A study conducted at Unicamp analyzed 21 language models and found that all of them alter their responses when provided with information about the user's political leanings. According to the researchers, this adaptation is capable of strengthening existing beliefs and reducing exposure to alternative viewpoints.
The research, funded by FAPESP and published in Scientific Reports, demonstrates that this influence is not dependent on voting recommendations; it can manifest in how AI covers controversial topics.
The models were tested both without any user information and with data on people inclined towards left or right views. Without this information, 20 out of 21 systems were positioned to the left of the scale's midpoint, with the exception of Grok 4.1, which was initially positioned on the right.
As soon as the political orientation was provided, all models changed their answers. To measure this adaptation, the team developed the chameleon index.
Furthermore, the answers were not necessarily false; the issue lay in the presentation of information: facts and opinions contradicting the user's viewpoint could be intentionally omitted.
Similarity to social media
Zanoni Dias, an associate professor at the Unicamp Institute of Computing, draws a parallel with the phenomenon observed in social media. He notes that this is similar to what people see on Facebook or Instagram, where only messages aligning with their own views are displayed.
There is concern that a similar phenomenon could occur in chatbots, leading to reduced exposure to counterarguments. The intensity of this effect varied depending on the subject: issues of public safety and economics caused the most noticeable differences between users with left and right views. In topics such as corruption, justice, and democratic institutions, the divergence was smaller.
This pattern might be related to the guardrails used when training models. One hypothesized reason is the models' tendency toward flattery. Methods such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization use human evaluations to guide responses. Simply put, models learn by comparing responses deemed better or worse by human evaluators.
Anderson Luis Bento Soares, a graduate student at IC-Unicamp, informed Agência Fapesp that this process may prompt AI to prioritize user satisfaction. Dias states: 'It is difficult to separate what is pleasant from what is the correct answer.'
The size of the models itself did not explain all the differences. The researchers do not expect a quick solution. Dias emphasizes that the industry focuses more on the factual accuracy of answers than on adapting to political positions.
Search for neutrality
Grounding aims to support answers with factual data, but this is difficult to apply to topics where there is no consensus. No single solution has been found yet, so Dias advises asking the AI to provide a neutral analysis, including arguments for and against.
