Automated decision-making as domination

Abstract

Machine learning ethics research is demonstrably skewed. Work that defines fairness as a matter of distribution or allocation and that proposes computationally tractable definitions of fairness has been overproduced and overpublished. This paper takes a sociological approach to explain how subtle processes of social-reproduction within the field of computer science partially explains this outcome. Arguing that allocative fairness is inherently limited as a definition of justice, I point to how researchers in this area can make broader use of the intellectual insights from political philosophy, philosophy of knowledge, and feminist and critical race theories. I argue that a definition of injustice not as allocative unfairness but as domination, drawing primarily from the argument of philosopher Iris Marion Young, would better explain observations of algorithmic harm that are widely acknowledged in this research community. This alternate definition expands the solution space for algorithmic justice to include other forms of consequential action beyond code fixes, such as legislation, participatory assessments, forms of user repurposing and resistance, and activism that leads to bans on certain uses of technology.