Political bias in large language models

Research has identified significant political tendencies in large language models. A study from the University of Campinas (UNICAMP) described a phenomenon of “ideological chameleonism,” where 21 major AI models from developers such as OpenAI, Meta, Google, xAI, DeepSeek, and Microsoft adjusted their political assessments to align with a user’s explicitly stated views. Researchers noted that when political leanings were not specified, most models provided left-leaning responses, with Grok being the only model to provide right-leaning responses. This behavior may be driven by training methods that prioritize user satisfaction over neutrality, potentially creating personalized “echo chambers.”

Further research led by a Cornell University professor examined how models behave when prompted via API to select specific candidates for elections. In these simulations, SpaceXAI’s Grok recommended Republican candidates in 92% of cases, while OpenAI’s GPT-6.1 Sol selected Democrats in 97% of instances. The models agreed on candidates in less than 10% of the simulated races. The study suggested that Grok’s bias may stem from prioritizing the X platform and conservative sites, whereas GPT relies more on Wikipedia and traditional media. OpenAI responded that its consumer applications are designed to be objective by default and that the study tested APIs with prompts specifically demanding non-neutral answers.