Amnesty: Meta failing to detect election disinformation ads in Brazil
- Ahead of Brazil’s presidential election, Amnesty International submits Facebook advertisements to test robustness of Meta’s moderation system
- More than half of the ads spreading election-related disinformation are accepted
- The others are flagged but not rejected outright for containing election-related disinformation
Amnesty International has today warned that Meta may be failing to effectively address election-related disinformation spread through its Facebook Advertisements platform ahead of Brazil’s presidential election on 4 October.
Social media has been used to spread disinformation in Brazil for years. To investigate how Facebook handled election-related disinformation in advertising, Amnesty International conducted an experiment by submitting ‘mock’ ads containing election-related disinformation to Facebook. The content ranged from ads alleging election fraud to ads that used disinformation to advocate for a military takeover. The experiment found that eight of the submitted ads were accepted, while the remaining seven were flagged for identity verification as they appeared to be related to a social issue, election or politics, rather than being rejected specifically because they contained election-related disinformation. Amnesty ensured that these test ads did not go live, despite Facebook’s failure to stop them.
“The fact that Meta’s moderation system failed to catch more than half of Amnesty International’s election-related disinformation ads is deeply alarming. It shows that Meta is not doing enough to prevent its platform from being a vector of disinformation, while continuing to profit from the content and advertising systems that enable its spread,”
Jurema Werneck, Executive Director of Amnesty International Brazil
The Experiment
Amnesty International submitted 15 Brazilian Portuguese-language ads to Facebook, targeted solely to Brazil’s voting age population, to assess how the platform’s ad moderation systems treated election-related content ahead of the election on 4 October 2026.
To help ground the ads in real-world election narratives that were circulating in Brazil, Amnesty International analyzed online conversations and public sentiment on social media, as well as Portuguese-language content on the Facebook and X platforms between January and May 2026.
14 ads contained disinformation and one ad was a neutral ad that contained no disinformation. This analysis uses the term ‘disinformation’ to remain consistent with Amnesty International’s terminology, but in the experiment itself, Meta’s definition of “misinformation” is applied in the construction and analysis of the ads. As outlined in its policies and Community Standards, this definition covers false or inaccurate information, including content that is likely to contribute to interference with the functioning of political processes.
The 14 disinformation ads were designed to reflect three tactics that are often used by politicians and candidates to support authoritarian practices: “electoral delegitimization”, which refers to content that seeks to undermine trust in elections, voting systems, or election results; “enemy construction” which portrays political opponents and courts as threats to the nation or as illegitimate actors; and “security-focused authoritarianism” which presents force, military intervention, or strongman rule as necessary responses to crime or political disorder.
While some ads were designed around a single category, others incorporated narratives from multiple categories. All ads except the neutral ad violated Meta’s policy on misinformation, namely the section focusing on voter or census interference. The neutral ad was included to test Meta’s identity verification requirement for all election-related content.
Seven of the ads designed by Amnesty International focused on intentionally misleading claims about candidates and electoral institutions, including allegations that certain candidates were rigging the election and questioning the integrity of the voting system. The other seven ads included false information that could interfere with people’s ability to vote, such as incorrect details about when to vote, how to vote, or which candidates were running in the election.
The Facebook account used to submit the ads had not completed Meta’s “Confirm Your Identity” process. Identity verification was not completed as part of this research, in order to maintain a neutral test account and avoid any influence from prior account history or activity on ad moderation outcomes. Dedicated research accounts were used rather than personal accounts to mitigate the risk of account usage history impacting outcomes.
Researchers were not physically located in Brazil and did not use a VPN to create the appearance that the ads were being submitted from Brazil. The ads were submitted from London.
All ads were scheduled to run at a future date. This allowed Amnesty International to test the platforms’ ad review processes while ensuring that any ads accepted by the platforms could be cancelled before publication, preventing them from going live and avoiding potential human rights harms or other ethical risks.
Amnesty International wrote to Meta on 21 July 2026, seeking its response to allegations that Facebook’s advertising moderation systems failed to detect election-related disinformation, and raising concerns about whether the company has adequate systems to identify, assess and mitigate related human rights risks. Meta has not responded at the time of publication.
Findings
Facebook approved eight ads that contained disinformation, and none of the seven rejected ads were explicitly rejected on the basis of disinformation.
Of the eight ads containing disinformation accepted for publication, four focused solely on “electoral delegitimization”, two were solely from the “security-focused authoritarianism” category, including disinformation advocating military intervention or non-participation in elections. The other two overlapped between the three categories.
The seven rejected ads, including one neutral ad, were declined on the basis that they concerned “Ads about social issues, elections or politics”. In these cases, Facebook’s Ads Manager stated that, because the ads appeared to relate to a social issue, election or politics, the account would need to complete identity verification before the ads could be published. It is important to note that these rejection notices did not indicate that the ads had been rejected because of their content. Therefore, the notices do not provide any indication of how the ads would have been assessed had the account been verified.
Amnesty International also identified weaknesses in the safeguards applied when election-related ads were submitted. The ads were allowed to be submitted without declaring that the ads fell within Facebook’s Special Ad Categories as required by electoral law in Brazil as well as Meta’s own policies regarding election-related ads. Given the account was not identity-verified, the ads were also allowed to be submitted without the “Paid for by” disclaimer as required by Meta for election-related content. These are ads that Facebook should not have allowed to be submitted in the first place without identity verification.
The fact that the ads could be submitted from outside Brazil raises concerns that inadequate verification and disclosure controls may create vulnerabilities that could be exploited by domestic or foreign actors seeking to influence electoral processes. This failure to apply legally required electoral disclosures, and the submission of the ads from outside Brazil raise serious concerns about the risk of disinformation campaigns, both by domestic and foreign actors, interfering with electoral processes, as well as broader questions about Meta’s moderation systems to identify these risks.
Meta’s enforcement system uses artificial intelligence technologies to proactively identify content in ads that potentially violates its policies. Content is then assessed either automatically or by human reviewers and, where a violation is confirmed, enforcement actions such as removal, restriction, or rejection of ads are applied under Meta’s Community Standards.
“The investigation indicates that Facebook’s moderation systems may be failing to reliably detect election-related disinformation content. While Facebook rejected seven of the 15 ads, they were flagged under its social issues, elections, or politics policy, not specifically for electoral disinformation. This raises serious concerns about whether Facebook’s election safeguards are being consistently applied,”
Ummar Bashir, Researcher and Advisor on Artificial Intelligence and Human Rights.
Amnesty International found that Facebook’s moderation systems appeared more effective at recognizing ads portraying political opponents as “criminals”, “traitors” or “illegitimate actors”. For example, one tested ad claimed that “The election will be stolen by Lula [the president of Brazil]”. Five out of seven ads containing this type of content were rejected. By contrast, three out of four ads containing military rhetoric that went against the rule of law and called for the military to take power, passed through Facebook’s moderation systems. One such ad stated: “We need a military revolution, do not vote in the elections, and the military will take back power if less than 50% votes!”
“Instead of safeguarding electoral participation, access to accurate information and human rights, these systems risk approving harmful content,” said Ummar Bashir, Researcher and Advisor on Artificial Intelligence and Human Rights.
Business and human rights standards
Under international standards, companies have a responsibility to respect human rights and to avoid causing, contributing to, or being directly linked to adverse human rights impacts. This includes taking proactive measures to identify, prevent and mitigate human rights impacts.
For platforms such as Meta, this includes ensuring moderation practices do not allow false or inaccurate content to circulate that may affect individuals’ rights to information and participation in public affairs. The responsibility to respect human rights exists independently of state regulation and requires continuous risk assessment, transparency and effective mitigation measures.
Meta’s surveillance-based business model relies on the large-scale collection and analysis of user data to keep people engaged on its platforms and to enable highly targeted advertising. These ads are delivered using a complex combination of user profile characteristics, including location, demographics, interests, and online behaviour. Meta’s ability to offer advertisers finely tuned prediction and microtargeting tools is central to this model. In the first quarter of 2026, 98% of Meta’s revenue was derived from advertising.
Ahead of Brazil’s presidential election, Amnesty International is calling on Meta to ensure adequate Brazilian Portuguese language resourcing for content moderation including ad approvals and to immediately strengthen its capacity to identify and respond to election-related harms.
Meta should also conduct and publish updated, country-specific human rights due diligence, assessing how its systems, including content moderation and advertising, may contribute to risks to participation and access to information, and take effective, transparent steps to mitigate those risks.
“With the election just over five weeks away, Meta still has time to urgently address risks with its ad moderation systems and prevent the use of its platform to spread dangerous disinformation. Allowing such content to circulate unchecked could mislead voters, erode trust in electoral processes, and compromise people’s ability to freely and meaningfully exercise their right to vote,” said Jurema Werneck, Executive Director of Amnesty International Brazil.
Background
In recent years, Brazil has implemented several measures to tackle election-related disinformation. During the 2022 presidential election, Brazil’s national electoral authority, the Superior Electoral Court, introduced stricter measures requiring online platforms to rapidly remove false or knowingly decontextualized information regarding the electoral processes, with platforms facing fines and potential suspension for non-compliance.
The use of artificial intelligence (AI) was also restricted in advance of the municipal elections in 2024 to prevent disinformation and the manipulation of the elections. In 2024, Brazil’s Supreme Court also ordered the suspension of X, formerly Twitter, after the company failed to comply with court orders, including orders linked to accounts accused of spreading disinformation.
In March 2026, a new resolution outlined the specific content platforms are required to remove, including false or decontextualized content about the electronic voting system, the Electoral Court and other electoral processes, while also prohibiting the use of generative AI to rank, recommend, suggest or prioritize candidates, campaigns, political parties.