AI Credit Models Risk Automating Past Discrimination, Researchers Warn

Traditional credit scores are backward-looking and disadvantage “thin-file” borrowers, prompting lenders and start-ups to test AI models that analyse thousands of data points — though regulators and researchers warn those same models risk automating past discrimination without new safeguards.

Sophie Hope discovered the limits of the current system when a mortgage application was denied after her credit report caused her score to plummet.

“When I first checked my credit score, it was excellent. But then my mortgage application was suddenly denied by the lender because of my credit report. It was a massive shock as I’d never had bad credit. When I looked into it, my score had gone to ‘very poor’,” she said.

Her case illustrates a broader structural problem. Three firms — Experian, Equifax and TransUnion — produce credit scores in the UK using largely the same historical data. “A credit score is backward looking. It’s looking at what has happened in the past, which doesn’t necessarily predict the future,” said Funmi Olufunwa.

“But when somebody’s looking to lend you money, they’re looking at what actually is affordable.” Traditional scoring also disadvantages “thin-file applicants” — young people and recent immigrants to the UK who won’t have built up a credit history.

AI Models Offer Alternative to Traditional Scoring

In the United States, lending decisions often hinge on a credit score from FICO, which uses data from the big three credit assessment firms. AI-powered lending firms Upstart and Zest AI have both built businesses offering lenders an extra option when it comes to assessing a client’s credit worthiness.

AI Credit Models Risk Automating Past Discrimination, Researchers Warn
Photo: finwire.io

“Thanks to our machine-learning models that look at more than 2,500 variables, we believe we can make significantly more loans at low prices to borrowers from every demographic group,” Upstart said in a statement.

Sarah Tucker says: “AI has the potential to help lenders analyse much broader sets of information, identify patterns more quickly and build a more individual picture of someone’s affordability and risk.” She believes there is “a real opportunity to improve financial inclusion” if technology allows lenders to understand so-called thin-file applicants better, rather than simply applying increasingly rigid rules.

Bias and Accountability Concerns

She stresses that firms must apply existing regulations and establish new guardrails to ensure that AI does not simply perpetuate the biases of the past.

“Not putting any safeguards in place will fortify, foster, and enhance discrimination.” Upstart said in a statement: “We’ve invested in creating objective, automated evaluations to identify bias, and we’ve incorporated those evaluations into our model training process so it evolves to become even more accurate and more fair over time.”

Evangeline Atkinson had a frustrating experience over her credit score. “I was at university, paying rent, paying bills, as I had been doing for many years. I tried to get a £30-a-month phone contract, and they said no,” she recalls.

“When I dug into it, it was because I was completely invisible to the system. No record could be found for me at any of the big three credit reference agencies.”

That experience inspired her to try to change the system. In 2021, Atkinson founded a credit reference agency with her best-friend from school, who had also experienced difficulties getting credit. “It’s very sensitive data and it would be very easy for the models to be inferring various kind of protected characteristics.” Their firm, Noggin HQ, was authorised by the FCA in June.

Incumbents Adopt AI Tools

Perhaps conscious of change, the big three credit scoring firms have all started experimenting with AI. Experian announced a partnership with OpenAI in August 2026, letting users check their credit score inside ChatGPT.

Upstart, founded in 2012 and based in San Mateo, California, first tried to fund borrowers through income-share agreements before pivoting to personal loans in 2014. Its models weigh education and work history alongside traditional credit data, and the company went public in late 2020. Over 70% of its loans are now funded with no human interaction.

“I wonder if older cohorts, who may not use technology as much, will be quite as comfortable with it,” says Olufunwa.