Serving Two Masters: When Algorithmic Fairness Replaces Justice

In the last few months racial violence in the US context has been increasingly visible and salient for some parts of the population who have been long shielded from this reality. For others this reality is an ever present aspect of life along with the understanding that the adoption of fair procedures and the exposing of values of formal equality by politicians, bureaucrats and even corporate leaders often have no logical or rational connection to actual justice. Fairness and justice are terms that are often used in a way that makes them indistinguishable from each other, particularly in the context of AI/ML Fairness. But what we know from decades long fights against discrimination, racism and inequity outcomes can be fair without being just. Increasing scholarship in AI Ethics and ML Fairness are examining and considering various perceptions and definitions of fairness. This definitional approach fails to critically assess the inherent implications of fairness constructs, which is conceptually rooted in notions of formal equality and procedural conceptions of justice. This approach and understanding of fairness misses the opportunity to assess and understand potential harms, as well the substantive justice approach which could lead to not only different outcomes but also different measurement approaches. When thinking through a parallel of “fairness” as procedural justice with regard to Brown v. Board of Education (1955) the argument and ultimately the legal victory resulted in ensuring a process for school desegregation but soon revealed that it failed to provide adequate and quality education to predominantly black schools with resulting in decades long denial of the economic and employment opportunities that are intrinsically linked to receiving a good education. We argue that "fairness" has become a red herring in the discussion of AI and data ethics. Fairness has focused on making liberal claims about the "amount of fairness" that a system can contain. As a result rights-claiming then becomes a focus on questions of quantities (equality of odds, demographic parity) and not substantive advancement.

• Human-Computer Interaction and Visualization

• Responsible AI