Learn the answer to the 2024 Biz Quiz.
Answer (Online Only)
The mitigation strategy used in the paper works best on approval decisions, where it almost completely eliminates bias. It does this by increasing the approval gap for Black applicants with a high credit score by 5%, but it works especially well at the low end of credit scores, increasing approvals by almost 15% to erase the racial gap.
The mitigation strategy works when asking ChatGPT to suggest interest rates, but not as well: a 0.2% gap remains between Black and White applicants.
The reason for this difference is…unknown! And the subject of ongoing research for Professor Bowen and his research partners. LLMs are new technology and their workings are opaque. They suspect that the bias ingrained into ChatGPT might manifest differently for a yes/no approval question than asking for a continuous number; the latter could allow more subtle forms of bias to seep into the reply.