Interpretable Algorithms as a Potential Solution to CFPB’s Guidance on AI-driven Credit Denials

In September 2023, the Consumer Financial Protection Bureau (CFPB) issued guidance on the use of artificial intelligence in issuing credit denials, a prevalent practice among lenders.  The CFPB explained that when denying a credit application, lenders must provide a substantive reason behind the denial.  The guidance further explains that the Equal Credit Opportunity Act (ECOA)—a law that outlaws discrimination against credit applicants based on protected classifications—and its implementing regulation, Regulation B, prevents lenders and creditors from relying on a checklist of reasons to deny a credit application if the reasons do not “specifically and accurately” indicate the principal reason(s) for the adverse action.  These reasons, found in Regulation B’s definition of “data to be collected and reported,” are provided in the sample forms.  According to the CFPB, “creditors cannot state the reasons for adverse actions by pointing to a broad bucket.”  As one example, the CFPB explained that “if a creditor decides to lower the limit on a consumer’s credit line based on behavioral spending data, the explanation would likely need to provide more details about the specific negative behaviors that led to the reduction beyond a general reason like ‘purchasing history.’”  Doubling down, the CFPB further commented that “creditors that simply select the closest factors from the checklist of sample reasons are not in compliance with the law if those reasons do not sufficiently reflect the actual reason for the action taken.”

This guidance could pose a challenge for certain creditors because the sophisticated algorithms typically used by creditors—sometimes referred to as “black box” algorithms—may not reveal the substantive reason that was the basis for the denial.  Black-box algorithms, like other types of artificial intelligence, apply statistical transformations to convert input data into an actionable output (in this case, a credit denial).  However, there is no visibility into the determinative factors that led the algorithm to transform the input data into the output.  Thus, under a black-box credit algorithm, the credit provider only has visibility into the array of input factors which led to the decision: credit score, income, etc., but not necessarily the determinative factor.

One potential consideration for creditors who rely on black-box lending algorithms is to incorporate certain algorithm-interpretation techniques into their business practices.   Indeed, CFPB commissioner Rohit Chopra issued a statement in June 2023 in which he commented that “automated models can make bias harder to eradicate…because the algorithms used cloak the biased inputs and design in a false mantle of objectivity.  …  [I]nstitutions…have to take steps to boost confidence in valuation estimates and protect against data manipulation.”  Examples of such steps include Local Interpretable Model-agnostic Explanations (LIME) or Shapley Additive exPlanations (SHAP) values to convey the rationale behind credit denial decisions.  LIME/SHAP are interpretability-enhancing methods which harness statistics to increase the transparency of black-box AI models.  In effect, these methods can rank and isolate the most determinative factors in a credit decision, regardless of the original AI model.  Both the LIME and SHAP methods enhance interpretability and reveal the determinative factors that lead to a credit denial decision, with some limitations.  Both methods perturb individual variables between otherwise similar credit applications.  For example, say Applicant 1 and Applicant 2 submit identical applications, except that Applicant 1 has an income of $50,000 / year and Applicant 2 has an income of $65,000 / year.  Applicant 2 is accepted and Applicant 1 is rejected.  LIME will assign a probability (say, 45%) that income is responsible for the behavior of the underlying black box credit model.  Similarly, SHAP values will be assigned to the input factors and rank income among the potential factors by their determinative impact on the output.

In sum, by implementing interpretability-enhancing methods such as LIME and SHAP, creditors may be able to better identify and disclose to consumers the rationale behind black-box algorithms, and minimize their risk of an ECOA violation.