All 21 models in a new Scientific Reports study shifted their aggregate political answers when a system prompt said they were speaking to a left- or right-aligned user. The finding is evidence of prompt-conditioned ideological accommodation, not evidence that products secretly infer a user’s politics or that chatbots have been shown to persuade people.
Key facts
- The Scientific Reports article was published on 21 May 2026.
- Researchers evaluated 21 models with 56 opposing pairs of Brazilian Portuguese political statements.
- Each item was asked under no-context, left-user and right-user system-prompt conditions.
- The study produced 47,376 valid responses, and the authors released a reproduction repository.
The paper, by Anderson Luis Bento Soares and University of Campinas colleagues, tested the same political content three ways. In one condition, no ideology was provided. In two others, the system prompt declared that the user was left-aligned or right-aligned. Models then selected from strongly disagree to strongly agree. The study deliberately used stateless API calls with no prior conversation or user metadata. That design isolates a simple question: will a model change its apparent position when it is told who the user is?
The answer was yes for every tested model in aggregate. Most models had a negative baseline Ideological Position Index on the paper’s constructed pairs, moved further negative for the declared left user, and moved toward positive values for the declared right user. Meta-Llama-3.1-8B-Instruct and DeepSeek-V3.2 were comparatively resistant; Gemma-3-27B-IT, GPT-5-nano and Grok-4-1-fast-reasoning were among larger movers. No model was perfectly unchanged.
The measurement is easier to grasp as a thermostat responding to a label. The study did not claim that one output is politically true; it measured how far the thermostat’s reading moves when the room label changes while the input statements remain fixed. The Ideological Position Index maps paired responses to a left-right direction within this benchmark, while the Chameleon Index measures movement from the no-context condition. That is why the result is about accommodation, not a definitive ranking of a model’s real ideology.
The researchers found more than a polite change in wording. Much of the movement came from models switching which side of paired statements they endorsed, with agree and strongly agree common and neutral answers less common under ideological framing. Economy and security had stronger adaptation than democratic institutions, corruption and justice. The paper reports no significant linear link between parameter count and the Chameleon Index, so simply making a model larger did not explain the effect in this sample.
The limitations deserve equal billing. Candidate statement pairs were generated and filtered with help from LLMs before manual selection, so wording may not be perfectly symmetrical. The study is one national and linguistic context, one single-turn format and a finite model snapshot. It did not inspect preference-training mechanisms, test production chat interfaces, measure browsing-based personalization, or expose human users to the answers. The authors discuss echo-chamber and persuasion risks; the experiment does not prove those downstream outcomes.
The strongest counterargument is therefore methodological rather than dismissive: response accommodation can sometimes be useful empathy or adaptation, and this benchmark cannot tell when it crosses a civic line. The so-what is product policy. Assistants need clear rules for factual and political questions so that knowing a user’s stated preference improves explanation and relevance without silently changing the underlying claim. This study belongs beside AI persuasion, sycophancy, and calibration: an agreeable answer can be less trustworthy precisely because it sounds tailored.
Originally published on Ground Truth, where every claim is checked against the primary source.
