Towards a Theory of Proxies (for Race)

César (they/he) is a Ph.D candidate in Philosophy at Stanford University. They specialize in political philosophy and their interests lie at the intersection between democratic theory, environmental ethics, AI ethics, and philosophy of race. In their dissertation, César examines how climate change pushes us to rethink the limits of democratic legitimacy.

Attend this talk via ZOOM

Talk: Towards a Theory of Proxies (for Race)

Abstract: The question about what counts as a proxy for race or other protected categories has become newly relevant with the use of machine learning models (MLMs) in decision-making. What is the relation, if any, between a feature F (e.g., education), encoded in a given model M, and race (which may not be encoded in M) such that F is a proxy for race and a decision made on the basis of F counts as a decision made on the basis of race? If there is an answer to the “proxy question” and there is such thing as proxies for race, then decisions using those proxies may count as racial discrimination. In particular, MLMs would exacerbate the phenomenon of “proxy discrimination” insofar as race and other legally protected characteristics remain predictive of certain outcomes: models will seek for proxies for those characteristics in order to make their predictions. The literature offers several responses to the proxy question. In this paper, I identify the strengths and weaknesses of four of these responses: statistical approaches, thick constructivist approaches, explanatory approaches, and deflationary approaches. This assessment allows me to identify the desiderata for a theory of proxies; I then propose a theory that fulfills those desiderata. I will argue that the most promising theory of proxies combines a statistical approach with some form of thin constructivist account of how race causes individual and collective decision-making. Finally, I emphasize the strengths of my theory vis-à-vis the available accounts and anticipate potential objections.