Post

WP thread: “Measuring and Mitigating Racial Bias in LLM Mortgage Underwriting” with Don Bowen, McKay Price, and Ke Yang Our audit study asks AI to assess simple mortgage applications (real HMDA data w/ randomized race and credit scores) 🦶🦶🏼🦶🏿 Download lukeste.in/llmmortgage 1/

LLMs recommend denying more loans and charging higher interest rates to Black applicants Especially at low credit scores/riskier loans Simple prompt engineering can help mitigate gaps

I.e., ChatGPT (we also assess other models) • Claims it’s unbiased… • …but it 𝗶𝘀 biased against Black applicants. Particularly at low credit scores. (We also have this for other risk measures.) • We can partly close the racial gap just by asking for unbiased responses. 3/

LLMs recommend denying more loans and charging higher interest rates to Black applicants They would, on average, need credit scores ~120 points higher than white applicants to receive the same approval rate; ~30 higher to get same interest rate 4/

LLM’s mortgage racial bias biggest for applications with lower credit scores [or high DTI, high LTV]: • The Black–white approval rate gap is ~56% greater for low-score applicants than at average (13.3pp vs. 8.5) • Interest rate gap ~32% greater (47bp vs. 35) 5/

We find anti-Black bias in mortgage underwriting recommendations from a number of LLMs Black borrowers with low credit scores suffer the most 6/

(Training data incl. long history of racial disparities in mortgages, plus prompts may trigger bias learned elsewhere) We give access to explicit race information, which should make it easy to 𝘢𝘷𝘰𝘪𝘥 discrimination if LLMs can ignore it, as they know they should… 7/

Somehow, just asking LLM to be unbiased • Eliminates approval recommendation gap (on average and across different credit scores) • Reduces average racial interest rate gap by about 60% (from 35bp to 14), with even larger effects for lower-credit-score Black applicants 8/

Thanks and apologies to @lukestein.com for the thread copy. Can't improve on this!