If the biased models replace human decision making, then it just has to be shown that the models are less biased than humans which may not be that high of a bar to pass.