Ask around and you will hear it stated as fact: this chatbot leans left, that one leans right, and somewhere a team of engineers decided it should. The suspicion is understandable. These systems talk in a confident, even-handed voice, so when one declines to praise a politician or frames an issue a certain way, it is easy to read intent behind it. The idea that a model has been handed a party line, though, is mostly a misreading of how it came to sound the way it does.
Where the slant actually comes from
No major lab ships a model with a voting record attached. What political texture these systems have seeps in from three less dramatic places. First, the training data, scraped from a web that is not politically balanced to begin with. Second, the fine-tuning and safety rules, where human preferences about tone and harm shape which answers get reinforced. Third, and this is the part that is only now coming into focus, the feedback of the people using them.
That third source is the subject of a recent preprint, "Political Sorting Can Drive AI Models Apart Through User Feedback," from a group including Petter Tornberg and Christopher Bail. It is worth saying plainly that this is a preprint, not yet peer reviewed, and that parts of it rest on simulations rather than released products. With that caveat in hand, its argument is striking.
The feedback loop
The researchers describe what they call a centrifugal alignment spiral. People with different politics tend to choose different models. Those models then learn from the feedback of the users who stuck around, which nudges them further in the direction their audience already leans. That drift, in turn, shapes who picks which model next. Round and round.
To test it, the team fine-tuned models on synthetic feedback skewed toward Democratic or Republican preferences, then checked how far the paired versions diverged on survey questions they had never been trained on. The gap ran from 12 to 41 percent of questions, always moving in the expected political direction, and sometimes spilling into topics the training had deliberately left out. One more finding cuts the other way, and matters: when feedback was pooled across groups rather than siloed, the divergence largely disappeared.
So, is the fear justified?
Partly, and not in the way most people mean. There is real, measurable evidence that models can hold and express ideological leanings, and that those leanings can be pushed around. What there is no evidence for is the cloak-and-dagger version, a hidden instruction that makes a model campaign for one side. The more plausible risk is quieter and, frankly, more uncomfortable: that our own habit of sorting ourselves could hand each camp a chatbot that mostly agrees with it, with the machine drifting along behind the user rather than leading.
That would not be a conspiracy. It would be a mirror. And it points to a more useful question than "which way does it lean?" Try asking whether a model will tell you something you did not want to hear. A tool that only ever flatters your priors is doing you a disservice, whichever direction those priors run.
Commentarii · 0