On Testing for Discrimination Using Causal Models

Consider a bank that uses an AI system to decide which loan applications approve. We want ensure the is fair, is, it does not discriminate against applicants based on predefined list of sensitive attributes, such as gender and ethnicity. expect there be regulator whose job certify bank’s fair or unfair. consider issues will have confront when making decision, including precise definition fairness, dealing with proxy variables, what we call allowed variables salary decision depend, despite being correlated variables. show (among other things) problem deciding fairness defined co-NP-complete, but then argue that, in practice should manageable.