A Problem With AI-Based Credit Models

A Problem With AI-Based Credit Models

February 6, 2024

By: Tyler Brown

The use of technology does not excuse unlawful discrimination, according to federal regulators. But as artificial intelligence (AI) models make their way into underwriting, they introduce the potential for discrimination that fair lending laws and banks’ compliance practices aren’t designed to handle. AI’s capabilities and the murkiness of models may get in the way of lending decisions that are sound, legal, and explainable. Regulators are watching: In 2023, four federal agencies published a statement about enforcing fair lending laws and regulations in the age of AI. Their bottom line: Complexity and opacity aren’t a defense against violations of fair lending regulations.

A series of laws first passed in the 1960s addressed a problem of an age long past: One in which limited information to inform loan decisions allowed too much discretion to loan officers. Today’s problem is the opposite: As the Bank Policy Institute argues, there is too much data, and regulations based on fair lending laws were not designed with today’s wide availability of demographic and behavioral data and raw computing power in mind.

Even with objective analysis based on a lot of data, bias exists as “disparate impact.” That occurs when an underwriting model has a “disproportionately adverse impact on applicants from a protected class, unless the practice meets a legitimate business need that cannot reasonably be achieved by means that are less disparate in impact,” according to the Bank Policy Institute.

As the banking industry has computerized, it’s made do with computational underwriting models that account for that compliance challenge. Models have reportedly been simple enough to be explainable, generate the legally required rationale for credit decisions, and allow banks to test systems for compliance. But AI poses challenges to that status quo because AI-based credit underwriting systems do not always transparently or accurately provide the specific reasons for adverse credit decisions, putting them out of compliance.

As lenders assess challenges with AI-based underwriting, they have two big issues to keep in mind, according to the Consumer Financial Protection Bureau (CFPB):

Black box algorithms: Fair lending laws require banks to explain “the specific reasons for denying an application for credit or taking other adverse actions, even if the creditor is relying on credit models using complex algorithms.” If nobody knows exactly how the model works, including its developers, it may not be possible to comprehend the algorithm’s decision-making process or determine its biases.
Digital redlining: Statistical models can infer characteristics protected by fair lending laws, and intentionally or not, those characteristics can be reflected in underwriting decisions. As the Brookings Institution argues, there are “too many sources of data that can hide as proxies for illegal discrimination.” The Equal Opportunity Credit Act specifically prohibits discrimination based on “race, color, religion, national origin, sex, marital status, age” and several other factors.

Despite real issues, the ability to evaluate more data for credit decisioning using AI systems with the appropriate guardrails is an opportunity for lenders. Traditional credit scoring and reporting has its own biases, including potential borrowers’ need for a credit history to acquire more credit. Holistic analysis using AI models that includes “alternative” credit scoring data made available through open banking, for example — like bill and rent payments and spending patterns — may make credit scoring fairer. But solving the problem of bias in AI systems will need to happen first. That, ultimately, will require solving for bias in data the models are trained on — which itself is complicated to balance.