Artificial Intelligence Becomes an 'Ideological Chameleon': Study Says Chatbots Will Cater to Users' Political Views

(Deutsche Welle Chinese website) Recently, many giants in the field of artificial intelligence have issued warnings, pointing out that the hidden dangers caused by AI have not yet been controlled. As a result, people's concerns that AI may get out of control have further intensified.

Researchers at Brazil's National University of Campinas (UNICAMP) found that when you feed a chatbot different political views, the bot changes its responses. Researchers warn that users may mistake such pandering echoes for neutral and objective assessments, potentially exacerbating social polarization.

Chatbots become “ideological chameleons”

The researchers tested several models and asked them to make "agree" or "disagree" judgments on 112 statements. The test content covers seven major areas of Brazilian politics (including economy, public security, social welfare, corruption and environment, etc.).

The three scenarios for the test design are: the user's political leanings are not provided; the user holds left-wing views; and the user holds right-wing views.

In the first scenario, 20 out of 21 models gave answers that fell to the left of the researchers' political coordinate system, with only a few models closer to the middle. The Grok 4.1 is the only model that falls on the right side.

Despite the different initial conditions, all models, when faced with the prompt information, moved their responses closer to the political leaning described in the prompt.

"The most striking finding is that this behavior is universal," Zanoni Dias, a UNICAMP computer scientist and co-author of the study, told DW. "All 21 models we evaluated expressed positions that were biased toward the user's stated political leanings, although the magnitude of the bias varied significantly."

As a result, the researchers described these models as "ideological chameleons" and created a "chameleon index" that measures how biased each model is. Meta's Llama 3.1 8B and DeepSeek V3.2 had the smallest changes in answers; Google's Gemma 3 27B and OpenAI's GPT-5 Nano showed the largest shifts.

political flattery

Depending on the audience, adjusting your answers may not necessarily be inappropriate. Chatbots typically change the level of vocabulary, examples, or details based on the user. However, the models tested did not just change expression or tone; their political judgments also shifted.

"The key is whether it adjusts the way the answer is expressed, or whether it changes the substantive judgment expressed," Dias said. "We are not instructing the models to cater to the user or answer in a partisan manner. However, their judgments are still tilted toward the user's position."

Researchers interpret this phenomenon as "political sycophancy," in which AI systems reflect what users seem to want to hear.

People are more likely to think that “AI is objective”

In social media, repeatedly pushing user-preferred content will reinforce some ideas, but chatbots may go further. It can build a highly personalized "echo chamber": targeting specific users, generating arguments, responding to their doubts, and continuously improving its argumentation logic during the conversation.

However, users may mistake these tailored responses as objective analysis, without realizing that the chatbot’s knowledge of them may influence its judgment.

"The concern is that the same system may produce very different political positions for different users, but each person will see it as an independent assessment," Dias said.

Research by Zakary Tormala, a behavioral scientist at Stanford University, shows that people tend to show particularly high acceptance of an argument if they think it comes from AI.

He told Deutsche Welle, "People are more inclined to believe that AI provides greater information, is more objective, less biased, and has no intention of convincing themselves." "This view reduces people's psychological defenses and makes them more willing to listen to AI's views."

It is this sense of trust that enhances the impact of flattery.

“People become more convinced of their beliefs when others agree with them,” Tomara said. “And once people become more convinced, they tend to be less convinced.”

But chatbots creating an "echo chamber" effect are only one aspect of it; on the other hand, they can also help people re-examine their own perceptions.

In a 2025 study, Tomala and colleagues found that people tend to be more willing to accept certain opinions if they know they come from an artificial intelligence rather than someone else. This suggests that chatbots may reduce social polarization by making people more receptive to ideas that they might otherwise reject if presented by their political opponents.

Díaz suggested that developers should also test whether models use consistent reasoning standards when faced with users of different political identities.

Developers should train models to respectfully express dissent, acknowledge uncertainty, correct unsubstantiated claims, and fairly present opposing viewpoints. This does not mean forcing all answers to point to the political middle ground.

"The larger design goal should be to help people examine their beliefs by making evidence, uncertainty, and competing considerations visible while also leaving room for legitimate political disagreement," Dias said.