Answers are grounded in this paper and cite their sources.
Mortgage underwriting is entering a new phase of technological change as general-purpose generative AI tools begin to appear in mortgage origination workflows. We develop an audit framework to evaluate how large language models process mortgage applications. Using real HMDA applications, we experimentally manipulate borrower race, race proxies, and credit quality, and prompt leading commercial LLMs to make approval and pricing recommendations. LLMs recommend more denials and higher interest rates for Black applicants than otherwise identical white applicants, with disparities largest for riskier loans and persisting when race is signaled through names or locations. A simple instruction to avoid bias substantially reduces disparities when race is explicit, partially reduces the name-based pricing gap, and does little against the geographic pricing gap. In matched HMDA–Freddie Mac data, race-blind LLM recommendations correlate with actual lender pricing and subsequent delinquency, though the delinquency association operates largely through observed credit scores. The results show how AI auditing can inform mortgage underwriting governance before general-purpose models become more deeply embedded in origination workflows.
No takes yet. Share an insight, caveat, or question.
Bowen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: