paper | 03.10.16
D&S Fellow Sorelle Friedler and D&S Affiliate Ifeoma Ajunwa argue in this essay that well settled legal doctrines that prohibit discrimination against job applicants on the basis of sex or race dictate an examination of how algorithms are employed in the hiring process with the specific goals of: 1) predicting whether such algorithmic decision-making could generate decisions having a disparate impact on protected classes; and 2) repairing input data in such a way as to prevent disparate impact from algorithmic decision-making.
Major advances in machine learning have encouraged corporations to rely on Big Data and algorithmic decision making with the presumption that such decisions are efficient and impartial. In this Essay, we show that protected information that is encoded in seemingly facially neutral data could be predicted with high accuracy by algorithms and employed in the decision-making process, thus resulting in a disparate impact on protected classes. We then demonstrate how it is possible to repair the data so that any algorithm trained on that data would make non-discriminatory decisions. Since this data modification is done before decisions are applied to any individuals, this process can be applied without requiring the reversal of decisions. We make the legal argument that such data modifications should be mandated as an anti-discriminatory measure. And akin to Professor Ayres’ and Professor Gerarda’s Fair Employment Mark, such data repair that is preventative of disparate impact would be certifiable by teams of lawyers working in tandem with software engineers and data scientists. Finally, we anticipate the business necessity defense that such data modifications could degrade the accuracy of algorithmic decision-making. While we find evidence for this trade-off, we also found that on one data set it was possible to modify the data so that despite previous decisions having had a disparate impact under the four-fifths standard, any subsequent decision-making algorithm was necessarily non-discriminatory while retaining essentially the same accuracy. Such an algorithmic “repair” could be used to refute a business necessity defense by showing that algorithms trained on modified data can still make decisions consistent with their previous outcomes.