Abstract
We evaluate LLM responses to a mortgage underwriting task using real loan application data. Experimentally manipulated race is signaled explicitly or through borrower name/location proxies. Multiple generations of LLMs recommend more denials and higher interest rates for Black applicants than otherwise-identical white applicants, with larger disparities for riskier loans. Simple prompt engineering can cost-effectively mitigate these patterns. Race-blind recommendations correlate strongly with real lender decisions and predict delinquency, but LLMs incorporate racial signals when available despite similar delinquency rates across groups. Our findings show potential costs of adopting this new technology in financial settings and raise important questions for regulators.

This publication has 31 references indexed in Scilit: