Gaslighting Ourselves: Racial Challenges of Artificial Intelligence in Economics and Finance Applications

Creating ethical, equitable and racially-just machine learning is currently impossible for six reasons, summarized by the acronym DO-GOOD : (1) data for training is historically racially biased, (2) ownership of data is disproportionately racially tilted, (3) guidelines /ethical algorithms have been largely lacking, especially in the space of racial equity, (4) opacity of algorithm design prevents testing for racial bias, and then improvement of racial equity, (5) over-reliance on math as a false racially-neutral tool, and (6) decision errors by humans who are subject to their own lived experiences and prejudices. For concreteness, each is demonstrated using examples from economics and finance applications, concluding with suggestions about how to address them (and indeed, how we are already tackling them): they require not only thoughtfulness but regulatory oversight to prevent harm.