AI Loan Approval and Lending Discrimination Explained

When a lender approves, prices, or declines an application, a statistical model increasingly does the first pass, and sometimes the last. AI loan approval can widen access for borrowers whom old scorecards overlook. It can also reproduce old patterns of lending discrimination through variables that look neutral, with less visible reasoning than a loan officer's. This article explains how the models work, what peer-reviewed studies and regulators have found, and where the law stands in late 2026.

Educational only, not financial advice. This article describes research and regulation. It does not recommend any lender, product, or investment, and it is not legal advice. Rules in this area are changing, so check current sources for any specific situation.

Key takeaways

• Automated lending models estimate the chance of default from application and credit data, then convert that estimate into approve, decline, or price decisions.
• A model can be discriminatory without ever seeing race or sex, because other variables can stand in for them. This is called proxy discrimination.
• Studies disagree on whether algorithms help or hurt. One found machine learning widened interest-rate gaps for Black and Hispanic borrowers; another found fintech lenders may discriminate less in rejections.
• U.S. creditors must give applicants specific reasons for a denial, and the CFPB has said a complex model is no excuse.
• Enforcement of disparate-impact liability is in flux: an April 2025 executive order told agencies to deprioritize it, and the CFPB amended Regulation B provisions on it in April 2026.

How an automated lending decision is made

A credit model is a function that takes in facts about an applicant and returns a number, usually an estimated probability of default. The lender then picks a cutoff. Scores above the line are approved; scores below it are declined, priced higher, or sent to a human.

The facts can include the items on the application (income, requested amount, employment), the credit bureau file (payment history, balances, length of credit history), and sometimes other data such as bank-account cash flow. The model might be a simple logistic regression scorecard, which adds up weighted points. It might also be a gradient-boosted decision tree ensemble, a method the site's explainer on XGBoost covers, or a neural network.

The more flexible the model, the more patterns it can find. That flexibility is the source of both the promise and the problem, as the research below shows. For a plain-language look at the building block most of these ensembles use, see how decision trees split data.

Why lenders automate at all

The scale of automation is hard to measure from public sources, but one academic count is useful. In a study of mortgage discrimination, Robert Bartlett, Adair Morse, Richard Stanton, and Nancy Wallace examined the 2,098 largest mortgage lenders from 2012 to 2018. They found that by 2018, 45% of them offered a complete online or app-based application process. They also noted that loan origination was becoming almost exclusively algorithmic.

Automation matters for fairness in two opposite ways. It removes some of the discretion that lets an individual loan officer act on prejudice. It also moves decisions into code that few people read and that can apply the same flaw to every applicant at once.

A short history: why scoring was meant to reduce bias

Statistical credit scoring was not invented to hide bias. The New York State Department of Financial Services (DFS) writes in its 2021 report on the Apple Card that statistical scoring, introduced in the mid-twentieth century, "was intended in part to reduce or eliminate unacceptable human biases from credit determinations."

Before the federal Equal Credit Opportunity Act (ECOA) was passed in the 1970s, DFS notes, racial, ethnic, and sex-based bias was inherent and even expected in creditworthiness decisions. Credit managers routinely applied their "judgment" or "intuition," and women were typically not offered credit independently and had to bring a male co-signer. ECOA first prohibited reliance on sex or marital status, and later on race, nationality, religion, age, and receipt of public assistance.

The evidence that human lenders discriminated is old and strong. In 1992, researchers Alicia Munnell, Lynn Browne, James McEneaney, and Geoffrey Tootell at the Federal Reserve Bank of Boston circulated Working Paper 92-7, "Mortgage Lending in Boston: Interpreting HMDA Data," later published in the American Economic Review in March 1996. Minority applicants were initially denied 2.7 times as often as white applicants. After controlling for applicant and property characteristics, the ratio narrowed to about 1.6 to 1. The authors reported that applicants with the same economic and property characteristics would face a 17% denial rate if they were minorities, compared with 11% for whites.

That study is the benchmark every later fairness claim is measured against. The question is whether algorithms closed that gap, or found new ways to preserve it.

What "better" can mean: the case for AI underwriting

The strongest argument for machine-learning underwriting is that it can approve people that older models wrongly reject. A traditional scorecard relies on a thin set of variables. A borrower with little formal credit history can look riskier than they are.

The Consumer Financial Protection Bureau (CFPB) gave the clearest public data point. In a blog post dated August 6, 2019, the Bureau described results from the lender Upstart under a CFPB no-action letter. Upstart's tested model approved 27% more applicants than a traditional model, with 16% lower average APRs for approved loans.

The CFPB attached caveats that matter. The results came from Upstart's own simulations, and the Bureau did not replicate them. The no-action letter is tied to Upstart's specific facts and circumstances, and it does not endorse any particular variables, modeling techniques, or products. The results reflect the combined effect of alternative data and machine learning on one lender's applicant pool.

In 2015, the CFPB published an estimate of "credit invisible" consumers, people with no credit record at all. A June 2025 technical correction said that estimate should be roughly cut in half, with an almost commensurate increase in records that exist but are unscored. Either way, a meaningful group of people is hard for traditional scorecards to evaluate, and that is the opening alternative data tries to fill.

How a fair-looking model can still discriminate

U.S. law recognizes two broad forms of discrimination in lending. NY DFS describes them this way: disparate treatment is the intentional use of prohibited classifications such as sex or marital status. Disparate impact refers to facially neutral policies that nevertheless produce unequal outcomes for members of protected classes.

Bartlett and his coauthors explain the legal test in a footnote: disparate impact occurs when a decision maker's practice does not expressly discriminate on a protected characteristic but disadvantages one or more groups without a legitimate business justification. Under disparate impact theory, proof of discriminatory intent is not required to establish liability, and the Supreme Court has held that the theory extends to the Fair Housing Act.

Proxy discrimination, explained with a recipe

Imagine a recipe that never lists an ingredient, yet the dish tastes strongly of it. A lending model can behave the same way. It never receives race as an input, but it receives other inputs that are correlated with race, and it learns to use them.

Residential segregation makes this concrete. Where someone lives can predict default through local economic conditions. It can also predict race, because neighborhoods in the United States remain segregated. A model that leans on location data can therefore be accurate and discriminatory at once. The same can be true of variables such as the type of merchant a person pays, the device used to apply, or the type of income. Some of those are legitimate risk signals; others are close stand-ins for protected traits. Telling them apart is a legal and statistical judgment, not a technical one.

This is where the distinction between correlation and causation does practical work. The site's explainer on how correlation differs from causation shows why a variable that predicts default does not necessarily cause it, and why that matters when the variable also tracks a protected group.

What the research actually found

Several well-regarded studies bear on whether algorithms reduce or increase disparities. They look at different things, so they do not contradict each other as much as it first seems.

Study	Data and method	Finding	What it does not show
Munnell et al., Boston Fed, 1992 (published AER 1996)	Boston-area mortgage applications, human underwriting	Denial ratio fell from 2.7 to 1 to about 1.6 to 1 with controls, but a gap remained	Says nothing about algorithms
Bartlett, Morse, Stanton & Wallace, working paper 2019 (Journal of Financial Economics 2022)	U.S. conventional mortgages sold to Fannie Mae and Freddie Mac	Both face-to-face and fintech lenders charged Latinx and African-American borrowers 6 to 9 basis points more on purchase mortgages and 2 to 4 on refinances	Rate pricing, not every stage of the application decision
Fuster, Goldsmith-Pinkham, Ramadorai & Walther, Journal of Finance 2022	U.S. mortgage data, comparing traditional and machine-learning default models	Black and Hispanic borrowers disproportionately less likely to benefit from machine learning	A simple model of the credit market, not observed lender decisions
The Markup analysis of 2019 HMDA data, 2021	More than two million conventional home-purchase applications	After controlling for 17 factors, Black applicants were 80% more likely to be denied than similar white applicants	Excludes credit scores, which public data omit

The pricing study

The Bartlett team used a clever feature of the U.S. mortgage system. Fannie Mae and Freddie Mac charge lenders a guarantee fee that depends only on where a loan sits on a grid of credit score and loan-to-value ratio. Because the GSEs absorb the credit risk, differences in interest rates between borrowers within the same grid cell cannot reflect differences in credit risk. They must reflect lender pricing choices.

Using that setup, the authors found discrimination of 6 to 9 basis points in purchase-mortgage markets and 2 to 4 basis points in refinance markets, costing African-American and Latinx borrowers about $750 million a year in extra interest. They report that "FinTech algorithms have not removed discrimination," while noting two silver linings: algorithmic lending seems to have increased competition or encouraged more shopping, and there are reasons to believe fintechs may discriminate less in application rejection. The numbers here come from the February 2019 working paper version; the published version is Journal of Financial Economics volume 143, issue 1 (2022), and readers who need final estimates should check it.

The machine-learning study

Andreas Fuster, Paul Goldsmith-Pinkham, Tarun Ramadorai, and Ansgar Walther asked a different question. What happens when a lender swaps a traditional default model for a machine-learning model? Using U.S. mortgage data, they found that Black and Hispanic borrowers are disproportionately less likely to benefit from the change. In their model of the credit market, machine learning widened interest-rate disparities both between and within demographic groups. They attribute this mostly to the greater flexibility of the models, rather than to use of otherwise excluded characteristics.

That last point deserves emphasis. Removing race from the inputs did not fix the problem, because a flexible model can recover similar information from many other variables.

The denial-rate analysis

In 2021, the news organization The Markup analyzed 2019 data collected under the Home Mortgage Disclosure Act (HMDA). Holding 17 factors steady across more than two million conventional home-purchase applications, it reported that lenders were 80% more likely to deny Black applicants, 70% more likely to deny Native American applicants, 50% more likely to deny Asian and Pacific Islander applicants, and 40% more likely to deny Latino applicants than similar white applicants. The gaps appeared in 89 metro areas.

Industry groups, including the American Bankers Association and the Mortgage Bankers Association, argued the analysis was based on incomplete public data. The public HMDA file omits credit scores, and the analysis could not include them. The Markup's analysis was journalism, not peer-reviewed research, and it could not separate algorithmic decisions from human ones. It is best read as evidence that disparities persisted in automated-era mortgage lending, not as proof of what any specific algorithm did.

The same reporting noted that Fannie Mae and Freddie Mac required lenders to use the Classic FICO score, based on 1990s data, with a minimum of 620, and that critics said this disadvantages people of color because it does not count on-time rent or utility payments. That was the position as reported in 2021, and requirements in this area have been under review since, so the specific rule should be checked against current sources.

A case study: the Apple Card investigation

The most publicized algorithmic-lending controversy was the Apple Card. In November 2019, a technology entrepreneur posted on Twitter that he had been offered a credit limit 20 times higher than his wife's, despite filing joint tax returns and living in a community property state. He also complained that a customer service agent could not explain the algorithm. A co-founder of Apple then said he, too, had been offered far better terms than his wife despite jointly held assets.

NY DFS investigated. In its March 2021 report, the department analyzed underwriting data for nearly 400,000 New York applicants, used regression analysis, and reviewed several thousand pages of records. It found no evidence of intentional discrimination and no evidence of disparate impact in the credit decisions. The Bank was also able to identify the factors behind each complainant's decision, such as credit score, indebtedness, income, credit utilization, and missed payments.

DFS did find problems of a different kind. It cited deficiencies in customer service and transparency, and noted that a six-month waiting period before applicants could appeal credit limits frustrated customers. Goldman Sachs raised the credit limits of both women to match their spouses' within days of the complaints surfacing, and removed the waiting period.

The report also made a point that is easy to miss. Federal law requires lenders to explain only credit denials, not the amount or terms of credit offered. DFS noted that couples with shared finances often have different credit files, and that differences in credit history, such as one spouse holding a mortgage and multiple cards while the other holds one card, can produce different offers without any sex-based rule.

The lesson is not that algorithms are cleared or condemned. It is that a statistical audit on a large data set can answer the disparate-impact question, and that a lender who cannot explain terms to customers invites suspicion even when the numbers hold up.

Explaining a decision: adverse action notices

If a lender declines an application or offers less favorable terms, U.S. law requires a notice. Under the CFPB's Regulation B, which implements ECOA, the creditor must notify an applicant within 30 days of receiving a completed application. The notice must be in writing and include either a statement of specific reasons or a disclosure of the right to request them.

The reasons have to be specific. The regulation's official commentary treats vague explanations as insufficient, such as saying the decision was based on internal standards or that the applicant failed to reach a qualifying score. The commentary does not require a specific number of reasons, but says that disclosing more than four is not likely to be helpful to the applicant.

Can a lender hide behind the model?

No, the CFPB said in Circular 2022-03, announced on May 26, 2022. Creditors must provide specific reasons for denials even when they use complex credit models. CFPB Director Rohit Chopra said companies "are not absolved of their legal responsibilities when they let a black-box model make lending decisions." The circular states that there is no exception for violating the law because a creditor is using technology that has not been adequately designed, tested, or understood, and warned that some model outputs may be unexplainable in a way that makes compliant notices impossible. A follow-up circular in September 2023 addressed the use of the sample checklist reasons in Regulation B; the CFPB's page for it is now marked as archived.

How explainability tools help

Explainability techniques try to attribute a prediction to input features. SHAP, introduced by Scott Lundberg and Su-In Lee in a 2017 paper, defines a class of additive feature importance measures and shows that one member of it has a unique set of desirable properties. In lending, such tools can turn a complex model's output into a ranked list of reasons.

Explanations are not the same as fairness. A model can give perfectly clear reasons for a decision that rests on a discriminatory proxy. The site's article on mechanistic interpretability describes the harder problem of understanding what a model has actually learned, which goes beyond ranking input features.

Why "fair" has no single definition

Even with perfect intentions, fairness in prediction involves trade-offs that are mathematical rather than moral. Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan defined three fairness conditions central to public debates about algorithmic classification and proved that, outside of very narrow special cases, no method can satisfy all three at once. Moritz Hardt, Eric Price, and Nathan Srebro proposed a different criterion, equality of opportunity, and illustrated it with a case study of FICO credit scores.

A hypothetical example shows the tension. These numbers are constructed for illustration and are not data from any lender.

	Group A	Group B
Applicants	1,000	1,000
Who would actually default	100 (10%)	200 (20%)
Flagged "high risk" by the model	143	429
Default rate among those flagged	40%	40%
Default rate among those not flagged	5%	5%
Share of non-defaulters wrongly flagged	About 9.5%	About 32%

The score is perfectly calibrated: a flag means a 40% default rate in both groups, and no flag means 5%. Yet because group B has a higher underlying default rate, far more of its reliable borrowers get flagged. Equalizing the false-flag rates would require breaking the calibration, and vice versa. That is the type of conflict Kleinberg and his coauthors proved.

The practical upshot is that lenders and regulators must choose which definition of fairness to prioritize, and that choice is partly a legal and policy question. It also explains why a single fairness metric cannot certify a model. The site's explainer on precision and recall covers the related error trade-offs in classification.

Where U.S. regulation stands in late 2026

The regulatory picture changed sharply in 2025 and 2026, and it is still unsettled.

Date	Development	What it says
May 26, 2022	CFPB Circular 2022-03	Complex models do not excuse the requirement to give specific denial reasons
September 19, 2023	CFPB Circular 2023-03	Addresses adverse action notices and the sample forms in Regulation B; page now archived
April 23, 2025	Executive Order 14281	States a policy of eliminating disparate-impact liability "in all contexts to the maximum degree possible"; directs agencies to deprioritize enforcement of statutes and regulations to the extent they include disparate-impact liability
April 22, 2026	CFPB final rule amending Regulation B	Amends provisions related to disparate impact, discouragement of applicants, and special purpose credit programs; effective July 21, 2026

Sources: CFPB; White House.

Two points follow. First, the duty to give specific reasons for a denial sits in the text of Regulation B and was not described as changed by the sources reviewed for this article. Second, the status of disparate-impact liability under ECOA is in motion. The CFPB's page for the 2026 rule states that it amends disparate-impact provisions but does not detail the changes, so this article does not characterize them. Executive Order 14281 also told agencies to review pending investigations and existing consent judgments that rely on disparate-impact theories. Readers should consult the Federal Register notice and current legal commentary for the details.

State laws, private lawsuits, and the Fair Housing Act operate separately from federal agency priorities. How these interact with the new federal approach is not settled in the sources used here.

What responsible model design looks like

Researchers and practitioners describe a set of practices that reduce risk. This is a description of the field, not a compliance checklist.

Test outcomes by group. Compare approval rates, pricing, and error rates across demographic groups, and investigate gaps that cannot be explained by legitimate underwriting factors. The core idea is the same one the Bartlett team used: compare similar borrowers and ask why outcomes differ.

Review the features. Examine variables such as location, merchant type, and device for their role as proxies, and consider whether a less discriminatory alternative would do the same predictive job.

Keep humans in the loop for close calls. A hybrid design can let a model handle clear approvals and clear declines while routing borderline cases to a person who can be held accountable.

Monitor after launch. Populations and economies change, and a model that was balanced at launch can drift. The site's piece on what model drift means in production explains how to detect it.

Document the choice of fairness metric. Given the trade-offs above, it matters which definition a lender chose and why.

Common misconceptions

"If the model never sees race, it cannot be biased." The Fuster team's result runs the other way: the widened gaps came mainly from model flexibility, not from use of excluded characteristics. Bias can enter through correlated variables and through the data the model learned from.

"Algorithms are objective." A model inherits whatever patterns are in its training data, including those produced by past human decisions. The site's article on cognitive biases covers the human side of that inheritance, and its overview of AI ethics debates covers the wider argument.

"Human underwriters are fairer." The Boston Fed study and decades of fair-lending enforcement show human lenders can discriminate too. The Bartlett study found that face-to-face and fintech lenders both charged more, and suggested fintechs might discriminate less in rejections.

"More approvals means a fairer system." Approval rates can rise for everyone while the gap between groups holds or widens. Upstart's reported 27% increase in approvals is not a statement about group disparities.

"An explanation proves a decision was fair." Explanations show the factors behind a decision. They do not show whether those factors are proxies.

How this connects to the wider credit system

Lending models do not operate in isolation. They sit in a system where households carry debt, wealth is unevenly distributed, and credit access shapes who can buy a home. For international context on household leverage, see the site's ranking of countries with the highest household debt, and for the distributional backdrop, its explainer on wealth inequality.

The same modeling techniques appear elsewhere in finance. The site covers machine learning credit scoring, AI in insurance underwriting and pricing, and reinforcement learning for portfolio optimization, and each raises its own versions of the accuracy, transparency, and fairness questions discussed here.

The takeaway

AI loan approval is neither a fix for lending discrimination nor a new cause of it. The best-documented finding is that disparities in mortgage lending persist under both human and algorithmic decision making, and that making models more flexible can widen gaps for some groups even when race is not an input. The best-documented protection is procedural: lenders must give specific reasons for denials, and a complex model is not an excuse. Whether disparate-impact liability will remain a strong enforcement tool is now an open question in the United States, which makes the lender's own testing, and the public's access to data like HMDA, more important rather than less.

FAQ

Q: How does AI loan approval differ from traditional underwriting? A: Traditional scorecards use a limited set of variables with fixed weights. AI models, including gradient-boosted trees and neural networks, can use many more variables and capture non-linear patterns. That can help borrowers with thin credit files, as in Upstart's reported 27% increase in approvals, but it also makes decisions harder to explain.

Q: What is an adverse action notice? A: It is the notice a creditor must send when it denies credit or offers less favorable terms. Under Regulation B, it is due within 30 days of a completed application and must include specific reasons or a disclosure of the right to request them. Vague reasons such as "internal standards" are not enough.

Q: Can a lender defend a denial by saying the algorithm decided? A: Not under the CFPB's May 2022 Circular 2022-03. CFPB Director Rohit Chopra said companies are not absolved of their legal responsibilities when they use a black-box model, and the circular says there is no exception for technology that has not been adequately designed, tested, or understood. Some outputs may be unexplainable, which could make compliant notices impossible.

Q: Do AI systems treat borrowers more fairly than human loan officers? A: The evidence is mixed. A 2019 study found face-to-face and fintech mortgage lenders both charged Latinx and African-American borrowers more, with some reason to think fintechs discriminate less in rejections. A 2022 study found machine learning widened rate gaps for Black and Hispanic borrowers. Results depend on design, data, and testing.

Q: What is proxy discrimination? A: It occurs when a model uses variables correlated with a protected characteristic, such as race, to produce outcomes that disadvantage that group, even though the characteristic is not an input. Under disparate-impact theory, intent is not required to establish liability, though the legal status of that theory is being revised in 2025 and 2026.

Q: What did the Apple Card investigation conclude? A: New York's Department of Financial Services reviewed data on nearly 400,000 New York applicants in its March 2021 report. It found no evidence of intentional discrimination or disparate impact, but it did find deficiencies in customer service and transparency.

Sources

• Consumer Financial Protection Bureau, CFPB acts to protect the public from black-box credit models using complex algorithms (Circular 2022-03, May 26, 2022).
• Consumer Financial Protection Bureau, Update on credit access and the Bureau's first No-Action Letter (August 6, 2019).
• Consumer Financial Protection Bureau, 12 CFR 1002.9, Notifications.
• Consumer Financial Protection Bureau, Equal Credit Opportunity Act (Regulation B) final rule (April 22, 2026).
• Consumer Financial Protection Bureau, Technical correction and update to the CFPB's credit invisibles estimate (June 2025).
• The White House, Restoring Equality of Opportunity and Meritocracy (Executive Order 14281, April 23, 2025).
• New York State Department of Financial Services, Report on Apple Card Investigation (March 2021).
• Bartlett, Morse, Stanton & Wallace, Consumer-Lending Discrimination in the FinTech Era (working paper, February 2019); Journal of Financial Economics 143(1), 2022, DOI 10.1016/j.jfineco.2021.05.047.
• Fuster, Goldsmith-Pinkham, Ramadorai & Walther, "Predictably Unequal? The Effects of Machine Learning on Credit Markets," Journal of Finance 77(1), 2022, DOI 10.1111/jofi.13090.
• Munnell, Browne, McEneaney & Tootell, Mortgage Lending in Boston: Interpreting HMDA Data, Federal Reserve Bank of Boston Working Paper 92-7 (1992); American Economic Review 86(1), 1996.
• The Markup, The Secret Bias Hidden in Mortgage-Approval Algorithms (August 25, 2021).
• Lundberg & Lee, A Unified Approach to Interpreting Model Predictions (2017); Hardt, Price & Srebro, Equality of Opportunity in Supervised Learning (2016); Kleinberg, Mullainathan & Raghavan, Inherent Trade-Offs in the Fair Determination of Risk Scores (2016).

---

Read next:

• Robo-Advisors vs. Human Financial Advisors: Who Wins in the Age of AI?
• Overfitting and Underfitting: The Bias-Variance Tradeoff
• Why Sample Size Changes Confidence