[Finance](https://bfcs.ai/industry/finance/) · [Safer](https://bfcs.ai/safer/) # Can AI Make Lending Fairer? In our reading, credit models are easier to audit than loan officers, which makes bias measurable, and one peer-reviewed study suggests it can be reduced at relatively low cost. But a model that passes a fairness test on launch day can drift, be recalibrated by a vendor or widen a disparity through its own flexibility. Fairness in lending is a reading that has to keep being taken. Updated **October 2026** · 7 min read · Evidence-graded · 8 sources ## The BFCS verdict Every article scored on the same four dimensions | Dimension | Verdict | Why | | --- | --- | --- | | Better | Partly · Moderate confidence | Better prediction, unevenly shared · Peer-reviewed US mortgage research found Black and Hispanic borrowers were less likely to gain from a switch to machine learning, so a more accurate model is not automatically a fairer one. | | Faster | Partly · Low confidence | Decisions are quick; proper testing is not · We found no independent study in this review that measures decision speed against fairness outcomes. In our reading, the time saved in scoring is partly spent on the testing, monitoring and explanation work regulators expect. | | Cheaper | Partly · Moderate confidence | Fairness constraints appear affordable · A peer-reviewed comparison on real credit data found algorithmic discrimination can be reduced to a reasonable level at relatively low cost to profit. Pricing disparities, by contrast, cost US minority borrowers over $450 million a year. | | Safer | Conditional · Governance-gated | Only with continuous outcome testing · Bias can be measured and reduced, but in a sample of 23 licensees ASIC found very few proactively identified bias risks or said they actively tested for bias (more showed mitigation when asked), and some did not appear to test outcomes on an ongoing basis. The EU AI Act classes credit scoring as high risk and, once those rules apply, requires lifecycle risk management and post-market monitoring. | ## The problem, in numbers Lending data carries a measurable gap. In 2023 United States mortgage data, 17.1% of Black applicants and 12.7% of Hispanic white applicants for a home purchase loan were denied, against 9.7% of Asian applicants and 6.8% of non-Hispanic white applicants. (Source 1) These are raw rates, not adjusted for credit history or loan size, and the report notes that Black and Hispanic white borrowers also had lower median credit scores. (Source 1) A raw gap does not prove discrimination, but in our reading it is the history a model trained on past lending decisions learns from. ### US home purchase mortgage denial rates by applicant group, 2023 (Source 1) Black or African American · 17.1% Hispanic white · 12.7% Asian · 9.7% Non-Hispanic white · 6.8% Share of applications denied, excluding withdrawn and incomplete files. Unadjusted for credit or loan characteristics. Bar widths are our calculation, scaled to the highest rate. Price matters too. A peer-reviewed study of US mortgages found that risk-equivalent Latinx and Black borrowers paid significantly higher interest rates on GSE-securitised and FHA-insured mortgages, costing them over $450 million a year. (Source 2) In Australia, the evidence is about governance rather than outcomes. ASIC reviewed 624 AI use cases across 23 banking, credit, insurance and financial advice licensees, chosen as likely AI users rather than as a representative sample. Only 12 of those licensees had AI policies, guidance or checklists that mentioned fairness, discrimination or bias, and only 10 had documented requirements on telling consumers when AI was involved. (Source 5) None appeared to have specific arrangements for a consumer to contest an AI decision. (Source 5) ### The evidence, in four numbers 624 · AI use cases reviewed by ASIC · across 23 banking, credit, insurance and advice licensees, as at December 2023; ASIC chose licensees it judged most likely to use AI, so the sample is not representative 12 of 23 · licensees whose AI policies referenced fairness, discrimination or bias · same ASIC review; 10 had documented requirements on telling consumers about AI use 0 · licensees that appeared to have specific arrangements to contest an AI decision · same ASIC review; some noted the concept in principle over $450m · extra interest paid each year by risk-equivalent Latinx and Black US borrowers · GSE-securitised and FHA-insured mortgages (Bartlett et al., Journal of Financial Economics, 2022) ## What improves, and why In our reading, a credit model has one property a loan officer lacks: it can be audited. You can rerun it on the same applicants, change one input and watch what moves, and compare outcomes across groups at scale. That makes bias testing in lending more achievable than testing human judgement. Outcome data makes lenders comparable. In the US pricing study, FinTech lenders' rate disparities were similar to those of other lenders for one class of mortgage, but lower for FHA-insured loans in certain years. (Source 2) Measurement shows where a disparity is smaller and where it is not. Fixing what the test finds need not be ruinous. A peer-reviewed comparison of fairness techniques on real credit data concluded that algorithmic discrimination "can be reduced to a reasonable level at a relatively low cost" to profit, and that several fairness criteria can be approximately satisfied at the same time. (Source 4) > A fairness test is a reading taken on one day. Fairness in a lending book is a property of the system over years, so the test has to keep running. ## What the evidence actually says Peer-reviewed research warns that accuracy and fairness can pull apart. Fuster and colleagues compared traditional and machine learning default models on US mortgage data and found Black and Hispanic borrowers were disproportionately less likely to gain from the switch. In their market model, machine learning increased rate disparities both between and within groups. (Source 3) They traced the effect mainly to the models' extra flexibility, rather than to their use of characteristics that traditional models leave out. (Source 3) In our reading, a disparity can widen without any new sensitive input, so checking the input list is not enough; only testing outcomes will show it. A literature review by the UK regulator's researchers finds bias enters through data shaped by historical exclusion and through modelling choices, and that technical fixes can cut accuracy or shift bias onto other groups, so they need context and human review. (Source 8) Then there is time, which in our reading is where one-off certification fails. ASIC found some licensees knew that bias in their data could affect outcomes but did not appear to test for a disparity in outcomes on an ongoing basis, and a small number reported testing only before and at deployment, relying on trigger-based reviews after that. (Source 5) Its case studies show what can go wrong. One licensee's credit default model was described in an internal report about 10 months after launch as a black box, with poor governance and a lack of a monitoring process, and stayed in use for several more months before a simpler model replaced it. (Source 5) At another, a vendor recalibrated the credit scores and the licensee recorded that there had not been time for thorough testing. (Source 5) Better practice was routine weekly monitoring that caught unexpected scores, then root cause analysis and a check for consumer impact. (Source 5) ## Where it falls short - **Flexibility can widen gaps.** In the US mortgage research, machine learning widened rate disparities mainly through greater flexibility, (Source 3) so fairness has to be measured on outcomes. - **Vendors change the model under you.** In our reading, a third-party recalibration is a new model for fairness purposes. One licensee recorded that time did not allow thorough testing of such a change. (Source 5) - **Explanations are still owed.** US Regulation B requires the specific principal reasons for a refusal; saying an applicant missed a qualifying score is not enough. (Source 6) The CFPB withdrew its 2022 circular applying this to complex algorithms in May 2025, pending a review, but the requirement remains in the regulation itself. (Source 6) - **Little Australian outcome data.** We found no published Australian data on lending outcomes by protected attribute, so local lenders cannot benchmark against a public baseline. ## Where governance decides the outcome The rules increasingly assume fairness is monitored, not certified. ASIC says the Australian framework is technology neutral: licensees must provide services efficiently, honestly and fairly, should weigh whether AI brings a risk of unfairly biased or discriminatory treatment, and should monitor and regularly review their compliance measures. (Source 5) The EU AI Act classes credit scoring of individuals as high risk, treats risk management as a continuous lifecycle process, and requires providers to analyse performance data throughout the system's lifetime. (Source 7) The AI Omnibus regulation moved those high-risk rules for Annex III uses, which include credit scoring, to 2 December 2027. (Source 7) In the UK, the FCA has said that firms using AI in a way that embeds or amplifies bias, leaving some groups worse off, might not be acting in good faith, as the Consumer Duty requires, unless the differences can be objectively justified, and the Equality Act 2010 also applies. (Source 8) In our reading, lenders should name an owner for fairness results, set the threshold that triggers review, and give customers a real route to challenge a decision. ## How to think about adoption Treat a fairness assessment as the start of a monitoring programme, not a certificate. Before launch, test outcomes across the groups you can measure, using approval rates, pricing and error rates, and document the trade-offs you accept. After launch, rerun the same tests on a fixed schedule and every time the data, the cut-off or the vendor's model changes. When results shift, find the cause before adjusting thresholds; ASIC called simply amending thresholds back within risk appetite, without root cause analysis, poorer practice. (Source 5) Make sure every refusal can be explained in specific terms, and tell customers when a model was involved. One peer-reviewed study suggests the fairness constraint is affordable. ASIC's review suggests ongoing outcome testing is where the gaps are. ## How we read this evidence BFCS.ai does not fabricate figures. The outcome evidence here is from the United States, where mortgage data by race is public; we found no published Australian data on lending outcomes by protected attribute, so the Australian evidence is about governance practice. Three journal studies are cited from their published abstracts only, because the publisher pages refused automated access. ## Sources 1. Consumer Financial Protection Bureau, "Data Point: 2023 Mortgage Market Activity and Trends," December 2024, section 3.3 and Table 4A. Home purchase denial rates in 2023: [files.consumerfinance.gov](https://files.consumerfinance.gov/f/documents/cfpb_2023-mortgage-market-activity-and-trends_2024-12.pdf) **What we took from it:** all applicants 9.4%, Black or African American 17.1%, Hispanic white 12.7%, Asian 9.7%, non-Hispanic white 6.8%. Base: closed-end first-lien home purchase mortgages secured by site-built single-family principal residences; denial rate is denied applications divided by all applications, excluding withdrawn and incomplete files. The rates are not adjusted for credit or loan characteristics; the report notes (section 3.2 and Table 3A) that Black and Hispanic white borrowers who took out home purchase loans continued to have lower median credit scores than other groups. Bar widths are our calculation, scaled to the highest rate. This was the most recent edition of the report we could locate. 2. Bartlett, R., Morse, A., Stanton, R. & Wallace, N., "Consumer-lending discrimination in the FinTech Era," Journal of Financial Economics 143(1), 2022, pp. 30-56 (peer-reviewed), as summarised in the published abstract. [ideas.repec.org](https://ideas.repec.org/a/eee/jfinec/v143y2022i1p30-56.html) **What we took from it:** Risk-equivalent Latinx and Black borrowers paid significantly higher interest rates on GSE-securitised and FHA-insured loans, with larger gaps in high-minority-share neighbourhoods, costing them over $450 million a year; FinTech lenders' rate disparities were similar to those of other lenders for GSE mortgages but lower for FHA mortgages issued in 2009 to 2015 and for FHA refinance mortgages issued in 2018 to 2019. (publisher DOI 10.1016/j.jfineco.2021.05.047; the publisher page refused automated access, so figures are from the abstract only). 3. Fuster, A., Goldsmith-Pinkham, P., Ramadorai, T. & Walther, A., "Predictably Unequal? The Effects of Machine Learning on Credit Markets," Journal of Finance 77(1), 2022, pp. 5-47 (peer-reviewed), as summarised in the published abstract. [ideas.repec.org](https://ideas.repec.org/a/bla/jfinan/v77y2022i1p5-47.html) **What we took from it:** Using US mortgage data, Black and Hispanic borrowers are disproportionately less likely to gain from the introduction of machine learning; in a simple equilibrium credit market model, machine learning increases disparity in rates between and within groups; the changes are attributed primarily to the greater flexibility of machine learning rather than its use of characteristics excluded from traditional models. (publisher DOI 10.1111/jofi.13090; the publisher page refused automated access). 4. Kozodoi, N., Jacob, J. & Lessmann, S., "Fairness in credit scoring: Assessment, implementation and profit implications," European Journal of Operational Research 297(3), 2022, pp. 1083-1094 (peer-reviewed), as summarised in the published abstract. [ideas.repec.org](https://ideas.repec.org/a/eee/ejores/v297y2022i3p1083-1094.html) **What we took from it:** Compares fairness processors in a profit-oriented credit scoring setting on real-world data; several fairness criteria can be approximately satisfied at once; separation is recommended as the measure of scorecard fairness; fair in-processors give a good balance between profit and fairness; algorithmic discrimination can be reduced to a reasonable level at a relatively low cost. (publisher DOI 10.1016/j.ejor.2021.06.023; the publisher page refused automated access). 5. Australian Securities and Investments Commission, REP 798, "Beware the gap: Governance arrangements in the face of AI innovation," 29 October 2024. [download.asic.gov.au](https://download.asic.gov.au/media/mtllqjo0/rep-798-published-29-october-2024.pdf) **What we took from it:** 624 AI use cases from 23 AFS and credit licensees in banking, credit, insurance and financial advice, as at December 2023, drawn from licensees ASIC identified as most likely to be using AI, so not representative of AI use generally or of the sectors (pp. 4 and 38; detailed information reviewed for all use cases of smaller licensees and an ASIC-selected subset for larger ones; ASIC did not test consumer outcomes of individual models, p. 38). Only 12 licensees had AI policy documents, guidance or checklists referencing fairness or related concepts such as discrimination or bias; only 10 had documented requirements on disclosure to consumers; no licensee appeared to have implemented specific contestability arrangements for AI (p. 19). Very few licensees proactively identified algorithmic bias risks in their responses about particular use cases or indicated they actively tested for bias, in some cases likely because bias was less relevant to the use case; more showed they had considered and mitigated the risk when ASIC specifically asked; one indicated it did not test for bias; some were aware of possible bias in data sets but did not appear to test for a disparity in outcomes on an ongoing basis; a small number reported testing only at the pre-deployment and deployment stages and relying on trigger-based reviews after deployment, while most licensees were monitoring their models with widely varying practices (pp. 21 and 23). Case studies: a credit default model described in an internal report about 10 months after deployment as a black box, with poor governance and a lack of a monitoring process, used for several months more before replacement by a simpler model (p. 7; the internal report also said the model had been stable); a vendor recalibration of credit scores where the licensee recorded that time did not allow for thorough testing, against a licensee whose routine weekly monitoring caught unexpected scores and led to root cause analysis and a check for consumer impact; amending thresholds without root cause analysis is described as poorer practice (p. 23). The framework is technology neutral; licensees should consider whether AI brings risks of unfairly biased or discriminatory treatment to their obligation to act efficiently, honestly and fairly, and compliance measures should be documented, implemented, monitored and regularly reviewed (p. 34). 6. Consumer Financial Protection Bureau, Regulation B, 12 CFR 1002.9(b)(2): [consumerfinance.gov](https://www.consumerfinance.gov/rules-policy/regulations/1002/9/) CFPB, Consumer Financial Protection Circular 2022-03, "Adverse action notification requirements in connection with credit decisions based on complex algorithms," 26 May 2022: [consumerfinance.gov](https://www.consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms/) CFPB, "Interpretive Rules, Policy Statements, and Advisory Opinions; Withdrawal," 90 FR 20084, 12 May 2025: [govinfo.gov](https://www.govinfo.gov/content/pkg/FR-2025-05-12/html/2025-08286.htm) **What we took from it:** the statement of reasons for adverse action must be specific and indicate the principal reasons; statements that the applicant failed to achieve a qualifying score on the creditor's credit scoring system are insufficient. the requirements apply equally to all credit decisions regardless of the technology used. withdraws 67 guidance documents, including Circulars 2022-03 and 2023-03, which should not be enforced or otherwise relied on by the Bureau while its review is ongoing; the notice says the withdrawal is not necessarily final. The text of 12 CFR 1002.9(b)(2) was unchanged in the eCFR as at 29 October 2026. 7. Regulation (EU) 2024/1689 (AI Act), text as published by the European Commission AI Act Service Desk. [ai-act-service-desk.ec.europa.eu](https://ai-act-service-desk.ec.europa.eu/en/ai-act/annex-3) Article 9(2): [ai-act-service-desk.ec.europa.eu](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-9) Article 72: [ai-act-service-desk.ec.europa.eu](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-72) European Commission, "AI Act" policy page (updated 19 October 2026): [digital-strategy.ec.europa.eu](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) Regulation (EU) 2026/1744 (Digital Omnibus on AI), 8 July 2026, Official Journal: [eur-lex.europa.eu](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng) **What we took from it:** Annex III point 5(b): AI systems intended to evaluate the creditworthiness of natural persons or establish their credit score are high risk, except those used to detect financial fraud. the risk management system is a continuous iterative process planned and run throughout the entire lifecycle, requiring regular systematic review and updating. providers must establish a post-market monitoring system that actively and systematically collects, documents and analyses relevant performance data throughout the system's lifetime. credit scoring is given as an example of a high-risk use; high-risk requirements include high-quality datasets to minimise risks of discriminatory outcomes; the AI Omnibus entered into force on 27 July 2026, and the rules for high-risk use cases in certain sensitive areas (Annex III) have been extended to 2 December 2027. the application of Sections 1 to 3 of Chapter III is set to 2 December 2027 for AI systems classified as high-risk under Article 6(2) and Annex III. 8. Financial Conduct Authority, "Artificial Intelligence (AI) update: further to the Government's response to the AI White Paper," 22 April 2024, paragraphs 3.24 to 3.26 and 3.33. [fca.org.uk](https://www.fca.org.uk/publication/corporate/ai-update.pdf) FCA Research Note, "A literature review on bias in supervised machine learning," first published 11 December 2024: [fca.org.uk](https://www.fca.org.uk/publications/research-notes/research-note-literature-review-bias-supervised-machine-learning) **What we took from it:** The Consumer Duty addresses discrimination harms by requiring firms to take account of customers' different needs, including those with protected characteristics; firms using AI in a way that embeds or amplifies bias, leading to worse outcomes for some groups, might not be acting in good faith unless differences in outcome can be justified objectively; firms are also subject to the Equality Act 2010. bias arises from data issues, including historical practices of exclusion, and from modelling choices; technical mitigation may affect model accuracy or have unintended consequences for bias against other groups, and should be supplemented by consideration of context and human review. Research notes may not represent the FCA's position. Putting this to work ## Working through this in your own organisation? BFCS.ai is published by Inhouse CX. We help small and mid-sized organisations decide where AI earns its place, and put the governance in place that makes it safe to use. [Talk to Inhouse CX →](https://inhouse.cx/engage/) Continue the thread - [Can AI Detect Fraud Better?](https://bfcs.ai/articles/can-ai-detect-fraud-better/): More in Finance · FINANCE · BETTER · 7 min · Read → - [Can AI Cut Compliance Costs?](https://bfcs.ai/articles/can-ai-cut-compliance-costs/): More in Finance · FINANCE · CHEAPER · 7 min · Read → - [Can AI Make Benefits Decisions Fairer?](https://bfcs.ai/articles/can-ai-make-benefits-decisions-fairer/): Also Safer · GOVERNMENT · SAFER · 7 min · Read → --- Canonical page: https://bfcs.ai/articles/can-ai-make-lending-fairer/ This is the Markdown version of the page, generated from its HTML at build time. Site guide for agents: https://bfcs.ai/llms.txt · Every page: https://bfcs.ai/sitemap.xml