I mean, how much censorship (or artificial curation) would you need to avoid an ML model giving "the climate warrior version" of questions about whether the world was getting warmer?!
Obviously categorizing beliefs into "neutral" and "not neutral" is impossible. Your statement is a classic example of the false consensus effect -- everyone thinks they are neutral.
There is one distinction that matters a lot, and that is empirical vs. moral. Also known as vs. ought (https://en.wikipedia.org/wiki/Is%E2%80%93ought_problem).
When exploring empirical claims, one can aspire for neutrality. Even if perfect neutrality is not possible, it's certainly possible to do better or worse at it.
Moral claims cannot really be evaluated neutrally, I agree. But an AI striving for neutrality could attempt to avoid taking sides on contentious issues.