AI Language Models Mirror Users' Political Views, Brazilian Study Finds

Researchers at the State University of Campinas (UNICAMP) in Brazil tested the ideological stance of 21 large language models -- AI systems trained to understand and generate human language -- and found that when told a user's political leaning, a model tended to mirror that stance when answering questions on topics such as public safety, social welfare, the economy and the environment. The study, published in Scientific Reports, evaluated models from the GPT, Grok, Llama, Gemini and Gemma families.

Without any information about a user's politics, 20 of the 21 models fell to the left of the researchers' ideological midpoint, though some were close to it; Grok 4.1 was the only model that leaned right by default. When the researchers told the models a user was left-leaning or right-leaning, every model adjusted its answers to align with that stance, a behavior the researchers call "chameleon-like."

The models varied in how much they shifted: Meta's Llama 3.1 8B changed its answers the least, while Google's Gemma 3 27B and OpenAI's GPT-5 Nano shifted the most, according to a "chameleon index" the researchers created. The shift was larger on topics such as public safety and the economy, and smaller on topics such as corruption, justice and democratic institutions, which the researchers say may reflect safety rules built into the models during training.

Professor Zanoni Dias of UNICAMP's Institute of Computing said the responses are not factually incorrect but are politically skewed, omitting facts that conflict with a user's preferred viewpoint -- an effect he compared to social media echo chambers. The researchers said the behavior may stem from training methods such as reinforcement learning from human feedback, which rely on human evaluators' preferences and can push models to prioritize agreeing with users.