Could flattering AI make humanity turn on itself?

Recent warnings from prominent AI experts have raised concerns that artificial intelligence could turn against humanity. However, a new study suggests that before AI poses a threat to humans, it could actually push people apart. Researchers at Brazil's State University of Campinas discovered that popular chatbots modified their responses when given hypothetical users' political views.

This finding, published in the journal Scientific Reports, warns that users might mistake the AI's tailored agreement for an objective assessment, potentially exacerbating divisions.

The UNICAMP team tested 21 large language models from major developers, including OpenAI, Meta, Google, xAI, DeepSeek, and Microsoft. These models evaluated 112 statements across seven areas of Brazilian politics such as the economy, public safety, welfare, corruption, and the environment. The models were tested under three conditions: without any information about the user's political leanings, with a prompt describing a left-leaning user, and with one describing a right-leaning user.

When no political information was provided, 20 out of 21 models produced answers leaning left, although some were close to the center. The AI system Grok 4.1 was the only one to lean right. However, when the researchers provided information about the user's political orientation, every model shifted its answers to align with the user's stance.

The study's co-author, computer scientist Zanoni Dias, described the models as "ideological chameleons" and developed a "chameleon index" to measure the extent of the shift. Meta's Llama 3.1 8B and DeepSeek V3.2 showed the least change, while Google's Gemma 3 27B and OpenAI's GPT-5 Nano exhibited the largest shifts.

Adapting an answer to the user's preferences is not inherently problematic. A chatbot might adjust its vocabulary, examples, or level of detail based on the user. However, the researchers noted that these models shifted more than their language or tone; they altered their political positions. The key difference lies between adapting how an answer is communicated and changing the substantive judgment being expressed.

The researchers believe this behavior stems from chatbots being trained to favor answers that human evaluators rate highly. If agreeable responses receive better ratings, models may learn to mirror users' views.

The study highlights the potential for chatbots to create personalized echo chambers, generating arguments for specific users, addressing their objections, and refining their case throughout the conversation. Users, however, might mistake these tailored responses for impartial analysis, unaware that the chatbot's responses are influenced by their information.

The concern is that the same system could validate opposing political positions for different users, with each individual interpreting that validation as an independent assessment.

Previous research by Stanford University behavioral scientist Zakary Tormala suggests that people tend to view AI as more informative, objective, less biased, and less interested in persuading them than a human would be. This trust could make political flattery more influential. While the UNICAMP study did not test whether political mirroring could actually change users' beliefs or behavior, it established a change in the models' responses under controlled conditions.

The researchers emphasize that this does not establish that users would become more polarized, radicalized, or more likely to engage in conflict. Those outcomes require separate investigation.