The deployment of machine learning models in consumer credit decisions has generated mixed evidence on fairness outcomes through 2025 and 2026, with research from the Consumer Financial Protection Bureau and the Federal Reserve Bank of Philadelphia producing nuanced findings that complicate the narrative on both sides of the debate. Some ML credit models demonstrably expand access to underserved borrowers while others reproduce or amplify existing disparities, with the difference often coming down to model design choices rather than the underlying technology category.

The core technical debate centres on what data the models use and how they weight it. Traditional credit scoring built around FICO and similar systems relies on a limited set of variables including payment history, credit utilisation, length of credit history, types of credit, and new credit inquiries. Modern ML models can incorporate hundreds or thousands of additional features ranging from cash flow patterns in deposit accounts to phone bill payment timing to e-commerce purchase histories. The expansion of input features creates both opportunities for accuracy improvement and risk of indirect discrimination through proxy variables.

A January 2026 study from the Federal Reserve Bank of Philadelphia analyzed 1.8 million credit decisions across 12 major US lenders comparing ML models to traditional scoring. The headline finding was that ML models approved 14 percent more applications overall while maintaining equivalent or lower default rates. However, the approval rate increase was unevenly distributed. Hispanic applicants saw a 22 percent approval rate increase, Black applicants saw an 11 percent increase, and white applicants saw a 9 percent increase. The disparity widening for some demographic groups masks meaningful variation in approval improvement across racial categories.

Upstart, a publicly traded fintech using ML credit scoring for personal loans, has been one of the most studied cases. The company's credit model approves approximately 173 percent more borrowers than a traditional FICO-based equivalent at the same risk level, according to its own validation data. The Consumer Financial Protection Bureau granted Upstart an unusual no-action letter in 2017 specifically permitting its ML approach, with periodic reviews continuing through 2025. CFPB analysis found that Upstart's model produced approval rate improvements concentrated in lower-income and minority borrowers without proportional default rate increases, providing empirical support for the inclusivity case.

The discriminatory side of the debate also has substantial evidence. A November 2025 study from the National Bureau of Economic Research analyzed mortgage application decisions across 22 major US lenders and found that ML-driven systems produced 8 to 14 percent higher rejection rates for minority applicants in geographic areas with sparse historical data. The mechanism appears to be that ML models perform less reliably in data-sparse contexts and default to more conservative decisions, which disproportionately impacts demographic groups historically underrepresented in lending data.

Regulatory responses have varied across jurisdictions. The CFPB's 2024 guidance on AI in credit decisions required lenders to maintain explainability documentation for adverse decisions, in line with existing Equal Credit Opportunity Act requirements. The EU AI Act, in force since August 2024, classifies credit scoring as a high-risk AI application requiring extensive documentation, third-party audits, and ongoing monitoring. Singapore's Monetary Authority and Hong Kong's HKMA have issued similar guidance through their Veritas and AI Ethics frameworks, focusing on traceability of decisions and human-in-the-loop oversight.

Industry response to regulatory pressure has divided lenders into two camps. The first camp embraces transparency, investing heavily in explainable AI tooling, third-party model audits, and rigorous adverse-impact testing. Lenders in this camp include Upstart, Affirm, Capital One, and several large credit unions. They argue that the marginal cost of compliance is more than offset by the regulatory tailwind and the ability to confidently expand into previously underserved segments.

The second camp uses ML models more cautiously, often as secondary inputs to traditional scoring rather than primary decision systems. JPMorgan, Wells Fargo, and Bank of America generally fall in this camp, with their consumer credit decisions still anchored on FICO scores supplemented by ML-driven risk overlays for specific decision types. The risk-averse approach limits the upside but also limits regulatory exposure if model bias claims emerge.

Geographic and product variation further complicates the picture. Credit card and personal loan decisions appear to benefit most from ML expansion, with broad agreement across studies that approval rates can increase without default rate degradation. Mortgage decisions show more mixed results, particularly in areas with thin local data. Auto loan and small business loan decisions sit in between, with the specific lender's data quality and model architecture mattering more than the underlying technology.

For consumers seeking to understand whether they should expect better or worse outcomes from ML credit decisions, the practical answer depends on circumstances. Borrowers with thin or unconventional credit histories, including immigrants, gig workers, and young adults, often see meaningful approval rate improvements from ML models that incorporate cash flow data and alternative variables. Borrowers with established traditional credit profiles see marginal improvement at best, since traditional models already perform well in those segments.

For policy makers and lenders, the evidence supports a calibrated approach rather than blanket optimism or pessimism. ML credit models can demonstrably expand access while maintaining or improving repayment performance, but only when implementation includes rigorous bias testing, transparent adverse-impact reporting, and ongoing monitoring. Models deployed without these safeguards can produce or amplify discrimination even unintentionally. The question of whether AI credit scoring is fairer or more discriminatory does not have a single answer because the answer depends entirely on how the model is designed and operated.

The 2026 trajectory points toward continued expansion of ML credit decisions with strengthening regulatory oversight. The lenders who get the implementation right are likely to capture profitable underserved market segments while improving overall credit quality. The lenders who deploy without adequate fairness testing risk regulatory action, customer attrition, and reputational damage. The technology is not the variable. The implementation discipline is.