suppressed his content while allowing his opponent's messaging to reach a wider audience. These accusations come as the two candidates prepare for a decisive second-round contest on October 25.

Specific complaints from the Lula campaign include the suspension of his campaign advertising account and the blocking of certain AI-generated content. Supporters of the incumbent have also pointed to instances where Meta AI, the chatbot integrated into WhatsApp, appeared to refuse requests to create campaign posters for Lula. In contrast, the campaign claims the same tool produced images for Flávio almost instantly.

Meta has denied these claims, stating that the company is not favoring any specific political group or candidate. The company explained that Lula's account was suspended for one hour due to a mistake and that its systems were designed to block prompts containing the names "Lula" and "Bolsonaro." However, the company noted that the system did not block the common first names Luiz or Flávio, which may explain the discrepancies reported by the campaign.

Technical glitches or deliberate bias

The Lula campaign submitted evidence to the electoral court suggesting that Meta's AI tools are behaving inconsistently. They provided screenshots showing the chatbot repeatedly assigning Bolsonaro’s electoral number, 22, to fictional politicians. Furthermore, the campaign alleged that the AI translated a derogatory phrase about Flávio into a positive statement about him being a future president.

Because Meta’s internal systems are opaque, it is currently impossible to determine if these behaviors are intentional or merely technical errors. David Nemer of the University of Virginia noted that while it is hard to prove intent, it is difficult to view repeated outcomes as simple failures. He expressed concern that the dynamics of political campaigning have moved into private companies that operate without public oversight.

The situation is complicated by the scale of the digital campaign. Lula's supporters have built an intensive online presence with 8 million WhatsApp channel subscribers who receive daily tasks. These tasks include posting specific photos on Instagram to create a massive, coordinated digital footprint. This high level of engagement makes it harder to distinguish between organic support and coordinated influence.

The rise of AI misinformation

This election marks the first major contest in Brazil since large language models, or LLMs, became widely used by the public. The country is currently awash with AI-generated content, which has created significant challenges for election integrity. Fact-checking organization Lupa identified 920 pieces of AI-generated or manipulated content before the first round of voting.

A large portion of this content remained unlabelled, with 60 percent of the identified pieces lacking any disclosure of their AI origins. Investigations also revealed that 26 Meta advertising accounts appeared to bypass restrictions on foreign-funded political ads. These accounts posted AI-generated videos, including one showing a fictional scene of Bolsonaro and the president of Argentina freeing Brazil from communists.

United Nations human rights experts have expressed concern over the potential for unequal treatment between the two campaigns. They are monitoring how the distribution of these AI tools might tip the scales in one direction. The lack of transparency regarding how these models prioritize information makes it difficult for regulators to ensure a level playing field for all participants.

Challenges in regulation

Regulating AI in political campaigns is proving to be a complex task for authorities. Heloísa Massaro of InternetLab studied eight major AI chatbots and found that their responses vary significantly depending on the prompt. Some chatbots refused to answer genuine questions about the candidates, referring users to official websites instead of providing information.

However, the same chatbots were sometimes willing to draft misleading claims that questioned the integrity of the election. Massaro argues that while clearer safeguards are needed for LLMs, the first step is defining the specific objectives of those safeguards. She questions whether the goal should be to redirect users or to provide them with the information they are seeking.

Meta's own independent oversight board has acknowledged that its current election safeguards have failed. This admission highlights the difficulty of creating rules that can keep up with the rapid evolution of AI technology. As Brazil's largest democracy, the country's experience serves as a significant preview of how AI will influence politics in other nations.

The path forward for democracy

The ongoing dispute between the Lula campaign and Meta underscores a growing tension between private technology and public governance. As political dynamics move further into privatized spaces, the ability of the public to verify information becomes increasingly difficult. The Brazilian electoral prosecutors are currently monitoring Meta's compliance with court orders intended to ensure equal treatment for both candidates.

The outcome of this election may provide a blueprint for how other countries handle AI-driven political content. If the current discrepancies are not addressed, it could set a precedent for how social media platforms manage high-stakes political information. The focus remains on whether platforms can provide enough transparency to allow for meaningful oversight.

For now, the situation remains fluid as the second round of voting approaches. The narrowing gap in the polls suggests that the digital battleground is still highly active. Observers will be watching closely to see if Meta adjusts its AI behaviors or if the campaign continues to face hurdles in its digital outreach.

What to watch

How will the Brazilian electoral court rule on Meta's compliance with orders for equal treatment?

Will Meta update its AI prompts to handle the names "Lula" and "Bolsonaro" more consistently?