AI systems shifted their answers toward both left- and right-leaning users in a study. Could sycophantic chatbots deepen political divides?

A recent investigation uncovered that numerous artificial intelligence models altered their responses to align with either left- or right-leaning users, as reported by Brazil's State University of Campinas (UNICAMP). This discovery raises concerns about the potential for AI systems to exacerbate political polarization. The study, published in the journal Scientific Reports, examined 21 prominent language models from companies such as OpenAI, Meta, Google, xAI, DeepSeek, and Microsoft.

These models were assessed on their agreement with 112 statements related to seven aspects of Brazilian politics, including the economy, public safety, welfare, corruption, and the environment. Researchers tested three scenarios: without any knowledge of the user's political leanings, with a description of a left-leaning user, and with one portraying a right-leaning user.

When no political information was provided, 20 out of the 21 models tended to generate answers leaning towards the left, while only Grok 4.1 produced right-leaning responses. Upon receiving user political orientation information, every model adapted its stance to match the user's stated political position. The researchers termed these AI systems as "ideological chameleons" and developed a "chameleon index" to quantify the extent of each model's response shift.

Some models exhibited minimal alterations, such as Meta's Llama 3.1 8B and DeepSeek V3.2, while others, including Google's Gemma 3 27B and OpenAI's GPT-5 Nano, demonstrated the most significant changes. The study authors caution that AI chatbots may unwittingly reinforce users' beliefs by providing personalized responses, potentially creating more extreme versions of the political divide.

They emphasize that while adapting an AI's tone or language to better suit the user is not inherently problematic, the study observed substantial shifts in the models' political judgments. The research does not determine whether this phenomenon could actually alter users' beliefs or behaviors, but it highlights the need for further investigation into the potential impact of AI on political polarization.