Lenders using screening systems that make similar mistakes can create a more divided credit market, with some borrowers paying more even when individual lenders’ assessment accuracy is unchanged, according to an October economics working paper.
The National Bureau of Economic Research paper, by Maryam Farboodi, Péter Kondor and Pablo Kurlat, examines how lenders’ decisions affect the applicants left for their competitors. It models a risk that can arise when institutions rely on overlapping data or similar analytical systems.
This is a theoretical study, not an investigation establishing that a named bank or artificial intelligence product has overcharged borrowers. It has not been peer reviewed.
Screening can be accurate overall and still miss the same people
Credit screening is the process of assessing whether an applicant is likely to repay. A false rejection turns away someone who would have repaid; a false acceptance approves someone who subsequently defaults.
Two lenders can each have a satisfactory overall accuracy rate while repeatedly misjudging the same applicants. Alternatively, their errors can occur among different people. The paper compares these possibilities while holding a lender’s accuracy at a given screening precision constant.
The model also separates repayment quality from how difficult a borrower is to assess. An applicant can be creditworthy but hard to evaluate with the available information. The authors call this difficulty “opacity”. It does not mean the borrower is concealing misconduct.
Imagine, as a hypothetical example, a young business with irregular but genuine contract income. Several automated assessments might struggle with that pattern. A different assessment using additional records might reach a better-informed decision, although neither a human reviewer nor another algorithm is automatically correct.
Rejected applications change the next lender’s market
Applicants rejected by a low-rate lender may approach another lender charging more. The second institution therefore assesses a pool shaped partly by the first institution’s decisions.
Economists call this a pool externality: one company’s screening changes the opportunities and risks facing other companies, beyond its own customers.
When errors are independent across lenders, the model produces a relatively smooth range of lending services and rates. When errors overlap in the particular way the authors model, lenders instead specialize in three segments.
One segment offers low rates with relatively basic screening. Another spends more on precise assessments of harder-to-evaluate applicants and charges more. A third lends without screening at very high rates.
All borrowers eventually receive credit in the baseline model. Its highest-cost segment should therefore not be described as proof that applicants are completely excluded from borrowing. Access to affordable credit and access to any credit are separate outcomes.
Lower rates within a segment can coexist with a higher average
A counterintuitive result concerns the average cost of borrowing. Under correlated errors, rates can be lower at a given screening skill level while more lending shifts toward high-rate segments.
The average borrower may consequently pay more. A lower price within each comparable category does not establish a lower market-wide average when the proportions of borrowers in those categories change.
This is a possible model outcome, not a measured increase in today’s loan rates. Actual markets also involve regulation, capital requirements, collateral, customer relationships and other features that a simplified model cannot reproduce in full.
For lenders evaluating a new system, the implication is that an aggregate approval rate or accuracy score can leave important questions unanswered. Which applicants are repeatedly rejected, what additional evidence would change the assessment, and how much does the model depend on the same inputs as competing systems?
Better technology can also lower borrowing costs
The paper does not conclude that advanced screening is harmful. In one scenario, cheaper assessment of the most opaque borrowers allows specialized lenders to serve customers who would otherwise rely on expensive, indiscriminate credit.
The resulting competition can also lower rates for borrowers remaining with high-cost lenders.
A different scenario considers data-sharing arrangements, such as open banking, that make information about relatively transparent borrowers cheaper to assess. Their effects spread through the remaining applicant pool. Depending on what information and analytical capability become available, the highest-cost borrowers can benefit or lose.
The direction of the technological improvement therefore matters. A tool that helps lenders understand previously difficult applications does something different from a tool that lets them compete more cheaply for applicants already easy to assess.
Regulators have identified related concentration risks
The Financial Stability Board’s 2024 report on AI in finance identified service-provider concentration, market correlations, data quality and model governance among potential vulnerabilities. Those broader concerns provide context for the paper, rather than proof of its modeled outcomes.
A 2024 Bank of England and Financial Conduct Authority survey found that 75% of its 118 financial-services respondents were already using AI. The survey covered many activities, including internal operations. It did not show that three-quarters of loan decisions were automated or establish the current adoption rate in 2026.
Using multiple suppliers may reduce some dependencies, but different brand names alone do not demonstrate different information or error patterns. Assessing that requires evidence about the systems and the outcomes they produce.
Borrowers need a comparable view of the offer
For a business seeking finance, complete, consistent accounts and an explanation of unusual cash flows can help a lender understand the application. They cannot guarantee approval or a particular rate.
A comparison should record the interest basis, fees, security, repayment schedule and any personal guarantee. A lower advertised rate can still accompany a costly arrangement if charges or repayment conditions differ.
Our coverage of how going public changes companies’ borrowing relationships examines separate evidence about the role of information in finance. It does not suggest that an IPO is an appropriate route for every business.
The new paper leaves a specific question for lenders and their customers: does an additional assessment bring useful new information, or reproduce the same uncertainty? More decisions do not necessarily provide more independent judgments.
