Brazil Runoff Election Day: Meta AI Refuses to Generate Lula Images, Court Doubles Fine to 2 Million Reais per Day

On October 25, 2026, Brazil’s presidential runoff election day, Meta is paying a court fine equivalent to about $200,000 per day—not for spreading disinformation, but because its own AI system gave asymmetric answers when generating campaign images of the two candidates.

The starting point of the incident is clear and reproducible: when WhatsApp’s built-in Meta AI assistant received an image-generation prompt containing the name of Luiz Inácio Lula da Silva and his ballot number “13,” it refused to execute it; while the same type of prompt with the name of opponent Flávio Bolsonaro plus ballot number “22” generated campaign images normally. This discrepancy was publicly tested and documented by Brazilian users, and was also confirmed by a “confession” from Meta’s own system—according to Tech Policy Press, Meta AI’s automated response acknowledged: “An automated safety system blocked this request... This violates the core principles of Brazil’s election law.”

Meta’s Technical Explanation Does Not Hold Up

Meta’s official explanation is that this was a technical flaw triggered by a “generic noun plus ballot number” combination. Specifically, the system briefly blocked image requests containing Lula’s name and ballot number, while Flávio’s combination “worked for one more day” before being similarly blocked. The company claims it reported to the court on October 20 that remediation had been completed, and temporarily blocked both candidates’ names and ballot numbers together. Meta denies any political bias and characterizes the problem as a technical malfunction affecting “all sides of the political spectrum.”

The problem is that this explanation conflicts with internal evidence already disclosed. According to the Brazilian newspaper Folha de S. Paulo, internal communications among Meta employees showed that the company’s algorithm was “more favorable to Bolsonaro,” and that Lula’s Instagram account was “restricted and negatively affected” during the election period. Specific manifestations of the restrictions included disappearing share buttons, reduced content reach, and missing reaction and repost functions. Meta responded that the internal discussion “concerned only the severity of the problem, not political intervention,” but this statement itself acknowledges that such discussion did occur.

The content-filtering rules of AI systems are usually not publicly disclosed, and outsiders cannot directly verify whether the “flaw” was truly accidental. This is precisely the key to the case: when a black-box system produces asymmetric outputs for two candidates at a critical moment, “technical malfunction” and “systemic bias” are almost impossible to distinguish under evidentiary standards.

Court Escalates Penalties, Doubles Fine

Judge Marco Aurélio Bellizze of Brazil’s Superior Electoral Court (TSE) issued a preliminary ruling on October 15 requiring Meta to treat both candidates equally, with a daily fine of 500,000 reais (about $98,700) for noncompliance. On October 21, the court found that Meta had not fully implemented the preliminary ruling and doubled the daily fine—according to Novara Media, the actual penalty had reached about $200,000 per day.

The court’s specific requirements covered three areas: equal treatment of the two candidates’ content distribution; application of the same review standards to AI image-generation requests; and public disclosure of content moderation parameters involving political and election content. The last requirement essentially places Meta’s AI content-filtering rules under a legal oversight framework.

In their complaint, Lula campaign lawyers Gean Ferreira and Evelyn Santos characterized the issue as “the platform illegally exercising algorithmic selectivity during a heated presidential campaign,” and asked the court to order Meta to compensate the reach of restricted content to twice the normal level.

Tight Race, the Weight of Platform Power

This case is more than a technical dispute because the race itself is extremely close. First-round results showed Flávio Bolsonaro winning 47% of the vote and incumbent President Lula 45%, a gap within the margin of error. According to Bloomberg, in one of Meta’s largest revenue markets, reaching the right audience “could be the key factor determining the election outcome.”

This structural context means that even if Meta’s technical explanation is entirely valid, the asymmetric impact of the “brief flaw” already genuinely occurred in the most sensitive time window. Content-reach restrictions and AI image-generation refusals targeted the same candidate, in highly overlapping time periods, and in the same direction—regardless of whether the cause was technical error or something else, the political effect is the same.

This Is Not an Isolated Incident

During Brazil’s 2022 election, X, owned by Elon Musk (then still Twitter), was briefly banned in Brazil after refusing to comply with a court order to remove content. Four years later, Meta staged a structurally similar drama on the same stage: the platform refused cooperation on technical grounds, regulators applied legal pressure, and the two sides eventually entered a fine tug-of-war.

The difference is that this time the core of the dispute is generative AI, not manually reviewed content policy. The content-filtering rules of generative AI systems are jointly shaped by model training data, safety fine-tuning strategies, and keyword-trigger logic, and the interaction of these mechanisms is often difficult even for the development team itself to predict precisely. Meta’s acknowledgment that “the system acknowledged the violation” precisely shows that even its AI’s own response mechanisms exceeded the company’s expected range of control.

The case also touches on a more general engineering problem: AI content moderation systems are usually optimized for the platform’s most common “problem content” (violence, pornography, disinformation), while handling logic for specific combinations such as “candidate names + ballot numbers in electoral politics” is often patched after the fact rather than designed in advance. When platform user scale reaches Meta’s level, every combination that has not been systematically tested can become a minefield in real political scenarios.

Independent Assessment

Meta’s “technical flaw” claim is not logically impossible—AI systems giving inconsistent outputs for similar inputs is a common defect. But between “impossible” and “unbelievable” there is an important evidentiary gap: internal employee communications, the one-directional content-restriction pattern, and the fact that remediation was deemed “insufficient” by the court together constitute a doubt that the company cannot eliminate through statements alone.

Even if Meta ultimately proves that its system had no subjective bias, this case already shows that an AI content-filtering mechanism unable to prove its own fairness is itself a risk in high-stakes political scenarios. Brazil’s 2026 election will become a reference point—after AI has become deeply embedded in information dissemination, how regulators require platforms to bear legal responsibility for algorithmic asymmetry.