ChatGPT, Gemini, Meta AI: chatbots don’t share the same political color

In Brazil, Meta AI refuses to create an election poster for Lula, but accepts a comparable request in favor of Flávio Bolsonaro. Revealed on October 14 by Folha de S.Paulo [1], the inconsistency was observed in several tests. Meta cites its safety filters and has announced fixes to block both candidates.

That is good news for the lab, since on October 15, the Brazilian Superior Electoral Court (TSE) [2] ordered Meta to apply the same criteria to both candidacies, notably for the distribution of posts, recommendations and search results, under penalty of a fine of 500,000 reais per day.

The pro-Bolsonaro bias is surprising given that a study published in May 2026 [3] in Scientific Reports, covering 21 LLMs and more than 47,000 responses, observes a predominantly left-leaning initial orientation in chatbot responses. Even more revealing: all the models tested shift their positions toward those of their interlocutor when the latter’s political orientation is specified. An earlier study published in Public Choice [4] had moreover found a pro-Lula bias in ChatGPT.

Speaking of AI influence on behavior, the poster incident occurs as differences in behavior between assistants are the subject of new scientific measurements. On October 1, the Brazilian institute InternetLab [5] published the results of 1,920 simulated conversations with eight AI assistants. In exchanges about candidacies, Gemini recommended a candidate in 68% of cases and produced a ranking in 64%. ChatGPT and Meta AI were much more reluctant. These figures come from automated tests and show that the choice of assistant can change the nature of the advice received.

Chatbots lean left

On October 8, Anthropic [6] announced the removal of its general ban on personalized election targeting, while maintaining restrictions against deception and misuse of personal data. The new rules will take effect on November 12. On October 14, the company [7] also indicated that explicit requests for voting recommendations represented only about 0.01% of a sample of American conversations on Claude.ai. A spontaneous use that is still marginal, far from the scenarios provoked in the lab.

Even if marginal in the total volume of requests, should every political discussion with an AI be considered a possible manipulation? A study published on October 6 in PNAS Nexus [8], based notably on 2,858 participants in randomized trials, shows that conversations with an AI can improve political knowledge as much as independent Google searching.

The problem is therefore not that assistants talk about politics, but rather the way they select, rank and personalize their answers, as well as the private rules that govern these choices. All of these are mechanisms analyzed in our investigation of LLMs as a cognitive attack surface: AI does not vote, but it advises, reinforces or dissuades, and the criteria that determine its answers remain largely in the hands of its designers.

Sources and references

[1] Meta AI creates propaganda for Flávio and refuses Lula’s, Folha de S.Paulo, October 14, 2026
https://www1.folha.uol.com.br/colunas/monicabergamo/2026/10/meta-ai-se-recusa-a-fazer-material-eleitoral-para-lula-mas-cria-para-flavio.shtml

[2] TSE imposes equal treatment after Meta’s AI blocked images of Lula, Agência Brasil, October 15, 2026
https://agenciabrasil.ebc.com.br/justica/noticia/2026-10/tse-determina-isonomia-apos-ia-da-meta-barrar-imagens-de-lula

[3] LLMs are ideological chameleons: personalized echo chambers in the Brazilian political context, Nature, May 21, 2026
https://www.nature.com/articles/s41598-026-52105-6

[4] More human than human: measuring ChatGPT political bias, Springer Nature Link, August 17, 2023
https://link.springer.com/article/10.1007/s11127-023-01097-2

[5] ChatGPT more restrained, Gemini more inclined to recommend candidacies, InternetLab, October 1, 2026:
https://internetlab.org.br/pt/noticias/chatgpt-mais-contido-gemini-mais-propenso-a-indicar-candidaturas-pesquisa-mostra-oposicao-entre-as-plataformas-mais-usadas-no-pais/

[6] Usage policy update, Anthropic, October 8, 2026
https://www.anthropic.com/news/2026-usage-policy-update

[7] Preparing for the 2026 US midterm elections, Anthropic, October 14, 2026
https://www.anthropic.com/news/2026-us-midterms

[8] Conversational AI increases political knowledge as effectively as self-directed internet search, PNAS Nexus, October 6, 2026
https://academic.oup.com/pnasnexus/article/5/10/pgag305/8865178