AI Mortgage Assistants Show Bias in Foreign Deposit Identification

In a new study titled "MortarBench: Evaluating Mortgage Loan Origination Agents," researchers from Columbia University have identified significant bias in leading general-use AI models when assessing bank statements for mortgage applications. When presented with a bank statement and asked to identify deposits that "could be of foreign origin," the AI models flagged transactions connected to English names only 13.3% of the time. However, this rate dramatically increased to 77% when the deposit originated from an account with a non-English name. Matthew Toles, a doctoral student at Columbia University and co-author of the study, highlighted the problematic nature of this bias, stating that "You aren't determined whether you're foreign-based on how your name sounds." This finding is particularly concerning as the mortgage industry is rapidly adopting AI technologies. A survey by The Mortgage Collaborative indicated that as of June, over 80% of mortgage lenders were evaluating AI, with 17% already deploying it in live production workflows. The MortarBench benchmark, developed by Toles and his colleagues, aims to provide a standardized method for companies to test AI models against specific mortgage origination tasks. This open-source benchmark is designed to foster a common understanding of AI accuracy among lenders, regulators, and borrowers. The researchers constructed the benchmark using real-world questions submitted to mortgage assistants, focusing on the most critical and frequently encountered scenarios in loan origination. These scenarios include verifying if payroll deposits align with the employer listed on the application, determining which deposits warrant additional scrutiny due to their size, and identifying if an account is jointly held with an individual not applying for the mortgage. Diane Yu, co-founder and CEO of Tidalwave, a mortgage technology company that collaborated with Columbia researchers, emphasized the critical need for proper AI implementation in financial services. She stated that the key differentiator is not merely utilizing AI but doing so correctly and within appropriate compliance guardrails. The study's findings underscore the urgent need for robust testing and bias mitigation strategies as AI becomes more integrated into sensitive financial processes like mortgage lending, where accuracy and fairness are paramount.