Realistically -- and equitably -- modeling the dynamics of group-level
disparities in machine learning remains an open problem. In particular, we
desire models that do not suppose inherent differences between artificial
groups of people -- but rather endogenize disparities by appeal to