Given the nebulous notion of a set of high quality talent (assume quality is a scalar and they're the people above a given threshold), it's pretty clear that:

Given the nebulous notion of a set of high quality talent (assume quality is a scalar and they're the people above a given threshold), it's pretty clear that: len(set(high quality talent) - set(ever employed at facebook)) < len(set(high quality talent) - set(refuses to work at facebook)) Note that "refuses to work at facebook" doesn't select for engineering talent, while "ever employed at facebook" explicitly does, meaning the average talent in the set containing facebook employees is higher. Given all this, you have to fill the roll (or why hire?) and you don't have infinite time (there is a cost associated with every day you've not filled the role relative to it being filled). Given those constraints, if you limit your pool due to prejudice the optimal outcome is for you to hire a lower quality person versus the situation if you were not prejudiced; if you hire the same quality person, it means you likely paid for it in another form, namely the cost associated with the role being unfilled longer.