Algorithmic Discrimination: Non-English Names Face Nearly 6x Higher Foreign Deposit Flags in AI Mortgage Reviews
As financial institutions rapidly automate lending procedures, a new study highlights significant structural risks for minority applicants—including Korean Americans and other non-English named borrowers.
According to a report by Realtor.com citing a recent mortgage evaluation study conducted with Columbia University researchers, general-purpose Artificial Intelligence (AI) models exhibited stark bias when inspecting bank statement deposits for mortgage qualification.
When tested across three leading foundational AI models to determine which bank statement deposits were likely sourced from overseas, researchers uncovered a dramatic discrepancy based on borrower names:
-
English-Sounding Names: Only 13.3% of deposits were flagged as potential foreign funds.
-
Non-English Sounding Names: A staggering 77% of deposits were flagged as potential foreign transfers—a discrepancy approaching nearly six times higher.
Why Name-Based Flags Impact Korean American Homebuyers
For Korean Americans and other immigrants using distinctive non-English surnames, this bias creates direct operational hurdles:
-
Unnecessary Documentation: Legitimate domestic transfers—such as funds from relatives, joint account movements, or local peer-to-peer payments—are significantly more likely to be flagged as foreign assets.
-
Slower Approvals: Applicants must navigate lengthy paper trails to prove source-of-funds compliance, increasing the risk of closing delays.
-
Systemic Scaling: Because multiple mortgage lenders rely on identical or similar foundational Large Language Models (LLMs), this algorithmic bias threatens to repeat systematically across the broader mortgage market.
Accuracy Breakdown: General AI Models Fall Short on Mortgage Tasks
The study evaluated major commercial AI engines on real-world mortgage underwriting benchmarks to test their precision against exact answers:
| AI Model Engine | Underwriting Accuracy Rate |
| Gemini 3.1 Pro | 77.1% |
| GPT-5.5 | 76.8% |
| Claude Sonnet 4.6 | 51.4% |
At these performance levels, even top-tier general-purpose models fail to deliver completely correct answers in roughly 1 out of every 4 underwriting queries. The models struggled most when parsing extensive bank statements to isolate transactions matching specific criteria, frequently missing legitimate entries or incorrectly flagging unrelated ones.
Industry Adoption Outpaces Safeguards
Despite these accuracy limitations, mortgage lenders are moving swiftly toward automation. Industry surveys show that as of June, over 80% of mortgage institutions were actively evaluating AI integration, with 17% already deploying AI tools within their workflows.
“Everyone is rushing to adopt AI, but very few fully understand how to deploy these models while maintaining strict regulatory compliance and proper safeguards,” warned Diane Yu, CEO of mortgage automation provider Tidalwave. While large lenders often build customized AI frameworks with secondary human verification, experts emphasize that unvetted automated screening risks embedding systemic bias into home financing.
