AI credit scoring bias: how algorithms decide who gets credit

AI credit scoring can repeat past bias when models learn from historical lending data or use proxies for race. Fuster and colleagues (2022) found Black and Hispanic mortgage borrowers were less likely to gain from machine learning models. Stanford HAI (2021) reported credit scores about 5 percent less accurate for minority borrowers across 50 million people.

EasWrk is a hosting platform, not a research institute. This essay summarises published sources, and the sources are listed at the end. More essays are on the EasWrk Research index.

AI credit scoring models and how they repeat past bias

Lenders use machine learning to predict who will repay a loan. Whether those models treat borrowers more fairly than older scorecards is still an open research question. The studies below give mixed answers.

Fuster, Goldsmith-Pinkham, Ramadorai and Walther (2022) compared traditional and machine learning default models on US mortgage data. They found that Black and Hispanic borrowers were disproportionately less likely to gain from the introduction of machine learning. In their simple model of the credit market, machine learning increased disparity in rates between and within groups. They attribute most of that change to the greater flexibility of the models.

Proxy variables in credit scoring

Fuster and colleagues (2022) describe two ways better statistics can change who wins and loses. One is greater flexibility in uncovering structural relationships. The other is triangulation of otherwise excluded characteristics. The second route is the proxy problem. A lender can remove race from a model and still let the model rebuild it from other inputs. A postcode is a common example, because where people live can track race. The healthcare essay covers the same proxy problem in medicine.

O'Neil (2016) calls opaque, widely used models that harm people "weapons of math destruction". Credit scoring is one of her examples.

AI lending bias against minority borrowers

Blattner and Nelson (2021) studied credit data on 50 million people supplied by one credit-score company. Stanford HAI (2021) reported the result: scores were about 5 percent less accurate at predicting default for minority borrowers than for non-minority borrowers, and about 10 percent less predictive for the bottom fifth of earners.

The researchers did not blame the algorithms. They said the data is less informative for these borrowers, often because they have thin credit files, and that one or two blemishes on a thin file can distort a score. Models tuned for minority and low-income borrowers performed about the same. They estimated that more complete credit files could close about half of the accuracy gap (Stanford HAI, 2021).

Bartlett, Morse, Stanton and Wallace (2019) studied US mortgage pricing. They found that FinTech algorithms also discriminate, but about 40 percent less than face-to-face lenders. They estimated that lenders charge Latinx and African-American borrowers 7.9 basis points more on purchase mortgages and 3.6 basis points more on refinance mortgages, about $765 million a year in extra interest. They also reported that FinTech lenders do not discriminate in loan approval.

Thin files matter here, so the size of the thin-file population matters. The Consumer Financial Protection Bureau (CFPB) estimated in 2015 that 25.9 million US adults, 11.0 percent, were credit invisible in December 2010. A June 2025 CFPB correction cut that figure to 13.5 million, or 5.8 percent. The same correction put the December 2020 figure at 7.0 million, or 2.7 percent (CFPB, 2025).

Algorithmic credit discrimination and the rules lenders face

In the United States, lenders send adverse action notices to rejected applicants under Regulation B, which implements the Equal Credit Opportunity Act (ECOA). The notice has to state the principal reasons for the decision.

The CFPB tested that duty against complex models. Circular 2023-03, released on 19 September 2023, asked whether creditors using artificial intelligence or complex credit models may rely on the checklist of reasons in the CFPB's sample forms when those reasons do not accurately identify the real reasons. The answer was no (CFPB, 2023). An earlier circular, 2022-03, covered adverse action notices for credit decisions based on complex algorithms.

In May 2025 the CFPB withdrew 67 guidance documents, and the list includes both circulars (Federal Register, 12 May 2025). The withdrawal covers the circulars. It does not amend Regulation B itself.

In the European Union, the AI Act (Regulation (EU) 2024/1689) lists AI systems used to evaluate the creditworthiness of natural persons or establish their credit score as high-risk, with an exception for fraud detection (Annex III, point 5(b)). The Act's recitals say such systems may perpetuate historical patterns of discrimination based on racial or ethnic origins, gender, disabilities, age or sexual orientation. A 2026 amending regulation, Regulation (EU) 2026/1744, set 2 December 2027 as the date from which the high-risk rules apply to Annex III systems.

Courts face a related question with risk scores in criminal justice. The legal essay on algorithmic risk scores covers it.

How finance came to run on algorithms

Finance took up algorithmic methods early. A 2020 survey of central banks found that most already discuss big data formally, and that applications span research, monetary policy, financial stability and supervision (Doerr, Gambacorta and Serena Garralda, 2021). Pasquale (2015) wrote about the secret algorithms that control money and information. This essay starts with credit because that is where the effects land on individual people. It then covers trading and systemic risk.

High-frequency trading and market fragility

The SEC's 2010 concept release on equity market structure said that estimates of high-frequency trading volume in US equities "typically are 50% of total volume or higher" (SEC concept release, 2010).

Hendershott, Jones and Menkveld (2011) used the NYSE's 2003 introduction of automated quote dissemination to estimate the effect of algorithmic trading on liquidity. For large stocks, they found that algorithmic trading narrowed spreads, reduced adverse selection and reduced trade-related price discovery. Liquidity in calm markets is one question. Liquidity under stress is another, and the 2010 flash crash is the best documented test.

What the 2010 flash crash showed about opaque trading systems

At 2:32 p.m. on 6 May 2010, a mutual fund complex started a sell program of 75,000 E-Mini S&P 500 futures contracts, valued at about $4.1 billion. The staffs of the CFTC and SEC reported that markets were already falling. By 2:30 p.m., buy-side liquidity in the E-Mini had dropped 55 percent, from nearly $6 billion in the early morning to $2.65 billion. Between 2:41 p.m. and 2:45:28 p.m., broad markets fell a further 5 to 6 percent, to intraday lows 9 to 10 percent down (CFTC and SEC staffs, 2010).

The report described "hot potato" trading. Between 2:45:13 and 2:45:27, high-frequency traders traded over 27,000 contracts, about 49 percent of total volume, while the same positions passed back and forth. Many individual securities and exchange-traded funds then traded at prices as low as one penny or as high as $100,000. Accenture fell from nearly $40 to one cent and recovered within seconds (CFTC and SEC staffs, 2010).

Kirilenko, Kyle, Samadi and Tuzun (2017) analysed audit-trail data for the E-Mini on 6 May and the three days before. Their abstract reports that the trading pattern of the most active intermediaries, classified as high-frequency traders, did not change when prices fell. In this essay's reading, the staff report and the paper describe a fragile market, not a single guilty party.

A single firm can also fail. On 1 August 2012, Knight Capital lost more than $460 million after code deployed in late July triggered a defective function in an automated order router. The SEC said the router sent more than 4 million orders in the first 45 minutes after the open to fill 212 customer orders. Knight agreed to pay $12 million to settle the SEC's charges in October 2013 (SEC, 2013).

Robo-advisers and retail investor risk

In February 2017 the SEC's Division of Investment Management published guidance on robo-advisers. It named three areas: disclosures to clients, the duty to obtain information from clients to support suitable advice, and compliance programs designed for automated advice. It suggested that a robo-adviser say that it uses an algorithm, describe what the algorithm does, and describe its assumptions and limitations. It noted that many robo-advisers rely solely on questionnaires to collect client information. The staff called its content suggestions and said that not every issue applies to every adviser (SEC, 2017).

Systemic risk and correlated AI behaviour

The Financial Stability Board (2024) identified four AI-related vulnerabilities that could raise systemic risk: third-party dependencies and service provider concentration, market correlations, cyber risks, and model risk, data quality and governance. It said existing frameworks cover many of these risks, but that more work may be needed to confirm they are sufficient.

Credit models show a small version of the same risk. Lenders that train on similar data can make similar errors, and the Blattner and Nelson result above suggests that shared data gaps produce shared mistakes. That is this essay's reading of the evidence. The FSB does not make that claim.