There is no “one-size-fits-all” approach to making AI-driven decisioning by financial institutions explainable and fair, but lenders want regulators to clarify how to measure the technology’s disparate impact and explainability.

Bank Automation News has been following the responses of lenders and vendors to an April request for information seeking input on AI use cases in financial services. The request was issued by the Federal Reserve and the Office of the Comptroller of the Currency, the Federal Deposit Insurance Corporation, Consumer Financial Protection Bureau and the National Credit Union Administration.

Quicken Loans

Quicken Loans said in a recent response to a federal request for public input that regulators should, “articulate expectations and requirements for measuring, monitoring, and mitigating fair lending risk when using AI/ML technologies.”

Early this year, Quicken created “Rocket Ethical Framework,” an internal framework that examines financial models through three lenses: data, for bias and security; methodology, for compliance with regulation; and usage, to create an audit trail. “It would also be helpful for the agencies to offer guidance on requirements regarding documentation, reproducibility, appropriate monitoring levels and explainability,” Quicken Loans noted.

MX

Meanwhile, financial data provider MX said in its response that major barriers to explainable AI include, a lack of industry-wide standards on measuring explainability and the “black box” nature of some models built for financial institutions by third-party vendors.

A lack of oversight on financial data usage also presents a challenge and the “vast volume of data generated and consumed by financial institutions, whether consumer permissioned or not, will be a huge challenge for explainability moving forward,” the MX response noted.

State Street

Custodian bank State Street enumerated the wide variety of use cases for which the $3.6 trillion financial institution uses AI models, including transaction processing, fund compliance, risk management and measuring the quality of net asset value calculations. In its response the bank echoed MX’s point regarding, the lack of explainability for models provided by third-party vendors, saying it is especially challenging due to the lack of access to underlying data.

State Street said its own AI -deployment is largely concentrated on operational efficiency as opposed to delegating decision making. Regulated financial institutions would greatly benefit from agency guidance on expectations for use, management, and oversight of third-party AI models.

TransUnion

Credit reporting agency TransUnion said in its response that AI and ML models tend to provide a better and more predictive separation of risk, which in turn can expand the pool of credit- worthy consumers. However, the absence of regulatory guidance could also “discourage firms from investing in AI and ML at scale.”

Accenture

In order to guard AI models from cyberattacks, data used to train models must be protected, “as attacks on training data could result in corrupt models from the outset,” said consulting firm Accenture in its response. It added that “financial institutions have reasonable evidence that suggests some discrimination may be occurring as an unintended consequence of using AI models,” further reinforcing the need to deploy effective detection and correction of bias along with fairness testing.

Most respondents to the federal request for information have urged agencies to take a flexible approach to regulating AI and to avoid prescriptive rulemaking, an approach echoed by industry trade group Bank Policy Institute. Arguing along similar lines, the Internet Association said in its response that regulators should adopt explainability requirements that account for factors like the severity of the use case, the importance of the goods or services involved and the possible impact on the affected person.

The Internet Association counts major tech firms like Google, Microsoft, PayPal, and Amazon among its members. Its filing added that when considering regulations surround AI, policymakers need to look at, “whether complex models that will improve society and people’s outcomes should be avoided because they cannot be fully explained.”