In December 2020, Upstart Network (“Upstart”), the NAACP Legal Defense Fund (“LDF”), and the Student Borrower Protection Center (“SBPC”) entered into an agreement to appoint Relman Colfax to serve as an independent fair lending Monitor to evaluate and make recommendations regarding the fair lending implications of Upstart’s lending platform, and to issue a series of periodic reports on its findings and recommendations. Those reports are published below, upon release.
Upstart’s lending platform relies on Machine Learning-based Artificial Intelligence (“ML” and “AI”) models and non-traditional applicant data—including data related to borrowers’ higher education—to underwrite and price consumer loans. LDF and the SBPC raised concerns with Upstart that the use of educational criteria can lead to discriminatory lending outcomes, particularly for communities of color, leading to the appointment of an independent fair lending Monitor.
An Initial Report, published in April 2021, provides context for the Monitorship, including legal and historical background regarding lending disparities and fair lending testing, a summary of Upstart’s lending program and model, and a description of studies and communications related to Upstart from the SBPC, LDF, and members of Congress.
On November 10, 2021, Relman Colfax published a Second Report. The Second Report addresses ongoing analyses of whether Upstart’s model results in a disproportionate adverse impact on any protected classes, and if so, whether there are alternative practices that maintain the model’s predictiveness but result in fewer disparities. This Report discusses results regarding disparities, as well as our methodologies for choosing among potential alternatives that result in fewer disparities. However, our evaluation of alternatives is ongoing and we are not yet in a position to determine whether a viable alternative model exists and, if so, what changes to Upstart’s Model we would recommend. This Report also analyzes whether variables used in Upstart’s model function as close proxies for protected classes. Our analyses suggest that Upstart’s input variables do not appear to be meaningful predictors of race and national origin. That said, given the AI/ML nature of Upstart’s Model, our methodologies have inherent limitations and we cannot eliminate entirely the possibility that proxies exist. Future reports will address further analyses, including whether we will make any recommendations to Upstart regarding the adoption of an alternative model that results in fewer disparities.
LDF, SBPC, and Upstart have jointly issued a release regarding the Second Report, available here.