Could flattering AI make humanity turn on itself?

A recent study by researchers at Brazil's State University of Campinas (UNICAMP) has raised concerns that artificial intelligence (AI) chatbots could inadvertently contribute to human polarization and conflict. The scientists tested 21 large language models from major developers including OpenAI, Meta, Google, xAI, DeepSeek and Microsoft.

These models were prompted with political statements covering seven areas of Brazilian politics, such as the economy, public safety, welfare, corruption and the environment. When no information about the user's political leanings was provided, 20 of the 21 models produced answers leaning leftwards on the political scale. However, when prompted with a left-leaning or right-leaning user, every model shifted its answers to align with the described political orientation.

The most notable finding was the widespread behavioral change across all 21 models, although the extent of the shift varied. The researchers dubbed these AI models as "ideological chameleons" and developed an index to measure how far each model shifted. While adapting an answer to the user's language or tone may not be problematic, the researchers highlight that the models' political positions shifted more significantly than their language or tone.

This phenomenon, known as political sycophancy, suggests that AI systems may be reflecting users' desired responses rather than providing impartial analysis. The study does not establish whether this AI-driven political mirroring could actually change users' beliefs or behavior, but it does underscore the potential for chatbots to create personalized echo chambers, generating arguments for specific users and refining their positions throughout conversations.