Equalizing Credit Opportunity in Algorithms: Aligning Algorithmic Fairness Research with U.S. Fair Lending Regulation

Legal frameworks exist but may not keep pace with technology: The paper identifies a critical gap between U.S. fair lending laws (designed to prevent discrimination) and how machine learning is actually being deployed in credit decisions. Regulators have not yet issued guidance on whether or how algorithmic fairness techniques should be used, leaving banks potentially operating in a legal gray area.

Machine learning can encode multiple types of unfairness: The research shows that algorithms can discriminate in different ways—some models might deny credit to protected groups at higher rates, while others might appear fair on the surface but use proxy variables (like zip code or education) that correlate with race or ethnicity. Banks could engage in illegal discrimination "potentially unwittingly" through these encoded biases.

Alternative data sources introduce new discrimination risks: The paper highlights that as lenders increasingly use non-traditional data (beyond traditional credit history) to train algorithms, new pathways for discrimination emerge. This underscores why regulators need updated guidance specific to modern machine learning practices in lending.

Computer scientists and policymakers are talking past each other: Fair machine learning research often lacks consideration of actual lending regulations and practical constraints, while regulators haven't engaged with technical fairness solutions. Bridging this gap is essential for developing credit models that are both legally compliant and technically sound.

Alignment between research and regulation is urgent for economic equity: Since unequal credit access perpetuates economic disparities between demographic groups, getting algorithmic fairness right in lending has real consequences for financial wellbeing and opportunity.