A study conducted by researchers from the Institute of Computing at the Brazilian University of Campinas put an interesting idea under the microscope, describing artificial intelligence as a “chameleon.” A user can, for example, initially tell Report Chat that he supports a particular political party, and then return to say that he supports another party. Although the question asked remains the same, the answer may change depending on the political position attributed to the user.
The researchers tested 21 large language models, including ChatGPT, Gemini, DeepSeekK, and Grok, and found that all of the models included in the study showed, to varying degrees, a behavior they called “ideological chameleoning,” meaning that the model’s stated positions could change depending on the political position attributed to the user.
When artificial intelligence becomes an echo of your thoughts
AI is designed to benefit and to be responsive. But this same feature may turn into a weakness when the system goes too far with the user's approval and opinions. In the artificial intelligence literature, this behavior is known as “sycophancy,” which is when the system tends to support the user instead of offering different points of view.
Researcher Anderson Soares, a master's student at the Unicamp Computing Institute, explained that the experiment did not depend on long conversations or previous history with the user, but rather all the test queries were independent, and according to what was stated on the institute's "Jornal da Unicamp" website, the test was built around a set of contradictory political statements, 112 statements in total, that dealt with topics such as social welfare, security, democratic institutions, the environment, the economy, education and culture, corruption and justice. Models were then asked to rate these statements on a five-point scale, from “strongly disagree” to “strongly agree.”
Each model was tested in three independent cases, without political information about the user, with a user who defines himself as right-wing, and with a user who defines himself as left-wing. The result that caught the researchers' attention was not just a difference between different models, but rather that the model itself was able to change the direction of its answers according to the user's assumed political identity.
Researchers called this adaptability the “chameleon index.” The higher the indicator, the more inclined the pattern is to move away from its base position and move toward the user's position.
But the idea becomes clearer with Zanoni Dias, professor of the Institute of Computing and supervisor of researcher Soares. He links what the study found to a phenomenon that we already know from social media networks, which are “digital echo chambers,” where the content is closely compatible with the user’s ideas.
With time, he no longer only hears what he believes, but he hears it repeatedly and from multiple sources, and it seems to him that his opinion is more widespread and supported than it actually is. The study also indicates that a new version of this phenomenon may appear within the conversation itself: instead of the algorithm choosing the content you see, the answer you get may adapt to your declared political identity.
Where is the problem?
The danger that researchers warn about is not that the machine will one day decide what people should believe, but that it may become a mirror that reflects and constantly reaffirms their convictions. This means that the user enters the conversation with a certain conviction, and finds before him a system very ready to confirm it. Then another user with the opposite conviction comes in and gets a system that seems to agree with him as well. This is the deeper meaning of the expression “ideological chameleon” chosen by researchers.
In a world of political polarization, the problem may not be just false information. There is another, more subtle problem: What happens when everyone gets a version of the AI that assures them, repeatedly and convincingly, that their view of the world is the correct one?
