AI and housing discrimination: the case of mortgage applications

Abstract

Issues surrounding bias and discrimination in housing markets have been acknowledged and discussed both in the literature and in practice. In this study, we investigate this issue specifically in the context of mortgage applications through the lens of an AI-based decision support system. Using the data provided as a part of the Home Mortgage Disclosure Act (HMDA), we first show that ethnicity bias does indeed exist in historical mortgage application approvals, where black applicants are more likely to be declined a mortgage compared with white applicants whose circumstances are otherwise similar. More interestingly, this bias is amplified when an off-the-shelf machine-learning model is used to recommend an approval/denial decision. Finally, when fair machine-learning algorithms are adopted to alleviate such biases, we find that the “fairness” actually leaves all stakeholders—black applicants, white applicants, and mortgage lenders—worse off. Our findings caution against the use of machine-learning models without human involvement when the decision has significant implications for the prediction subjects.