The promise of AI in lending is that it can see more, decide faster, and remove the human bias that has historically made credit harder to get for certain groups. The concern is that it can also encode new forms of discrimination — at scale, with less transparency, and with almost no way to appeal. The truth is it's doing both at once, and understanding the difference matters whether you're a borrower, a business owner, or just someone who cares about how financial systems work.
How Traditional Credit Decisions Were Made
To understand what AI is changing, it helps to know what it's replacing. For decades, most credit decisions in the US relied primarily on FICO scores — a three-digit number calculated from five factors: payment history, amounts owed, length of credit history, new credit, and credit mix. Lenders also reviewed income, debt-to-income ratio, and employment history.
This system was consistent and auditable — you could look at someone's FICO score and understand in broad strokes why they had it. But it had well-documented shortcomings. It excluded large segments of the population who had limited credit history: young people, recent immigrants, people who had avoided debt for cultural or personal reasons. And it carried historical bias baked in from decades of redlining and discriminatory lending practices that had left entire communities with less access to credit-building tools in the first place.
The average FICO score varies significantly by race and income — not because those are factors in the calculation, but because wealth and credit history are correlated with them in ways the historical financial system helped create. Traditional credit scoring wasn't actively discriminatory, but it preserved the downstream effects of systems that were.
What AI-Powered Credit Decisioning Does Differently
AI credit models can evaluate far more data points than a traditional FICO-based approach. Depending on the lender and jurisdiction, these systems may analyze bank account transaction patterns, rent and utility payment history, income volatility, spending behavior, employment data, and in some cases behavioral data like how long someone spent reading a loan agreement.
The appeal of this broader data set is real. Someone with a thin credit file — perhaps a recent graduate or a newcomer to the US — might be a reliable borrower whose creditworthiness is invisible to a traditional model. An AI system that can see their consistent rent payments, stable income deposits, and responsible spending patterns can make a more accurate risk assessment than one limited to formal credit tradelines.
Several fintech lenders have built their entire value proposition on exactly this. Upstart, for example, uses a model that incorporates education and employment history in addition to traditional credit factors, and claims its model approves significantly more borrowers at the same default rate compared to traditional models — with particular gains for thin-file applicants. Petal, another fintech credit card issuer, offers cards to applicants with no credit history by analyzing bank account data directly. These are genuine improvements in access for people traditional systems underserved.
Where the Fairness Problems Enter
Here's where the picture gets complicated. Analyzing more data doesn't automatically produce fairer outcomes — it depends entirely on what data is used, what the model is trained to predict, and whether the patterns it learns reflect actual creditworthiness or proxies for protected characteristics.
The risk is what researchers call proxy discrimination. Even when race, gender, or religion are explicitly excluded from a credit model, the model can learn to use correlated variables as substitutes. If certain zip codes, spending patterns at particular types of stores, or even the timing of financial transactions correlate with race or ethnicity — which many do, because economic segregation is real — a model trained on historical data can effectively discriminate without ever seeing a demographic variable. The model isn't programmed to discriminate. It learns to, from data that reflects a society that did.
This has been documented in practice. A 2019 study by researchers at UC Berkeley found that algorithmic mortgage lenders charged higher interest rates to Black and Latino borrowers compared to white borrowers with similar financial profiles — in some cases by as much as 5–9 basis points on average. The gap was smaller than with human loan officers, but it existed. The algorithm had absorbed something from the historical data it was trained on.
There's also the transparency problem. Traditional FICO-based decisions are explainable: a lender can tell you which factors hurt your score and by how much. Many AI models are essentially black boxes — they produce an output without any straightforward way to audit the logic. When a borrower is denied credit, they may receive a generic adverse action notice that doesn't illuminate what actually drove the decision. That opacity makes it harder for borrowers to challenge incorrect decisions and harder for regulators to identify discriminatory patterns.
What Regulators Are Doing About It
The regulatory response has been developing, though it's moved more slowly than the technology. In the US, the Equal Credit Opportunity Act (ECOA) and the Fair Housing Act prohibit lending discrimination based on race, color, religion, national origin, sex, marital status, or age — and these laws apply to AI-based credit decisions just as much as human ones. The issue is enforcement.
The Consumer Financial Protection Bureau (CFPB) and the Department of Justice have made AI credit discrimination an explicit enforcement priority. In 2022, the CFPB issued guidance clarifying that lenders using AI must still be able to provide specific reasons for adverse actions — "the algorithm said no" is not an acceptable explanation under ECOA. That guidance put pressure on lenders to build more interpretable models or add explainability layers on top of complex ones.
The CFPB has also signaled interest in using disparate impact analysis — looking at outcomes by protected group, not just inputs — to identify discriminatory patterns in AI credit models. This matters because proxy discrimination doesn't require discriminatory intent to be illegal; if the outcome is systematically unfair to a protected group, that can constitute a violation regardless of what the model was designed to do.
In the EU, the AI Act places credit scoring systems in a "high-risk" category, requiring transparency, ongoing monitoring, and the ability to explain individual decisions to affected parties. This is stricter than current US requirements and may push international lenders toward more auditable model designs.
What It Means for You as a Borrower
If you're applying for credit through a fintech lender or any institution using AI-based underwriting, a few things are worth knowing.
Under ECOA, you have the right to receive a specific reason if you're denied credit or offered less favorable terms. If a lender provides only vague or generic adverse action language, you can request more specificity. The CFPB has resources for filing complaints if you believe a credit decision was unfair or based on discriminatory factors.
If you have a thin credit file and have been rejected by traditional lenders, AI-based fintech lenders may genuinely offer more relevant assessments — but read the data permissions carefully when applying. Some lenders request access to bank accounts, employment history, or other data sources. Understanding what you're sharing and how it will be used is part of making an informed decision.
And if you're ever uncertain whether a credit denial was fair, you're entitled to your free credit report from each of the three major bureaus through AnnualCreditReport.com. Review it for errors — AI models trained on inaccurate input data will produce inaccurate decisions, and correcting a credit report error can change outcomes quickly.
The Bottom Line
AI in credit decisioning is neither a straightforward villain nor a clean solution. It has genuinely expanded access to credit for people who were invisible to traditional models — and that matters. It has also encoded new forms of discrimination that are harder to detect, harder to challenge, and spreading faster than the regulatory frameworks designed to contain them. Both things are true simultaneously, which is exactly what makes this one of the more important developments in modern finance to stay informed about.
The technology will keep evolving. What determines whether it ultimately makes lending fairer is whether regulators, lenders, and consumers demand transparency, accountability, and outcomes that hold up to scrutiny — not just speed and efficiency.
FAQ
Can an AI credit model legally discriminate against me? No — intentional discrimination based on protected characteristics is illegal under ECOA and the Fair Housing Act regardless of whether it's done by a human or a machine. However, proving unintentional disparate impact in an AI system is more difficult than in a traditional process, and enforcement is still catching up to the technology.
Do I have the right to an explanation if I'm denied credit by an AI system? Yes. Under ECOA, lenders must provide specific, actionable reasons for adverse credit decisions. "The model denied your application" doesn't meet this standard. If you receive an inadequate adverse action notice, you can file a complaint with the CFPB.
Are AI credit decisions more accurate than traditional FICO-based ones? In many cases, yes — AI models can assess risk more precisely for borrowers who are thin-file or unconventional. But accuracy in predicting defaults doesn't automatically mean fairness. A model can be more accurate overall while still producing systematically biased outcomes for specific demographic groups.
What data are AI lenders actually using? It varies by lender and what applicants consent to share. Common data sources include traditional credit bureau data, bank account transaction history, income and employment records, and sometimes alternative data like rent payments. Each lender discloses data usage in their terms and privacy policy — reading those disclosures before applying is worthwhile.
What should I do if I think I was discriminated against in a credit decision? Document the decision and the reasons provided. File a complaint with the CFPB (consumerfinance.gov) or the relevant state financial regulator. If you believe the discrimination was intentional or egregious, consulting a consumer protection attorney is worth considering. Enforcement of fair lending laws in AI contexts is an active area, and complaints contribute to the regulatory record that shapes future enforcement.
📚 Sources
Consumer Financial Protection Bureau – Adverse Action Notification Requirements and the Equal Credit Opportunity Act – https://www.consumerfinance.gov/compliance/supervisory-guidance/adverse-action-notification-requirements-under-the-equal-credit-opportunity-act/
Bartlett R, Morse A, Stanton R, Wallace N – Consumer Lending Discrimination in the FinTech Era. National Bureau of Economic Research (2019) – https://www.nber.org/papers/w25943
Federal Reserve – Report to the Congress on the Use of Alternative Data in Credit Underwriting – https://www.federalreserve.gov/publications/files/alternative-data-report-122019.pdf
CFPB – Innovation Spotlight: Providing Adverse Action Notices When Using AI/ML Models – https://www.consumerfinance.gov/data-research/research-reports/innovation-spotlight-providing-adverse-action-notices-when-using-ai-ml-models/
Upstart – How Upstart's Credit Model Works – https://www.upstart.com/about
European Parliament – EU Artificial Intelligence Act – High Risk AI Systems – https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence
