It's more likely to work the other way around. Suppose you train your system to try to match human hiring decisions, and some of your hiring managers have an un

It's more likely to work the other way around. Suppose you train your system to try to match human hiring decisions, and some of your hiring managers have an unconscious bias against members of group X, enough to make a small difference in probability of hiring. Training the model will probably find that bias, and the effect may be that belonging to group X, alone, will have a negative weight in the model. So unconscious biases may become explicit. There are ways around this: test the model by presenting credentials that are identical except for (is in group X) vs (is in group Y) and looking for major discrepancies.