Measuring and Mitigating Racial Disparities in Large Language Model Mortgage Underwriting
The paper investigates racial biases in LLM-based mortgage underwriting, quantifying disparities and evaluating bias mitigation strategies through prompt engineering.
What it examines
This paper audits racial disparities in mortgage underwriting decisions made by large language models (LLMs). By experimentally altering applicant race and credit scores, the study evaluates how LLMs compare to real lender behavior. It uses regression analysis and prompt engineering to detect and reduce bias for fair lending practices.
What it concludes
Results reveal that LLMs replicate and intensify existing racial biases in mortgage underwriting. Simple mitigation prompts significantly reduce these disparities, suggesting potential applications in automating unbiased loan approvals and shaping regulatory standards. Future research should enhance debiasing strategies and extend these methods to other financial decision processes.
Evidence objects
A study finds mortgage underwriting LLMs are biased: Black applicants require up to 120 extra credit points for approval and face higher interest rates than whites despite identical finances overall.
key_findings bullet 1 · key_findings · validation V0
Researchers employed a rigorous audit design combining real HMDA loan data, manipulated applicant demographics, and regression analysis with loan-fixed effects to reveal systematic racial biases in multiple large language models.
key_findings bullet 2 · key_findings · validation V0
A prompt engineering strategy instructing unbiased model decisions dramatically reduced racial disparities in loan recommendations, yet further research on calibration and external validity is crucial for ensuring fair financial outcomes.
key_findings bullet 3 · key_findings · validation V0
This innovative paper tackles critical credit market challenges by combining real loan data with experimental manipulation of race and credit scores, revealing AI bias in mortgage underwriting. Its original design and novel use of LLMs yield significant insights and practical mitigation strategies, offering compelling contributions to quantitative and computational finance.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
- … in various economic sectors, including credit markets. AI promises lower costs and … race and credit scores. By asking various leading commercial LLMs to recommend …
Source row: 1311 · abstract type: snippet