Machine + man: A field experiment on the role of discretion in augmenting AI-based lending models

Abstract

We assess the role of human discretion in lending outcomes using a randomized, controlled experiment. The lenders in our sample utilize a third party, machine-generated credit model as an input in their decision. We design a new feature for the credit-scoring platform – the slider feature – which invites lenders to incorporate additional discretion in their decision by adjusting the machine-based recommendation. We compare the loan outcomes for treatment lenders that randomly get the slider, relative to a control group. The treatment group's adjustments are predictive of forward looking portfolio characteristics – they show larger declines in future portfolio-level credit risk and larger increases in future sales orders, relative to the control group. The effects of our intervention are more pronounced when borrowers do not have social media accounts and in competitive markets. Our study provides insights about the role of human decisions, given the rapid evolution of machine-based lending models.