Fairness in lending is historically fraught, and experts worry that opaque AI models could inadvertently make discrimination worse. To avoid that outcome, researchers argue that combining new data with fairness constraints and better models could both advance equity and protect lenders.

A combination of sophisticated AI models and historically off-limits data could create more equity in lending. 

Researchers from the Federal Reserve Bank of Philadelphia published a paper that argues that sophisticated AI-based credit models should use data that’s currently barred due to the Equal Credit Opportunity Act – like applicants’ location information – to make lending fairer.  

Their basic rationale is that removing sensitive attributes like location hasn’t historically solved the problem of systemic credit gaps. Instead, those attributes should be deliberately integrated into modeling in a way that actively corrects their effects, the researchers say.   

For example, lowering credit thresholds for people in lower-income areas could be a “fairness constraint” that gets integrated into lenders’ modeling.  

To integrate these fairness constraints and drive more equitable results while also mitigating losses, lenders would need to use more complex AI-powered models that can predict risk better than older methods, researchers say. 

“It is crucial to combine the introduction of fairness constraints with better machine learning models,” Vitaly Meursault, one of the researchers, told American Banker. “That will allow lenders to predict default better and compensate the costs of the introduction of fairness constraints, while at the same time reducing credit access gaps to creditworthy consumers.” 

Ultimately, the paper posits that “if the trade-off is managed appropriately, incentives can change in a way that both fairness and profits can improve over time.” In other words, there is a path forward for integrating artificial intelligence into lending that makes the whole system fairer without increasing risk. 

Read more about the research and its potential implications at American Banker.