AlphaGo was trained on high-level human games, so had those biases. AlphaZero was only given the rules and trained from self-play, so didn't have any human biases in evaluating outcomes and outplays humans by a larger margin than AlphaGo.

We can't effectively/safely do that in other fields due to the time and danger to train by letting the machine make an (initially random) decision, record the outcome, repeat. I suppose we have to do the painstaking work of unbiasing like is done for important surveys but isn't easily done on the scale of training sets.