Second Opinions: Human Review Beyond Algorithmic Credit Pricing

We examine human intervention in AI-assisted lending using data from LendingClub, where a subset of borrowers is randomly selected for verification, enabling human review of algorithmic decisions. Exploiting this institutional variation and restricting attention to borrowers with identical observable credit profiles, we identify the causal effect of human intervention. We find that human verification leads to significantly higher interest rates, even among observationally identical borrowers. Moreover, human-induced interest rate adjustments predict ex-post default beyond the AI prediction, reflecting superior screening of latent borrower risk rather than higher interest rates driving borrower outcomes. We prove these findings as arising from human screeners processing the same borrower information differently and potentially applying conservative adjustments because of agency concerns or uncertainty about AI recommendations. Our results highlight the role of heterogeneous information processing in human-AI decision-making.