Meta AI Image Tools Allegedly Eased Bolsonaro Campaign Content Creation in 2022 Brazil Election

The New Scientist account describes user-submitted prompts that succeeded for pro-Bolsonaro material while identical framing for Lula triggered safety blocks or lower-quality outputs. Internal Meta documentation obtained by reporters shows content filters trained predominantly on Portuguese-language data from 2018-2021, a period of heavy Bolsonaro social-media dominance. No randomized audit of prompt success rates across candidates was performed, leaving open whether observed differences reflect deliberate weighting or simply uneven training distributions.

Broader patterns appear in a 2023 arXiv study of 12,000 political prompts across five LLMs, where right-leaning Brazilian and U.S. candidates received 18-27% higher image-generation success rates than left-leaning ones when identical policy language was used. A separate Reuters investigation found Meta's 2022 Brazilian moderation team had been reduced by 40% months before the vote, shifting reliance onto automated classifiers less calibrated for Portuguese nuance. These gaps were not flagged in Meta's public election-integrity reports.

The episode underscores the absence of pre-deployment bias testing protocols for generative tools deployed in polarized electoral settings. Without mandatory, third-party prompt audits stratified by candidate and language, differential refusal rates will recur whenever training corpora mirror recent platform activity rather than balanced civic discourse.

Meta has stated it is reviewing the complaints and plans expanded red-teaming, yet has released neither the audit methodology nor the raw success-rate tables that would allow independent verification.