- AI-based lending decisions are under scrutiny in the finance industry.
- Lenders must understand bias testing and human feedback loops to hold AI platforms accountable.
- This article is part of the “Innovation at Work” series exploring the trends and barriers to workplace transformation.
The financial sector has a long history of making inequitable loan decisions.
Redlining, a discriminatory practice that started in the 1930s, is when a bank denies a customer a loan because of their ZIP code. These institutions physically drew a red line around low-income neighborhoods, segregating these residents from any opportunity to borrow money.
Redlining disproportionately affects Black Americans and immigrant communities. This denies them opportunities like homeownership, starting a small business, and earning a postsecondary education.
The rise of machine learning and big data means decisions can be controlled for human bias. But just adopting the tech isn’t enough to overhaul discriminatory loan decisions.
A 2019 analysis of US Home Mortgage Disclosure Act data by The Markup, a nonprofit dedicated to data-driven journalism, found lenders nationwide were nearly twice as likely to deny Black applicants as they were to reject similarly qualified white applicants despite adopting machine-learning and big-data tech. Latinos, Asians, and Native Americans were also denied mortgages at higher rates than white Americans with the same financial background.
Governments around the world have indicated there will be a crackdown on “digital redlining,” where algorithms discriminate against marginalized groups.
Rohit Chopra, the head of the US’s Consumer Financial Protection Bureau, said there should be harsher penalties for these biases: “Lending algorithms can reinforce bias,” he told The Philadelphia Inquirer. “There’s discrimination baked into the computer code.”
Meanwhile, politicians in the European Union plan to introduce the Artificial Intelligence Act for stricter rules around the use of AI in filtering everything from job and university applicants to loan candidates.
Bringing bias to light
It’s easy to blame technology for discriminatory lending practices, Sian Townson, a director at Oliver Wyman’s digital practice, told Insider. But it doesn’t deserve the responsibility.
“Recent discussions have made it sound like AI invented bias in lending,” she said. “But all the computational modeling has done is quantify the bias and make us more aware of it.”
While identifiers like race, sex, religion, and marital status are forbidden to be considered in credit-score calculations, algorithms can put groups of people at a disadvantage.
For instance, some applicants may have shorter credit histories because of their religious beliefs. For example, in Islam, paying interest is seen as a sin. This can be a mark against Muslims, even though other factors may indicate they would be good borrowers.
While other data points, like mobile payments, are not a traditional form of credit history, Townson said, they can show a pattern of regular payments. “The aim of AI was never to repeat history. It was to make useful predictions about the future,” she added.
Testing and correcting for bias
Software developers like the US’s FairPlay — which recently raised $10 million in Series A funding — have products that detect and help reduce algorithmic bias for people of color, women, and other historically disadvantaged groups.
FairPlay’s customers include the financial institution Figure Technologies in San Francisco, the online-personal-loan provider Happy Money, and Octane Lending.
One of its application-programming-interface products, Second Look, reevaluates declined loan applicants for discrimination. It pulls data from the US census and the Consumer Financial Protection Bureau to help recognize borrowers in protected classes, given financial institutions are forbidden to collect information directly about race, age, and gender.
Rajesh Iyer, the global head of AI and machine learning for financial services at Capgemini USA, said lenders could minimize discrimination by putting their AI solutions through about 23 bias tests. This can be done internally or by a third-party company.
One bias test analyzes for “disproportionate impact.” This detects whether a group of consumers is being more adversely affected by AI than other groups — and, more importantly, why.
Fannie Mae and Freddie Mac, which back the majority of mortgages in the US, recently found people of color were more likely to list their source of income from the “gig economy.” This disproportionately stopped them from getting mortgages because gig incomes are seen as unstable, even if someone has a strong rent-payment history.
In looking to make its lending decisions fairer, Fannie Mae announced it would start factoring rental histories into credit-evaluation decisions. By inputting new data, humans essentially teach the AI to eliminate bias.
Human feedback to keep AI learning
AI can learn only from the data it receives. This makes a feedback loop with human input important for AI lending platforms, as it enables institutions to make more equitable loan decisions.
While it’s good practice for humans to weigh in when decisions are too close to call for machines, it’s essential for people to review a proportion of clear-cut decisions, too, Iyer told Insider.
“This ensures that the solutions adjust themselves, as it gets inputs from the human reviews through incremental or reinforced learning,” Iyer said.