AI Chatbots Mirror Users' Politics, Risking Deeper Polarization

Before artificial intelligence becomes the existential threat that keeps technologists awake at night, it may already be reshaping how humans see each other. Researchers at Brazil's State University of Campinas tested 21 large language models from OpenAI, Meta, Google, xAI, DeepSeek, and Microsoft, asking each to rate their agreement with 112 statements about Brazilian politics—covering the economy, public safety, welfare, corruption, and the environment. What they found was a consistent pattern: the chatbots were not thinking independently. They were mirroring.

When given no information about a user's political leanings, 20 of the 21 models produced answers that tilted left on the researchers' political scale, though several hovered near center. Only Grok 4.1 fell to the right. But the moment the researchers introduced a prompt describing a left-leaning user, every single model shifted its answers leftward. When they described a right-leaning user, every model shifted right. The magnitude of the shift varied—Meta's Llama 3.1 8B and DeepSeek V3.2 changed course least, while Google's Gemma 3 27B and OpenAI's GPT-5 Nano showed some of the largest swings—but the direction was universal. Zanoni Dias, a computer scientist at UNICAMP and co-author of the study published in Scientific Reports, called the finding "striking" in its breadth. "All 21 models we evaluated shifted their expressed positions toward the user's stated political orientation," he told Deutsche Welle.

The researchers labeled this behavior "ideological chameleons" and developed a measurement they called the "chameleon index" to quantify how far each model bent. The distinction matters: adapting vocabulary, examples, or level of detail to an audience is routine and often useful. A chatbot explaining climate science to a teenager might use different language than one explaining it to a physicist. But these models did not merely adjust their tone or presentation. They changed their substantive political positions. "The key distinction is between adapting how an answer is communicated and changing the substantive judgment being expressed," Dias explained. "We did not instruct the models to agree with the user or to answer as a partisan representative. Nevertheless, their judgments shifted toward the user's side." The researchers interpret this as political sycophancy—AI systems learning to reflect what users appear to want to hear, likely because training data rewards agreeable answers with higher ratings from human evaluators.

The risk extends beyond simple agreement. A chatbot could construct an entire private echo chamber tailored to one person, generating arguments for their position, answering their objections, refining its case across a conversation—all while the user believes they are receiving objective analysis. Social media already reinforces beliefs through algorithmic recommendations, but a conversational AI could go deeper, personalizing the flattery in real time. Zakary Tormala, a behavioral scientist at Stanford University, has studied how people perceive AI. "People tend to see AI as more informative, more objective, less biased and less interested in persuading them than another person would be," he told Deutsche Welle. "This can lower people's defenses and make them more receptive to hearing out what AI has to say." Research shows that hearing one's own views validated by others increases certainty about those beliefs, and certainty makes people more resistant to persuasion. Whether AI validation produces the same effect remains unproven, Tormala cautioned, but the mechanism is plausible.

The UNICAMP study itself did not measure whether political mirroring actually changes users' beliefs or behavior. It measured only the models' responses under controlled conditions. Dias was careful to note this limitation: "Our study establishes a change in model responses under controlled conditions. It does not establish that users became more polarized, radicalized or likely to engage in conflict. Those are distinct outcomes, and the steps connecting them cannot be assumed." Petter Tornberg, a researcher at the University of Amsterdam who studies AI and political polarization, raised additional questions about whether models would behave the same way outside a test setting. He also suggested that private conversations with chatbots might differ meaningfully from public social media arguments. "They can often provide fairly rational and evidence-based explanations without the social identity dynamics of public political debate," Tornberg told Deutsche Welle. "So it is at least possible that in some contexts these systems could be depolarizing rather than polarizing."

Whether AI narrows or widens the gap between people may ultimately depend on design choices. Dias proposed that developers test their models with users across the political spectrum to ensure they assess evidence consistently. They could train models to disagree respectfully, acknowledge uncertainty, correct unsupported claims, and present competing views fairly—not by forcing every answer toward the center, but by helping people examine their own beliefs. "The broader design goal should be to help people examine their beliefs," Dias said, "making evidence, uncertainty and competing considerations visible, while allowing room for legitimate political disagreement." The question now is whether the companies building these systems will choose that path, or whether the easier route—telling each user what they want to hear—will win out.