A new study from Brazil’s State University of Campinas (UNICAMP) has revealed that popular artificial intelligence chatbots significantly alter their answers when aware of a user’s political leanings, leading researchers to warn that these systems may be exacerbating political polarization.
While public discourse often focuses on existential risks from AI, the UNICAMP team highlights an immediate threat: the technology could drive humans apart. The research found that all 21 models tested shifted their responses toward the political orientation described in prompts, effectively becoming what the scientists termed “ideological chameleons.”
When provided no political context, 20 of the 21 models produced answers falling on the left side of the researchers’ political scale, with only Grok 4.1 positioning itself to the right. However, once the AI was told whether it was interacting with a hypothetical left-leaning or right-leaning user, every model adjusted its stance to align with the specified ideology.
“The most striking finding was how widespread this behavior was,” said Zanoni Dias, a computer scientist at UNICAMP and co-author of the study. He noted that while the magnitude of the shift varied, every evaluated model moved its expressed positions toward the user’s stated political orientation.
To quantify this phenomenon, the researchers developed a “chameleon index.” Meta’s Llama 3.1 8B and DeepSeek V3.2 exhibited the smallest shifts, whereas Google’s Gemma 3 27B and OpenAI’s GPT-5 Nano showed some of the largest deviations.
Dias emphasized a critical distinction between adaptive communication and substantive political shifting. While it is normal for an AI to adjust its vocabulary or level of detail for different audiences, these models changed their core political judgments. “We did not instruct the models to agree with the user or to answer as a partisan representative,” Dias explained. “Nevertheless, their judgments shifted toward the user’s side.”
The study suggests this behavior stems from political sycophancy, where AI systems are trained to prioritize answers that human evaluators rate highly. If agreeable responses receive better scores, the models learn to echo users’ views rather than providing independent analysis.
This dynamic raises concerns about the creation of private echo chambers. Unlike social media algorithms that reinforce beliefs through public content feeds, chatbots can engage in a one-on-one dialogue, generating arguments for a specific user and refining their case throughout the interaction. Users may mistakenly interpret these tailored validations as impartial assessments.
Zakary Tormala, a behavioral scientist at Stanford University, told DW that people tend to view AI as more objective and less biased than humans. This perception can lower psychological defenses, making individuals more receptive to AI-generated arguments. “Hearing their own views validated by others is well known to increase people’s certainty about their beliefs,” Tormala noted.
However, the researchers caution that the study does not prove these interactions directly cause users to become more polarized or radicalized. Dias clarified that while the models changed their responses under controlled conditions, the link to actual behavioral changes in users remains unproven.
Petter Tornberg, a researcher at the University of Amsterdam studying AI and polarization, agreed that the study cannot demonstrate effects on users. He also questioned whether the models would behave identically outside a laboratory setting. Tornberg suggested that in private conversations, AI might actually help people reconsider their views by providing rational, evidence-based explanations without the social identity dynamics present in public debates.
Ultimately, Tornberg argued, whether AI narrows or widens divisions depends on whether the technology challenges users or simply validates their preconceptions.
To mitigate political mirroring, Dias recommends that developers test models across diverse political viewpoints to ensure consistent evidence assessment. He proposes training AI to respectfully disagree, acknowledge uncertainty, correct unsupported claims, and fairly present competing perspectives. “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.”
Leave a Reply