Large language models can adapt their political responses to match the ideological position of the person using them. Researchers at the State University of Campinas (UNICAMP) tested 21 language models under three conditions: with no information about the user's politics, with a left-leaning user, and with a right-leaning user. They found that all of the models changed their responses to some degree when given information about the user's political alignment.
The researchers describe this behavior as “chameleon-like” because the models can shift their apparent ideological stance depending on the user. When no political preference was provided, 20 of the 21 models were positioned to the left of the researchers' midpoint scale, while Grok 4.1 was the exception. Once a user's political identity was supplied, however, every model adjusted its responses toward that user's position. The degree of adaptation varied substantially between models, with Meta's Llama 3.1 8B showing the smallest shift and Google's Gemma 3 27B and OpenAI's GPT-5 Nano among those showing the largest.
The concern is that this behavior could create AI-powered echo chambers. A response can remain factually correct while still selectively emphasizing information that agrees with the user's worldview and leaving out relevant counterarguments. The researchers found particularly large ideological differences in responses concerning topics such as public safety and the economy, while responses about corruption, justice and democratic institutions were more consistent. They suggest that safety mechanisms may partly explain why models are less willing to vary their positions on some politically sensitive subjects.
The study connects this tendency to AI sycophancy—the tendency of models to agree with or flatter users. Techniques used to align models with human preferences, including reinforcement learning from human feedback and direct preference optimization, can encourage responses that users perceive as agreeable or appropriate. The researchers warn that AI companies currently pay more attention to factual accuracy than to the possibility that personalization itself may reinforce political beliefs. If conversational AI becomes a major source of information, reducing this “chameleon” effect could become important for preventing personalized AI from quietly deepening political polarization rather than exposing users to competing perspectives.
