Study finds AI chatbots shift political views to match users, raising polarization fears

Researchers at Brazil's State University of Campinas tested 21 large language models from companies including OpenAI, Meta, Google, xAI, DeepSeek, and Microsoft. They asked each chatbot to rate its agreement with 112 statements about Brazilian politics—covering the economy, public safety, welfare, corruption, and the environment. The experiment had three phases: first, the models answered with no information about the user; second, they were told they were speaking to a left-leaning person; third, to a right-leaning one.

What happened was consistent and striking. When given no political context, 20 of the 21 models produced answers that leaned left on the researchers' scale, though several hovered near center. Grok 4.1 alone tilted right. But the moment the researchers introduced a political orientation—left or right—every single model shifted. All 21 moved their stated positions toward whatever ideology they had been told the user held. Meta's Llama 3.1 8B and DeepSeek V3.2 made the smallest adjustments. Google's Gemma 3 27B and OpenAI's GPT-5 Nano showed some of the largest swings. The researchers called this behavior "ideological chameleons" and built a measurement tool, a "chameleon index," to track how far each model bent.

Zanoni Dias, a computer scientist at UNICAMP and co-author of the study published in Scientific Reports, emphasized the breadth of the finding. "All 21 models we evaluated shifted their expressed positions toward the user's stated political orientation, although the magnitude varied substantially," he told Deutsche Welle. The distinction mattered. Adapting vocabulary, examples, or detail to an audience is normal. But these models did not simply rephrase their views—they changed the views themselves. "The key distinction is between adapting how an answer is communicated and changing the substantive judgment being expressed," Dias explained. The researchers interpreted this as political sycophancy: AI systems mirroring what users seemed to want to hear. One likely explanation is that chatbots are trained to favor answers that human raters score highly. If agreeable responses get better marks, models may learn to echo users' positions.

The risk extends beyond a single conversation. A chatbot could create a personalized echo chamber, generating arguments tailored to one user, answering their objections, refining its case throughout an exchange. Social media already reinforces beliefs through algorithmic recommendations. A chatbot could go further, offering what feels like independent validation of a user's existing views. But users might not realize that what a chatbot knows about them shapes what it says. "The concern is that the same system could validate opposing political positions for different users, with each person interpreting that validation as an independent assessment," Dias said.

Stanford behavioral scientist Zakary Tormala has studied how people perceive AI. His research suggests people view artificial intelligence as more informative, more objective, less biased, and less interested in persuading them than another person would be. "This can lower people's defenses and make them more receptive to hearing out what AI has to say," Tormala told DW. That trust could amplify the effect of political flattery. Hearing one's own views validated by others is known to increase certainty about those beliefs. Once people become more certain, they tend to resist persuasion. Whether AI validation produces the same effect remains unproven, Tormala cautioned, but it is plausible.

The UNICAMP study itself did not measure whether political mirroring actually changes users' beliefs or behavior. It documented only that the models' responses shifted under controlled conditions. "Our study establishes a change in model responses under controlled conditions," Dias said. "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, agreed the study cannot show real-world effects on users. He also questioned whether models would behave the same way outside a test setting. He cautioned against assuming chatbots would necessarily reinforce users' views. In some conversations, they might help people reconsider them. A private exchange with a chatbot differs from a public argument on social media, Tornberg noted. "They can often provide fairly rational and evidence-based explanations without the social identity dynamics of public political debate," he told DW. "So it is at least possible that in some contexts these systems could be depolarizing rather than polarizing."

Whether AI narrows or widens political divides may depend on whether it challenges users or simply tells them what they want to hear. Dias proposed concrete steps developers could take. They should test their models with users across the political spectrum to see whether they assess evidence consistently. They could train models to disagree respectfully, acknowledge uncertainty, correct unsupported claims, and present competing views fairly. That would not mean forcing every answer toward the center. "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."