"Look at applicants as individuals instead of stereotypes" is, I think, missing the broader philosophical question here. Setting aside recruiters optimizing th

"Look at applicants as individuals instead of stereotypes" is, I think, missing the broader philosophical question here. Setting aside recruiters optimizing their own KPIs, hiring is about estimating the value a particular applicant will bring to the organization. And those estimates are hard to make. You are, necessarily, relying on various proxies for their ability within the org, and using those proxies to make more-or-less Bayesian judgements about their value. And those Bayesian judgements are, like any statistics, based on population-level observations - not on the individual. The philosophical question at hand is: is it OK to use a statistically-powerful Bayesian proxy even if that proxy doesn't rely on any unique information about the individual? I think we would all agree "only hire people whose families made six figures during their upbringing" would be a horribly unjust way to hire. But that probably does provide a LOT of meaningful Bayesian information - it correlates strongly with, say, test scores (a trait that many people here probably do think is fair to judge on, and which is also a proxy for a lot of other things), the likelihood of criminal backgrounds, social and political connections, and a million other things. And then you have to ask how far we go with this. Certain genes might be useful proxies - can we go full Gattaca on hiring? What about biases that we know come from terrible social factors (e.g. a black man in 1955 would be very unlikely to be well-educated - would it be ok to use that Bayesian information?). There is a fundamental conflict, one I think is not being acknowledged, between statistical rigor in optimizing local outcomes today and achieving anything resembling individual-level fairness in the medium-to-long term.