Will using AI to trick credit into trading one type of bias for another?
The terms of the next loan you get may depend less on your creditworthiness and more on what a computer program thinks of your habits.
Digital lending is expected to double over the next three years, reaching nearly 10 percent of all lending in the US and Europe. There are now around 2,000 digital startups, many of which use artificial intelligence to analyze the data that is generated every day.
But government regulators are struggling to keep up with the pace of change.
Marc Stein who runs Underwrite.AI, writes algorithms that can teach themselves. This is known as machine learning, a type of artificial intelligence.
“When you shop on Amazon, they use some form of machine learning. This is how Amazon’s recommendation system works, ”he says.
The program learns from any correlation it finds, whether it’s determining a person’s favorite books or lying about their income on a loan application. And based on this information, it can predict whether the applicant is a good risk.
“When we looked at the delta between what people said, what they did, and what we could verify, it was very predictable,” says Stein.
Digital lenders pull in all kinds of data including purchases, SAT results, and public records like fishing licenses.
As part of the loan application process, some lenders have potential borrowers download an app that uploads an extraordinary amount of information, such as: For example, daily location patterns, the punctuation of text messages, or how many of your contacts have last names. This is called in the credit industry alternative dates, and it is mainly used to make decisions about short-term, high-yielding loans. But that will likely change soon.
“In 10 years there will be hardly a credit decision that does not have a hint of machine learning behind it,” says Dave Girouard, CEO of upstart, an online lender.
Armed with self-learning algorithms and new data, Upstart is trying to sell its technology to traditional banks.
“FICO and income, which is the sweet spot, so to speak, of what every consumer lender in the United States uses, are actually pretty biased towards the people themselves,” says Girouard.
Government research has found that FICO scores hurt younger borrowers and overseas borrowers because low-income people seek loans with higher interest rates. Girouard argues that new, smarter data can make lending fairer.
“The variables we introduce into this model actually reduce the biases inherent in most credit systems,” he says. “And that’s just one way of expanding access to credit.”
When artificial intelligence can tell good borrowers from bad ones by looking at things like browsing history, suddenly it doesn’t matter whether you live in a low-income neighborhood or your family has just immigrated. But it opens the door to new redlining versions of the 21st century.
Jo Ann Barefoot, a consultant and former regulator, provides this example:
“The question of whether the person pays for petrol at the pump or goes in,” she says. “The theory behind this is that people who pay indoors are more likely to be smokers and that smoking is said to be more closely related to poor creditworthiness.”
But what if you only buy M & Ms or prefer to use cash? An algorithm doesn’t know that. If left unchecked, computers can create all kinds of unintended distortion.
Some companies have linked nightly internet use to bad debt repayment. But should night owls pay higher interest rates?
There’s no evidence that this type of injustice is taking place, but Barefoot says regulators are starting to pay attention to it.
“And of course one of the tricky problems is that these novel models have not yet been tested by an economic downturn,” she says.
Barefoot fears that failing online loans in a recession could spread to the financial system. Ultimately, however, she calls digital lending “massively good”. Regulators seem to agree and are only now asking questions.