It is largely trained on online articles, which does have inherently a western media bias in the first place. Any censorship/filtering are done after the fact as part of the hosted service.

Reminder the models do not form their own opinion, they only calculated the most likely response after a question.