Artificial Intelligence Showcases Discriminatory Practices in Residential Loan Approvals Based on Race
In a groundbreaking study, researchers from Lehigh University, along with colleagues from Babson College and the College of Business, have shed light on the issue of racial bias in mortgage lending decisions made by Artificial Intelligence (AI) models [1].
The study, currently available as a working paper, reveals that AI models can reflect underlying data biases or societal inequities present in their training data, perpetuating discrimination in mortgage lending decisions [1]. To tackle this challenge, the researchers focused on instructing AI to avoid bias in decision-making as a potential solution [1].
The team manipulated race and credit score variables to determine their effects on mortgage approval decisions made by leading commercial large language models (LLMs), including OpenAI's GPT 4 (2023), ChatGPT 3.5 Turbo, Anthropic's Claude 3 Sonnet and Opus, and Meta's Llama 3-8B and 3-70B [1].
The findings confirmed that AI models, when not specifically instructed to avoid bias, would consistently recommend denying more loans and charging higher interest rates to Black applicants compared to otherwise identical white applicants [1]. However, when LLMs were instructed to ignore race, approval decisions for loans became indistinguishable between Black and white applicants across the credit spectrum [1].
Interestingly, ChatGPT 3.5 Turbo was found to show the highest discrimination, while ChatGPT 4 (2023) exhibited virtually none [1]. The study also found that Black applicants would require approximately 120 points higher credit scores on average to receive the same approval rate as white applicants [1]. Similarly, about 30 points higher credit scores were necessary for Black applicants to receive the same interest rate as white applicants [1].
The researchers also observed bias against Hispanic applicants, albeit to a lesser extent than against Black applicants [1].
The study underscores the importance of documenting and understanding biases in AI tools used in financial decision-making. Lenders and regulators are encouraged to proactively assess the fairness of LLMs and evaluate methods to mitigate biases [1].
AI can be programmed to avoid bias in decision-making, but comprehensive approaches, including fairness-aware training, auditability, and regulatory compliance, are necessary to meaningfully mitigate racial bias in mortgage lending decisions with AI technology [1].
Reference: [1] Price, M., Bowen III, D., Yang, K., & Stein, L. (2023). The Impact of AI on Mortgage Lending: A Study on Bias and Fairness. Working Paper.
In an episode discussing the widespread uses of AI in financial services, Donald Bowen III, one of the authors, shared insights from the study on the ilLUminate Podcast from the College of Business [2].
| Aspect | Details | |-------------------------------------------|---------------------------------------------------------------------------------------------| | Mitigation mechanism | Instructing AI to avoid bias avoids direct discriminatory indicators in decision logic[1] | | Challenge | Underlying data biases and lack of explainability may still produce biased outcomes[3] | | Testing with commercial LLMs | LLMs improve throughput and document handling but require specialized design to mitigate bias effectively[2] | | Effective mitigation in practice | Purpose-built AI mortgage platforms with fair lending monitoring ensure regulatory compliance and reduce bias[3] |
The study suggests that AI models can perpetuate discrimination in mortgage lending decisions by reflecting data biases or societal inequities [1]. To mitigate this, researchers propose instructing AI to avoid bias in decision-making as a solution [1]. Furthermore, the findings indicate that leading commercial large language models (LLMs) exhibit racial bias in mortgage approval decisions, but when instructed to ignore race, approval decisions for loans become indistinguishable between Black and white applicants [1]. This highlights the importance of proper programming and mitigation strategies to combat racial bias in AI technology used in finance.
