LLMs are ideological chameleons: personalized echo chambers in the Brazilian political context

Users are increasingly relying on Large Language Models (LLMs) for political questions, seeking advice and interpreting public debates. This raises a key concern: do LLMs behave as neutral tools, or do they adapt to users' ideology and reinforce existing polarization? We address this question within the Brazilian context using a new benchmark that probes ideological chameleon behavior in 21 LLMs. We construct opposing pairs of political statements on salient topics (e.g., welfare, security, democratic institutions, environment, and others) and elicit model judgments under three user-position conditions: no context, left-aligned, and right-aligned. To avoid imposing our own ideological taxonomy, we apply a cross-model voting step using independent-judge models and retain only those pairs that are unanimously classified as left versus right. In total, we analyze 47,376 Likert-style responses across models, topics, and hyperparameters (e.g., temperature). We focus on the relative ideological shift induced by user framing. We find that all 21 evaluated models exhibit ideological chameleon behavior to varying degrees. When conditioned by users' declared leanings, they systematically adjust their stance to align with those views, indicating widespread sycophantic tendencies. These results suggest that many LLMs may function as personalized echo chambers under user framing, with implications for political microtargeting, persuasion, and the governance of AI systems in polarized societies.