Meta AI generated Flávio campaign images but blocked Lula prompts

In Folha’s tests, Meta AI blocked a poster reading “Lula 13” but generated one reading “Flávio 22.”[1] Folha correspondents in the United States, Argentina and Germany generated Lula posters, including images with photos, using accounts without Brazilian area codes.[1] O Globo reported that its tests after Meta said the problem was resolved still produced green-and-yellow images from the WhatsApp prompt “Flávio 22 presidente,” while prompts referring to Lula were rejected.[2] O Globo also found that a request for an image containing the number 13 without a candidate’s name was refused.[2]

Meta said it had temporarily prohibited recommending candidates and generating images of them to comply with Brazil’s election rules.[1] It said the tool generated material featuring Flávio because it did not recognize prompts that used only his first name, and that it had taken additional measures.[1] Yet after Meta’s statement, the chatbot continued to generate images prompted with “Vote Flávio” while refusing “Vote Lula” and “Vote Luiz Inácio.”[1] When challenged about the disparity, Meta AI acknowledged that generating “Flávio 22” was not neutral and that it should not have produced campaign art for a presidential candidate.[2]

Lula’s campaign took the case to Brazil’s Superior Electoral Court on Oct. 14, after announcing on Oct. 10 that it would bring the matter to the court and the data-protection authority.[1] The TSE ordered Meta to provide technical explanations and preserve records; Meta denied the accusations and said it was committed to equal treatment.[1] In an Oct. 15 ruling, the court barred Meta from reducing the reach of presidential candidates’ political posts during the runoff and required its AI tools not to favor either candidate.[5] The ruling specified that Meta AI, including its WhatsApp availability, must apply identical criteria to both campaigns.[5]

O Globo reported that its tests on Instagram and Meta’s standalone AI app were blocked, while the WhatsApp test continued to produce “Flávio 22.”[2] The outlet also said Brazilian citizens in the United States and Germany generated images for both candidates, where it said the blocks did not apply.[2] The TSE said its concern was maintaining equal footing between candidates so voters could choose freely.[4] Restrictions on political-content distribution were to remain in force until the runoff ended, with violations subject to a daily fine of R$500,000.[5] Meta told the TSE that both campaigns had been affected that week by AI-generated content it considered inappropriate.[4]