They don't have any concept of their "personal" bias, so they'd imitate whatever training data they received that was tagged as not being biased, if there even was any.