Algorithmic Discrimination Is an Information Problem

Increasingly, algorithms make socially consequential decisions. This makes it crucial to determine when their decisions are discriminatory and how to prevent this outcome. While algorithmic decision-making has proven to be a challenge for traditional anti-discrimination law, there is an opportunity to regulate algorithms through the information that they are fed. But blocking information about protected categories will rarely protect these groups effectively because other information will act as proxies. To avoid disparate treatment, the protected category attributes cannot be considered; but to avoid disparate impact, they must be considered. This leads to a paradox in regulating information to prevent algorithmic discrimination. This paper addresses this problem. It suggests that, instead of blocking or allowing attributes in training data, we should modify them