AI assistants can inherit politics from their users

A study finds that feedback from strongly partisan user groups can push initially identical AI models toward different political answers. If people then choose assistants partly for political fit, user preferences and model updates could reinforce each other—unless providers pool or balance feedback and test their systems.

Today, asking two competing AI assistants the same political question can produce two polished answers, with no clear way to know whether the difference comes from the models’ capabilities, their rules, or the people whose feedback shaped them. Political sorting has usually changed which news and commentary people consume. It has not usually retrained the sources themselves. AI assistants could make that link more direct.

How feedback changes an AI assistant

A language model is first trained to produce text. After that, providers often adjust it using preference feedback: examples of answers that people like or dislike. This process is meant to make the assistant more helpful and better behaved.

That creates a possible feedback loop. In recommendation systems, what a person watches or clicks can change what the system shows them next. An AI provider could similarly use user reactions to change how an assistant answers later questions.

Political answers are not automatically a shared record of fact. They are generated text. A difference between two answers might involve factual claims, but it might also involve values, framing, or which opinions survey respondents preferred. A provider can reduce the influence of one audience by combining feedback from many users, keeping some rules fixed, and testing changes before releasing them.

When similar models begin to split

The researchers behind an arXiv preprint dated August 11, 2026, tested whether politically different feedback could push initially identical models in different directions. They made paired copies of several language models and fine-tuned one copy using synthetic feedback from an audience that was 90% Democratic and the other using feedback from an audience that was 90% Republican. The feedback came from partisan response patterns in OpinionQA, a survey-based dataset.

The result was a separation in political answers. Across repeated runs, the model trained mostly on Democratic feedback chose answers with greater Democratic relative to Republican survey support than its paired model. For one model, Qwen, the paired copies gave different answers to 40.9% of unseen questions where the partisan gap in the survey was large. The researchers also saw the effect on some kinds of issues that had not been included in the alignment training, although that transfer was less steady in the other models.

This does not show that commercial assistants already work this way. The researchers used synthetic feedback, survey-style multiple-choice questions, a 90/10 audience split, and a particular fine-tuning method. They did not test whether people would actually choose between the specific politically different assistants they created.

The paper then asks what could happen if they did. Its computer simulation represented users choosing among initially similar models while the models updated themselves from feedback supplied by their own audiences. If users partly chose a model because its politics suited them, the audiences became more sorted, and the models became more different. When feedback was pooled instead, the difference in choices fell to 3.8% in one diagnostic, compared with 31.9% when the streams stayed separate.

A different assistant at every kitchen table

If this pattern held in real products, two relatives could ask different assistants the same question about welfare or immigration before an election. Each might receive a smooth, confident answer that treated a different political position as the natural one. The assistants would not merely be separate echo chambers. Their users’ private reactions could help manufacture the political differences that later made people choose between them.

That could weaken the value of a general-purpose assistant as a shared intermediary. Users might think two products differ mainly in price or ability, while feedback routing had quietly shaped their political answers. Providers that pool and audit feedback could instead show that their systems are less influenced by one sorted audience. Independent evaluators could measure whether updates treat different audiences unevenly.

The future described by the paper depends on several things happening together: users must select systems partly for political fit, providers must update models from their own audiences, and the resulting differences must not be made visible or checked. Avoiding that outcome would require feedback to be pooled or balanced, updates to be constrained, and the released systems to face meaningful evaluation in real production pipelines.

Sources

arXiv, “Political Sorting Can Drive AI Models Apart Through User Feedback”