Modern AI has revolutionized the lending procedures of banks and fintech lenders helping them process loans faster and ensure credit access to those borrowers overshadowed by conventional credit assessments. But some important questions arise with this innovation: Are lenders able to explain the reasoning behind their AI’s decision to approve one applicant while rejecting another? And can they prove that this decision was just?
That question is amusing, but is no longer just theoretical. Upstart Holdings, the AI lending marketplace, received conditional approval from the Office of the Comptroller of the Currency in July 2026 to establish Upstart Bank, N.A., that can become the first nationally chartered bank built around AI-powered underwriting from the ground up. While banks applying for a charter must still seek approvals from the FDIC and the Federal Reserve, Upstart makes the disclosures about fair lending scrutiny of their models being an active risk for upcoming approval.
How AI underwriting works
Traditional underwriting methods depend on limited inputs, primarily credit score, income, and debt-to-income score, that are processed through a rule-based scoring system. With AI-powered underwriting processes, the range of inputs as well as approaches to modeling undergo significant changes. Using ML, one now has the opportunity to analyze thousands of different variables and take advantage of non-standard alternative data patterns, like education background, job history, or banking transaction cash flow signatures.
The practical effect is a model that can identify patterns linking applicant characteristics to repayment likelihood that a traditional scorecard would miss. Lenders using this approach argue it expands access to credit for so-called “thin file” borrowers who lack conventional credit history. The tradeoff is that the more complex the model is, the harder it becomes to state plainly why it reached a particular decision.
Why explainability is the harder problem
Speed and expanded access are the pitch most vendors lead with. However, the regulatory reality is less forgiving. US fair lending law, principally the Equal Credit Opportunity Act, requires lenders to give specific reasons for a denial and prohibits both intentional discrimination and facially neutral practices that produce a disparate impact on protected groups. A model that cannot show its reasoning, or whose reasoning correlates with race, national origin, sex, or age, exposes a lender to legal and regulatory risk regardless of its average accuracy.
This is where Upstart’s own record offers the most instructive real-world case study available, precisely because it comes from an independent monitor’s findings. In December 2020, Upstart entered an agreement with the NAACP Legal Defense Fund and the Student Borrower Protection Center to appoint the civil rights law firm Relman Colfax as an independent fair lending monitor of its lending platform, after the advocacy groups raised concerns that Upstart’s use of education-related data could produce discriminatory outcomes for communities of color.
Relman Colfax published four public reports between 2021 and its final report in March 2024. The monitor found no evidence that variables in Upstart’s model functioned as close proxies for protected classes and identified no pricing disparities, but it did find approval disparities affecting Black applicants. Upstart adopted nearly all of the monitor’s recommendations and enhanced its fair lending testing program. Where the parties could not agree was on the fix itself: the monitor proposed a less discriminatory alternative model, Upstart rejected that specific proposal on the grounds that it would unacceptably reduce model accuracy, and Upstart’s own alternative approach went unvalidated by the monitor, leaving the two sides at an impasse.
At the same time, since improvements were made, Upstart continuously claimed that its AI model could approve more applicants, including Black and Hispanic applicants, at lower APRs than a more traditional underwriting model. In 2023, those reported numbers were that the firm approves 116% more Black applicants and results in APRs that are 34% lower and 123% more Hispanic applicants and results in APRs that are 37% lower than conventional underwriting types. As of 2025, Upstart Access to Credit Report mentions 52% more Black applicants approvals which resulted in APRs that are 32% lower, and 46% more Hispanic applicants approvals with 31% lower APRs.
However, the conversation of Relman Colfax with Upstart sums up the tension at the heart of AI underwriting in one sentence: a lender can develop a model that avoids using explicitly protected traits and yet still creates discrepancies across different groups, and there is no consensus within the industry on how much accuracy a lender should give up in order to eliminate that gap. Relman Colfax’s final report was explicitly framed as guidance for the wider industry, not just Upstart, on how to align underwriting and fair-lending testing as more lenders adopt machine learning models.
What this means for lenders evaluating AI vendors
For banks and credit unions vetting AI underwriting vendors, the monitorship history points to a due-diligence checklist that goes beyond accuracy claims. Lenders should ask whether a model has been tested by an independent party for disparate impact, not just for the presence of protected-class variables. They should ask whether the vendor can produce adverse-action reasons that meet regulatory specificity requirements, and whether any proposed “less discriminatory alternative” has actually been validated rather than merely offered. Upstart’s pending bank charter puts these questions under a brighter spotlight, since a bank subsidiary answers to the OCC, FDIC, and Federal Reserve directly, rather than operating solely through partner-bank relationships.
What this means for consumers
For a consumer wondering why an AI-underwritten application was approved or declined, the practical answer is that the model likely weighed a broader set of data than a traditional lender would, and federal law still entitles that consumer to specific reasons for a denial. Group-level disparities in a monitor’s report do not explain any one borrower’s outcome, but they show why regulators and advocacy groups now treat explainability as inseparable from AI underwriting’s promised gains in speed and access.
