At Upstart, we are using artificial intelligence (AI) to rebuild the lending process so that it’s more affordable and inclusive. Central to this effort is ensuring that our lending platform is fair and accurate, and does not perpetuate many of the inequities of traditional credit models. As the pioneer of AI lending, we believe it’s important to articulate how we meet these goals, share our learnings, and hold ourselves accountable.
In a series of posts, we’ll dive into our efforts to avoid unlawful bias on our platform and expand access to credit through four areas of focus:
- How we detect and eliminate potential proxies
- Explainability and generating adverse action notices with AI
- Less discriminatory alternative model searches
- Our approach to fair-lending testing, and five ways we believe it could be improved for the benefit of the industry.
We believe our comprehensive approach to credit modeling already leads to more fair outcomes than current methods. Many of the most commonly used underwriting tools cannot effectively evaluate the diversity of factors necessary to determine the true risk of a borrower. Even worse, this inefficiency disproportionately harms underserved communities by assuming that large swaths of a particular community may pose significant risk, reducing their access to affordable credit and widening the economic divide.
We aim to solve these problems using advanced technology powered by machine learning and AI. Each year, we evaluate the ability of our personal loan model to underwrite applicants in comparison to a more “traditional” model. We’re pleased to report that our most recent research showed that the Upstart model could approve more applicants, including Black and Hispanic applicants, at lower APRs than a more traditional underwriting model. In comparison to the traditional model, the Upstart model:
- Approves 101% more applicants and results in APRs that are 38% lower.
- Approves 116% more Black1 applicants and results in APRs that are 36% lower.
- Approves 123% more Hispanic1 applicants and results in APRs that are 37% lower.
In a distressed macroeconomic environment, many lenders are forced to tighten their credit standards significantly, using blunt tools that harm access to credit in underserved communities. By lending with a more accurate, AI-driven model, lenders have the opportunity to continue lending to a broader set of consumers without taking excess credit risk. You can read more about the analysis that led to these conclusions in our 2023 Access to Credit Report, which can be found here.
We continue to seek out and establish best practices to ensure fairness in every step of the lending process. In the coming posts, we will explore what that looks like. By the end of this series, we hope to have shown the advantages of our approach, where we can improve, and how our industry-leading techniques are reducing the inequities that are so prevalent in traditional lending systems.
In our next post: Upstart’s patent-pending approach to detecting and eliminating proxies in underwriting models.
__________________________________________________________________________________________________
1. Upstart does not collect demographic information from its applicants. The demographic attributes are estimated via location and surname using the BISG methodology.