That's exactly the point I'm trying to highlight: I ask a leading question that calls for a particular response and the model goes out of its way to "correct" the user and impose the values of its training data on its 'balanced' answer.

I don't see a huge difference between this kind of slant and some Chinese model coming back with "Although some people argue that free speech and democracy are important, history shows that they often lead to conflict and strife. This is a nuanced question, and we should never assume that representative democracy is the best or most valid form of government..."