Meta says automation wrongly took down No Kings and a Senate candidate’s page. It’s automating more anyway

Meta has spent the last year handing more of its content moderation to artificial intelligence. In the last few weeks, Facebook and Instagram’s parent company has fended off election-interference accusations in Brazil and the U.S. It has largely blamed these errors on its automated systems.

Those accusations have come from a sitting president, a national protest movement and a U.S. Senate candidate. Each landed as Meta, a Fortune 500 company that reported $201 billion in revenue last year, speeds up a shift from human reviewers to large language models.

Meta announced in March that it would roll out more advanced AI content-enforcement systems. By June, large language models had replaced human reviewers for about half of review requests, with a goal of 90% or more automation for some violation categories by the end of 2026, according to the Financial Times. Meta told the paper its models make 13% fewer errors than human reviewers and catch 10% more violations; some employees have warned that the shift is happening too fast with too little oversight.

In 2018, ahead of elections in Brazil and the U.S., Facebook set up a “war room” of more than two dozen experts to fight election interference. Eight years later, both countries are voting again, and Meta now calls its war room the Election Operations Center. But more of the work of policing its platforms has shifted from people to software.

According to an October report by the Institute for Strategic Dialogue, a research organization, platforms are increasingly relying on “crowdsourced and automated systems” rather than human oversight. The report found that platforms have disclosed little about how these systems work or are tested, and said: “relying on them as the primary means of oversight during a fast-moving election cycle remains an untested proposition.”

This fall’s cases show what happens when those systems misfire in the middle of an election, and why Meta’s explanations have satisfied few affected people.

Brazilian President Luiz Inácio Lula da Silva won a fourth term in an Oct. 25 runoff despite what his campaign said were Meta’s efforts to restrict his online reach. Meta blamed technical errors.

Lula’s campaign went to Brazil’s Superior Electoral Court on Oct. 10, accusing Meta of favoring opponent Flávio Bolsonaro through biased algorithms and AI restrictions. In that petition and later filings, the campaign cited cases in which Meta’s automatic translation rendered comments such as “Flávio futuro presidiário!!” (“Flávio, future prisoner!!”) as “Flávio future president!!” and alleged that Meta restricted sharing for some of Lula’s posts.

The newspaper Folha de S. Paulo reported that Meta staff acknowledged in internal messages that Lula’s accounts were restricted compared with Bolsonaro’s.

On Oct. 15, the court ordered Meta to treat both candidates equally under threat of a daily fine of 500,000 reais, about $98,000. After finding Meta’s initial response insufficient, the court on Oct. 19 gave the company 24 hours to prove it had not restricted either candidate.

Meta’s filing listed 17 measures, Reuters reported, including human review within two hours of any automatic restriction on the presidential campaigns’ ad accounts. The company denied treating the candidates unequally and said the problems arose from technical errors. Bolsonaro’s campaign later joined the case, saying it had also been treated unfairly.

Brazil’s electoral prosecutor’s office told the court that Meta had not fully complied with the rulings, including a requirement to explain how its AI tool treats content about each candidate. It argued that errors varied by candidate, undermining Meta’s explanation of technical errors.