A language model produces the most probable continuation given what it has read. On a controversial question, that means it produces roughly the position most represented in its corpus. This is not a bias that could be corrected: it is the definition of what it does. The question is what that produces when millions of people consult the same average.
We have covered the impossibility of neutrality and the tendency to agree with you. Here is a distinct, and deeper, question.
What makes the problem real
A new idea is, by construction, a minority one. It is poorly represented in the corpus, often badly formulated by its early advocates, and frequently considered false at the moment it appears.
That is what the test we covered suggested: when models are asked to recover the idea of a scientific paper from its premises alone, scores fall between 3 and 15%. A system optimised for plausibility is structurally ill-suited to producing what was not plausible before someone thought it.
The risk is therefore not that AI imposes an opinion. It is more subtle: it makes consensus more accessible than its alternatives. You need a particular effort to get the minority position, while the dominant position arrives by default.
Three nuances that matter
This is not new. Textbooks, encyclopaedias, search engines and the media all produce a version of consensus. AI does not invent this dynamic, it accelerates and personalises it.
Consensus is often right. This is the point forgotten when dissidents are celebrated. For every minority voice that was right against everyone, there are thousands that were simply wrong. A tool that gives the majority position most often gives you the right answer.
You can ask for something else. A model is perfectly capable of laying out a minority position if explicitly asked. The problem is not that it cannot do so, but that it does not do so spontaneously.
Here is what seems most concerning to us, and it is not the content of the answers. It is that millions of people consult the same source to form an opinion. Previously, diversity came as much from fragmentation as from quality: everyone read different things, heard different versions, and those gaps produced debate. A single source, even an excellent and balanced one, mechanically reduces that variety. The danger is not that it is wrong, it is that everyone has the same opinion for the same reason.
What makes the problem more serious than before
Two differences from traditional sources deserve to be set out.
Personalised phrasing. An encyclopaedia says the same thing to everyone, and you know you are reading a general text. A model reformulates for you, which gives the impression of reasoning built for your case. The average position arrives dressed up as personal reflection.
The absence of visible disagreement. When you consulted several sources, divergences appeared. A model synthesises and produces a smooth position, where the field's real disagreements are flattened into a cautious balance. This is, as we noted, one of the markers of generated text.
What you can do
Three simple practices, and they work.
Ask for disagreement explicitly. Not "what is the view on X" but "what are the points on which specialists in X disagree, and why". The answer changes completely.
Ask for the best counter-argument. This is the most useful exercise you can do with these tools, and probably the least practised.
Keep sources that do not go through this. Reading someone with an avowed point of view, with their biases, brings something a synthesis does not: the sense that someone is defending something.
What to remember
The question in the title calls for a nuanced answer: AI does not lock anyone in, it makes one exit easier than the others. And what is easier ends up being what you do.
What is at stake is therefore not the censorship of minority ideas, but something more discreet: the disappearance of the effort once required to encounter a point of view you were not looking for. A library, a conversation, a newspaper exposed you to things you had not asked for. A tool that answers exactly what you ask never does that.
That is perhaps the most solid reason to keep frequenting sources that do not answer you.