Discriminatory Lending Algorithms: The Federal Probe

The Interagency AVM Rule: Compliance Realities Post October 2025

On October 1, 2025, the U. S. housing finance sector entered a new regulatory era. The long-awaited “Quality Control Standards for Automated Valuation Models” final rule, issued jointly by six federal agencies including the CFPB, OCC, and FHFA, is fully. This mandate transforms what was previously a patchwork of supervisory guidance into rigid federal law. Financial institutions, specifically mortgage originators and secondary market issuers, no longer have the option to treat algorithmic property valuations as proprietary “black boxes.” The rule enforces a strict liability standard: if an algorithm discriminates, the lender is responsible, regardless of whether the software was built in-house or purchased from a third-party vendor.

The regulation, finalized in July 2024, codified requirements under the Dodd-Frank Act. While the statute originally outlined four quality control standards, the agencies exercised their authority to add a fifth, mandatory pillar: compliance with applicable nondiscrimination laws. This addition forces lenders to actively test for impact on protected classes. Banks must maintain auditable evidence that their models do not replicate historical redlining patterns found in training data. The “set it and forget it” method to AVM deployment is legally dead. Compliance officers face the immediate load of retrofitting governance frameworks to meet these new federal standards or risking severe supervisory actions.

Table 1: The Five Pillars of AVM Compliance ( Oct 1, 2025)

Regulatory Pillar	Operational Requirement	Compliance load
Confidence Levels	Ensure high confidence in estimates produced.	Requires continuous statistical validation of error rates (e. g., PPE, FSD).
Data Integrity	Protect against manipulation of input data.	Mandatory “data hygiene” to prevent scrubbing or altering comparables.
Conflict of Interest	Segregate model operation from loan production.	Loan officers cannot influence AVM selection or override values without audit.
Random Sampling	Require random sample testing and reviews.	Regular back-testing of AVM values against physical appraisals to check accuracy.
Nondiscrimination	Comply with fair lending laws (The “Fifth Factor”).	Active impact analysis; testing for proxy variables (e. g., zip code bias).

The of this regulatory shift correlates directly with the industry’s reliance on automation. Data from Corporate Settlement Solutions (CSS) indicates that by 2024, lenders used AVMs or property condition reports for 34% of home equity loans, a significant jump from 25% the previous year. In the bridging loan sector, United Trust Bank reported AVM usage method 50% by early 2026. Speed and cost reduction drive this adoption. AVMs cost a fraction of a traditional $600 appraisal and return results in seconds. Yet, this efficiency creates a high-velocity risk environment. When an algorithm processes millions of valuations, a minor weighting error in the code can systematically undervalue minority-owned assets across entire metropolitan statistical areas (MSAs).

A sharp divide exists regarding the neutrality of these systems. Industry proponents, such as Veros Real Estate Solutions, released a 2024 year-end update claiming their models show “no evidence of widespread undervaluation” in minority neighborhoods. Their data suggests that metrics like the Mean Absolute Error (MAE) remain consistent regardless of a neighborhood’s racial composition. Regulators and independent researchers dispute this optimism. A November 2025 study from the University of Illinois found that algorithmic scoring models consistently penalized female borrowers and minority groups, even when controlling for credit risk variables. The study highlighted that “blind” algorithms frequently use proxy variables—such as location data or spending patterns—to reconstruct the very protected classes they are forbidden to consider.

The “Random Sample Testing” requirement (Pillar 4) presents the most immediate logistical hurdle for small and mid-sized banks. The rule demands that lenders compare a statistically significant sample of AVM valuations against physical appraisals or other value indicators. This creates a continuous feedback loop that legacy systems cannot support without manual intervention. For secondary market issuers, the liability extends to securitization determinations. If a bundle of mortgage-backed securities relies on AVMs that fail these quality control standards, the issuer faces chance repurchase demands and regulatory fines. The era of plausible deniability regarding vendor algorithms has ended; the lender owns the math.

Enforcement actions are expected to target institutions that fail to document their “Fifth Factor” testing. The CFPB has signaled that mere technical accuracy is insufficient if the model produces discriminatory outcomes. As the 2026 examination pattern begins, examiners can request the specific regression analyses used to clear these models for use. Banks unable to produce this documentation face a harsh reality: the cost of compliance is high, but the cost of a fair lending violation is existential.

References

• Federal Register (Aug 07, 2024): Quality Control Standards for Automated Valuation Models; Final Rule.
• Corporate Settlement Solutions (Feb 19, 2025): AVM Share Up, Bright Outlook for Home Equity Lending.
• University of Illinois (Nov 12, 2025): New research reveals widespread bias, in credit scoring and mortgage lending.
• Veros Real Estate Solutions (Jan 08, 2025): 2024 Year-End Update: AVM Performance: Is There Evidence of Racial Bias?
• Mayer Brown (Oct 09, 2025): Mortgage Originators and Secondary Market Participants Take Note: The Final Rule on AVMs Is.

Fourth Circuit Ruling: The Reinstatement of the Navy Federal Class Action

On February 9, 2026, the U. S. Court of Appeals for the Fourth Circuit issued a decisive ruling in In re: Navy Federal Mortgage Discrimination Litigation, reversing a lower court decision that had dismantled the largest fair lending class action in recent history. The appellate panel vacated the dismissal of class allegations against Navy Federal Credit Union, the nation’s largest credit union with over $170 billion in assets. This judgment establishes a binding precedent: plaintiffs challenging algorithmic underwriting systems need not identify the specific line of code causing bias at the pleading stage. They must only demonstrate a statistical sufficient to infer a widespread flaw.

The lawsuit originated from a December 2023 CNN investigation which analyzed Home Mortgage Disclosure Act (HMDA) data from 2022. The report exposed a 29-percentage-point gap in mortgage approval rates between White and Black applicants at Navy Federal, the widest among the top 50 U. S. mortgage lenders. The data revealed that the credit union approved 77. 1% of White applicants but only 48. 5% of Black applicants. More damning was the income inversion: Navy Federal approved a higher percentage of White borrowers earning less than $62, 000 annually than Black borrowers earning $140, 000 or more.

The District Court’s Dismissal

In May 2024, U. S. District Judge Leonie Brinkema of the Eastern District of Virginia granted Navy Federal’s motion to strike the class allegations. Judge Brinkema reasoned that mortgage underwriting decisions are “highly individualized,” relying on specific borrower credit profiles that vary too widely to support a common class action. She dismissed the treatment claims—allegations of intentional discrimination—citing a absence of direct evidence. While she allowed individual impact claims to proceed, the removal of the class method neutralized the financial threat to the credit union, as individual fair lending suits are rarely economically viable for plaintiffs.

The Appellate Reversal

The Fourth Circuit’s February 2026 opinion rejected Judge Brinkema’s logic. The appellate judges held that because Navy Federal uses a centralized, semi-automated underwriting engine to process all applications, the “commonality” requirement for a class action was satisfied. The court stated that requiring plaintiffs to pinpoint the exact mechanical failure of a proprietary “black box” algorithm before discovery would create an impossible barrier to civil rights enforcement. By reinstating the class claims, the court shifted the load back to lenders to prove their algorithms are neutral, rather than requiring outsiders to reverse-engineer them without access to the data.

Table 2. 1: Navy Federal Credit Union Mortgage Approval Rates (2022)

Applicant Race	Approval Rate	Denial Rate	vs. White
White	77. 1%	22. 9%	–
Latino	55. 8%	44. 2%	-21. 3 pts
Black	48. 5%	51. 5%	-28. 6 pts

Navy Federal has consistently defended its practices, arguing that the HMDA data used by CNN and the plaintiffs is incomplete. The credit union asserts that the public dataset absence credit scores, debt-to-income ratios, and accumulated assets—factors they claim explain the. Yet the Fourth Circuit noted that even when controlling for income and debt ratios available in the dataset, the racial gap remained statistically anomalous. The case returns to the district court for discovery, where Navy Federal can be compelled to hand over the precise inputs and weighting of its algorithmic models. This phase can determine whether the is a result of legitimate risk factors or a digital form of redlining in the code.

“The apparent impact is there for all to see—Black and Brown borrowers were denied mortgages far more frequently and charged higher rates than similarly situated White applicants. This ruling reaffirms that courts should not shut down civil rights cases before plaintiffs have any opportunity to access the evidence.”

The Enforcement Void: Federal Deregulation vs State AG Aggression

Between 2017 and 2021, a distinct regulatory vacuum emerged in the United States financial sector. While algorithmic lending tools grew in complexity and ubiquity, federal oversight stalled. The Consumer Financial Protection Bureau (CFPB), under the leadership of Mick Mulvaney and later Kathy Kraninger, dramatically reduced its enforcement footprint. Public records indicate that fair lending enforcement actions dropped by over 70% during this period compared to the previous administration. This retreat created an “enforcement void” where digital redlining and algorithmic bias could metastasize unchecked by federal authorities. Lenders rapidly adopted “black box” underwriting models, confident that federal regulators absence the political can to challenge the impacts of these new technologies.

Into this breach stepped a coalition of aggressive State Attorneys General (AGs). Recognizing that federal preemption did not shield lenders from state consumer protection and anti-discrimination laws, AGs in Pennsylvania, New York, Massachusetts, North Carolina, and California launched a series of high- investigations. These state-level actions shifted the liability, treating algorithmic bias not as a technical glitch but as a violation of state civil rights statutes. The message was clear: if Washington would not police the code, the states would.

The Trident Precedent and the Pennsylvania Pivot

The turning point for state-led enforcement arrived in July 2022 with the Trident Mortgage Company settlement. While the DOJ and CFPB were involved, the investigation was initiated and driven by the Pennsylvania Office of Attorney General under Josh Shapiro, alongside counterparts in New Jersey and Delaware. The probe revealed that Trident’s marketing algorithms and loan officer distribution deliberately avoided majority-minority neighborhoods in the Philadelphia-Camden-Wilmington region. The $20 million settlement was historic not just for its size, but for its structure: it required Trident to subsidize loans in the very neighborhoods its data models had excluded. This case established a blueprint for state AGs to prosecute modern-day redlining without waiting for federal permission.

Table 3. 1: Key State-Led Fair Lending Enforcement Actions (2022–2025)

State	Target Entity	Violation Type	Settlement / Action Date	Impact / Penalty
Pennsylvania / NJ / DE	Trident Mortgage Co.	Redlining / Marketing Bias	July 2022	$20 Million + Loan Subsidies
North Carolina	National Bank of PA	Redlining Minority Tracts	February 2024	$13. 5 Million Settlement
Massachusetts	Earnest Operations LLC	AI Underwriting Bias	July 2025	$2. 5 Million + Code Audit
California	Statewide Lenders	Algorithmic Discrimination	October 2025	New FEHA AI Regulations
Texas / DOJ	Colony Ridge	Predatory Lending / Targeting	February 2026	$68 Million Settlement

Massachusetts the Code: The Earnest Operations Settlement

While Trident addressed geographic redlining, the Massachusetts Attorney General’s Office took direct aim at the algorithms themselves. In July 2025, AG Andrea Joy Campbell announced a $2. 5 million settlement with Earnest Operations LLC, a student loan servicer. The investigation found that Earnest’s AI-driven underwriting model penalized applicants based on factors that served as proxies for race and immigration status, such as attending Historically Black Colleges and Universities (HBCUs) or absence specific citizenship documentation, even when creditworthiness was high. Unlike previous settlements that focused on human intent, this action targeted the outcome of the code. The settlement mandated a complete overhaul of the company’s algorithmic governance, requiring regular third-party audits to test for impact—a standard the federal government had hesitated to enforce broadly until the 2025 AVM Rule.

California’s Regulatory Firewall

California escalated the conflict by codifying enforcement into state law. On October 1, 2025, the California Civil Rights Department implemented new regulations under the Fair Employment and Housing Act (FEHA). These rules explicitly state that the use of AI or automated decision-making systems (ADS) does not shield companies from liability for discriminatory outcomes. The regulations place the load of proof on the user of the algorithm to demonstrate that a discriminatory model is “job-related” or a “business need” and that no less discriminatory alternative exists. This move preempted the “black box” defense in California courts, forcing lenders and landlords to unmask their proprietary models or face severe state-level sanctions.

The Colony Ridge Crackdown

The aggression of state prosecutors culminated in the massive Colony Ridge case. While the DOJ participated, the investigation was heavily fueled by state-level consumer protection claims. In February 2026, a $68 million settlement was reached to resolve allegations that the developer used predatory lending tactics specifically targeting Hispanic borrowers with bait-and-switch land sales and high-interest loans. The settlement required not just financial restitution but the construction of physical infrastructure and law enforcement facilities, marking a rare instance where a lending settlement directly funded municipal development to repair the harm caused by predatory financial extraction.

References

• Pennsylvania Office of Attorney General. (2022). AG Shapiro, Federal Partners Announce Historic $20 Million Redlining Settlement.
• Massachusetts Office of Attorney General. (2025). AG Campbell Secures $2. 5 Million Settlement with Student Lender Over AI Bias.
• California Civil Rights Department. (2025). New Protections Against Algorithmic Discrimination in Housing and Employment.
• U. S. Department of Justice. (2024). Justice Department and North Carolina Secure $13. 5 Million Agreement with National Bank.
• Consumer Financial Protection Bureau. (2022). CFPB and DOJ Order Trident Mortgage Company to Pay More Than $22 Million.
• Texas Office of the Attorney General. (2026). Attorney General Paxton Secures $68 Million Settlement Against Colony Ridge.

DOJ Policy Shift: The Termination of Redlining Consent Orders

The Department of Justice’s “Combatting Redlining Initiative,” launched in October 2021, represented the most aggressive federal crackdown on discriminatory lending in history. For nearly four years, the DOJ, in coordination with the Consumer Financial Protection Bureau (CFPB) and the Office of the Comptroller of the Currency (OCC), systematically dismantled the “black box” of proprietary lending algorithms and manual underwriting standards that disenfranchised majority-minority neighborhoods. By January 2025, this initiative had secured over $153 million in relief for communities of color. Yet, mid-2025 marked a distinct pivot in enforcement strategy: the quiet, systematic termination of consent orders for institutions that demonstrated “substantial compliance.”

This shift from public shaming to procedural resolution became undeniable in June 2025. Federal prosecutors moved to terminate consent orders against multiple financial institutions years ahead of schedule. The rationale was purely metric-driven: banks that rapidly deployed loan subsidy funds and corrected their algorithmic disparities were granted early release from federal oversight. This policy evolution suggests a new regulatory social contract—strict liability for initial violations, paired with an accelerated off-ramp for institutions that prove their remediation models work.

The June 2025 Termination Wave

On June 3, 2025, the DOJ filed a motion to terminate its September 2022 consent order with Lakeland Bank ( acquired by Provident Financial Services) more than two years early. The original order required Lakeland to invest $12 million in a loan subsidy fund for Black and Hispanic in the Newark, New Jersey area. By May 2025, the bank had not only disbursed the majority of these funds but had also opened new branches in previously redlined districts. The DOJ’s filing noted that the bank had “reached substantial compliance,” a phrase that would appear in multiple filings throughout the summer of 2025.

A similar pattern emerged with Patriot Bank. In January 2024, the Memphis-based lender agreed to a $1. 9 million settlement to resolve allegations of redlining in majority-Black and Hispanic neighborhoods. Less than 18 months later, in June 2025, the DOJ terminated the order. The speed of this reversal—from enforcement action to exoneration—signals that federal agencies are prioritizing the velocity of restitution over the duration of punishment. Institutions that can operationalize their penalty payments are being rewarded with a clean slate.

Key Redlining Settlements and Status (2021–2025)

The following table details the major settlements under the initiative, highlighting the between the initial financial penalties and the current operational status of the consent orders.

Federal Redlining Enforcement Actions: Financial Impact & Status

Institution	Settlement Date	Settlement Amount	Target Market(s)	Status (as of late 2025)
Trident Mortgage Company	July 2022	$24. 4 Million	Philadelphia, PA; Camden, NJ; Wilmington, DE	Terminated (June 2025)
OceanFirst Bank	Sept 2024	$15. 1 Million	Middlesex, Monmouth, Ocean Counties, NJ	Active
National Bank of Pennsylvania	Feb 2024	$13. 5 Million	Charlotte & Winston-Salem, NC	Active
Lakeland Bank	Sept 2022	$13. 0 Million	Newark, NJ Metro Area	Terminated (June 2025)
Ameris Bank	Oct 2023	$9. 0 Million	Jacksonville, FL	Terminated (2025)
Washington Trust Company	Sept 2023	$9. 0 Million	Rhode Island	Active
Park National Bank	Feb 2023	$9. 0 Million	Columbus, OH	Active
Patriot Bank	Jan 2024	$1. 9 Million	Memphis, TN	Terminated (June 2025)
The Mortgage Firm	Jan 2025	$1. 75 Million	Miami, FL	Active

The Trident Precedent: Non-Bank Compliance

The termination of the Trident Mortgage Company consent order is particularly instructive. As the redlining settlement against a non-bank mortgage lender, the July 2022 agreement carried a record $24. 4 million penalty. Trident, a subsidiary of Berkshire Hathaway, was accused of deliberately avoiding writing mortgages in minority neighborhoods across the Philadelphia metropolitan area. even with the severity of the allegations—which included internal emails containing racial slurs—the DOJ moved to terminate the order in June 2025.

This decision show a serious reality for non-bank lenders: the method for exit is purely financial. Trident had ceased lending operations but fully funded the $18. 4 million loan subsidy program and paid the $4 million civil penalty. The DOJ’s motion to dismiss the case with prejudice confirmed that “substantial compliance” applies even when the entity itself has exited the market, provided the financial restitution to the victimized communities is complete.

Continued Enforcement: The OceanFirst Mandate

While orders are, new enforcement actions remain rigorous. The September 2024 settlement with OceanFirst Bank demonstrates that the DOJ has not softened its intake criteria. OceanFirst agreed to pay over $15 million to resolve allegations that it redlined communities in New Jersey. The complaint detailed how the bank acquired other lenders and subsequently closed branches in majority-minority tracts while expanding in white neighborhoods.

The OceanFirst agreement requires the bank to invest at least $14 million in a loan subsidy fund. Unlike the earlier cases being terminated, OceanFirst is in the early stages of its remediation. The bank must also spend $700, 000 on advertising and outreach and $400, 000 on community partnerships. This bifurcation in the regulatory —active, heavy-handed monitoring for new offenders alongside early releases for remediated banks—creates a clear incentive structure. Lenders are being told that while algorithmic bias and redlining can incur severe penalties, the route to reputational rehabilitation is shorter than previously believed, provided the check clears and the loans flow.

“The Justice Department can continue to hold banks and mortgage companies accountable for redlining and to secure relief for the communities that continue to be harmed by these discriminatory practices.” — Attorney General Merrick B. Garland, September 18, 2024.

The juxtaposition of Garland’s stern rhetoric with the 2025 termination filings reveals the DOJ’s dual strategy. The “Combatting Redlining Initiative” is no longer just about punishment; it is about forcing capital into underserved markets. Once that capital is deployed, the federal government is can to step back, treating the consent orders as temporary corrective method rather than permanent punitive scars.

The Denial Gap: 2025 Baseline Data

As the Interagency AVM Rule took full effect in late 2025, the industry faced a clear statistical reality. Verified Home Mortgage Disclosure Act (HMDA) data analyzed by LendingTree and the Urban Institute for the 2024–2025 reporting period established a ” impact baseline” that federal regulators are using to benchmark algorithmic compliance. The numbers reveal a persistent, structural chasm in credit access that automated systems have thus far failed to close.

In 2024, the national mortgage denial rate for Black applicants stood at 19. 0%, nearly double the 11. 3% rate for the applicant pool. This 7. 7 percentage point gap represents a deterioration from previous years, contradicting industry claims that “colorblind” algorithms would naturally democratize lending. White applicants, by contrast, saw denial rates stabilize near 7. 3% in the same period. The is not a function of raw credit scores; when adjusting for income and debt-to-income (DTI) ratios, Black applicants remained 2. 1 times more likely to be denied than similarly situated White applicants.

Source: 2024-2025 HMDA Data Analysis; Urban Institute; Financial Times.

Metric	Black Applicants	White Applicants	Factor
2024 Denial Rate	19. 0%	7. 3%	2. 6x
Adjusted Denial Likelihood	2. 1x	1. 0x (Baseline)	2. 1x
“Other” Denial Reason	15. 5%	5. 5%	2. 8x
Avg. Denial Reasons per App	1. 22	1. 16	+5. 2%

The “Black Box” of Denial Reasons

A specific metric worrying regulators in the 2026 probe is the proliferation of the “Other” category in algorithmic denial codes. While “Debt-to-Income” (DTI) remains the primary reason for rejection across all groups—accounting for 33. 2% of Black denials versus 24. 9% for the general population—the unclear “Other” classification has surged.

Federal Reserve analysis of 2024 data indicates that Black applicants are 10. 0% more likely to receive a denial based on “Other” or unverifiable reasons compared to White applicants with identical profiles. This statistical anomaly suggests that machine learning models, when unable to find a traditional reason to reject a minority applicant, may be leveraging non-standard variables—such as educational background or “digital exhaust” proxies—to trigger a decline. This “Other” category has become a primary target for the new strict liability enforcement, as it frequently masks the specific algorithmic decision pathways that the October 2025 rule explicitly mandates must be explainable.

Case Study: The Navy Federal

The urgency of the 2026 metrics is underscored by the investigation into Navy Federal Credit Union, which provided a tangible example of algorithmic failure. 2024 data revealed the nation’s largest credit union approved 77. 1% of White mortgage applicants but only 48. 5% of Black applicants. This 29-percentage-point gap even among applicants earning over $140, 000, where White applicants earning under $62, 000 were approved at higher rates.

While Navy Federal attributed the gap to legitimate underwriting factors, the sheer magnitude of the statistical variance—the widest among the top 50 lenders—forced a re-evaluation of “neutral” algorithms. The was not limited to conventional loans; VA loan denial rates for Black borrowers at the institution were 28. 6%, compared to a national average of 13. 6% for Black VA applicants. These numbers serve as the “Exhibit A” in the current federal probe, demonstrating that without the new strict liability standards, algorithmic underwriting defaults to reproducing historical segregation patterns.

Valuation Bias and the Appraisal Gap

Beyond approval rates, the 2026 metrics examine the “appraisal gap” enforced by Automated Valuation Models (AVMs). Freddie Mac’s internal research, part of the public record for the probe, identified that homes in majority-Black neighborhoods were 12. 5% more likely to be appraised the contract price than homes in White neighborhoods.

In 2024, the Federal Housing Finance Agency (FHFA) flagged that automated tools frequently failed to account for “market condition” improvements in minority areas, freezing historical undervaluation into current price estimates. The ” Impact Metrics” for 2026 require lenders to track the ratio of undervaluation by census tract demographics. Early 2025 data shows that while the gap has narrowed slightly to 11. 8% following the rule’s announcement, the algorithmic tendency to devalue Black wealth remains statistically significant.

The Massachusetts Precedent: AI Bias Settlements in Student Lending

On July 10, 2025, the Massachusetts Attorney General’s Office (AGO) established a serious enforcement benchmark for the algorithmic lending sector. Attorney General Andrea Joy Campbell announced a $2. 5 million settlement with Earnest Operations LLC, a Delaware-based student loan refinancer. This action marked the successful application of state consumer protection laws—specifically Massachusetts General Law Chapter 93A—to penalize a lender for “educational redlining” within an automated underwriting model. The settlement dismantled the industry defense that proprietary algorithms are immune from impact liability, setting a template for the federal probes that followed later in the year.

The investigation focused on Earnest’s use of a “Cohort Default Rate” (CDR) variable. This metric, which calculates the average loan default rate of an applicant’s former university, functioned as a racially biased proxy. Because widespread inequities have historically underfunded Historically Black Colleges and Universities (HBCUs) and Hispanic-Serving Institutions (HSIs), graduates from these schools frequently attend institutions with higher aggregate default rates. By weighting this variable heavily, Earnest’s algorithm automatically assigned higher interest rates or rejection flags to Black and Hispanic applicants, even when their individual credit scores and income profiles matched those of white applicants from Ivy League or majority-white institutions.

State investigators also uncovered a “knockout rule” within the software that automatically rejected non-citizens, including permanent with valid green cards, before a human underwriter ever saw the file. This binary code filter violated the Equal Credit Opportunity Act (ECOA) and state anti-discrimination statutes. The AGO found that while the algorithm was marketed as a precision tool for risk assessment, it absence basic fairness testing. Earnest had deployed the model without conducting impact analyses to determine if the CDR variable disproportionately penalized protected classes.

Table 6. 1: The Cost of Educational Redlining (SBPC Forensic Audit Data)

Applicant Profile	University Type	Algorithm APR Offer	Lifetime Cost Penalty
24-Year-Old Analyst	NYU (Majority White)	16. 34%	$0 (Baseline)
24-Year-Old Analyst	Howard University (HBCU)	21. 29%	+$3, 499
24-Year-Old Analyst	New Mexico State (HSI)	19. 23%	+$1, 723

The settlement terms imposed a strict “Fair Lending Monitor” regime on Earnest, a requirement that has since been adopted in federal consent orders. The company was forced to remove the CDR variable and school ranking inputs from its underwriting logic. Furthermore, the agreement mandated the implementation of a corporate governance system specifically for AI. This system requires quarterly “fairness testing” of all predictive models and prohibits the deployment of any new algorithm until it passes a impact review. The AGO’s action criminalized the “launch, fix later” method common in the fintech sector.

This Massachusetts case originated from data initially surfaced by the Student Borrower Protection Center (SBPC), which had previously identified similar disparities in the algorithms of other fintech lenders like Upstart Network. While the CFPB had terminated Upstart’s “No-Action Letter” in 2022—signaling an end to the regulatory sandbox era—it was the Massachusetts settlement in 2025 that provided the financial penalty and enforceable corrective action plan. This precedent stripped lenders of the ability to claim that algorithmic bias was unintentional or unavoidable, establishing a strict liability standard for the software supply chain in student finance.

Automated Valuation Models: The Five Quality Control Standards

The regulatory framework governing Automated Valuation Models (AVMs) fundamentally shifted on October 1, 2025. While the underlying technology of algorithmic appraisal had been evolving for decades, the enforcement of the “Quality Control Standards for Automated Valuation Models” final rule marked the time federal law explicitly mandated how these black-box systems must be governed. Issued jointly by six federal agencies—including the CFPB, OCC, and FHFA—the rule codified requirements originally outlined in the Dodd-Frank Act, transforming voluntary guidance into strict legal obligations for mortgage originators and secondary market issuers.

The core of this regulation is a set of five specific quality control factors. Financial institutions can no longer simply purchase a third-party valuation tool and assume its output is valid. Instead, they are required to maintain rigorous, documented control systems that ensure their models adhere to these five statutory pillars. The four standards focus on technical integrity and operational security, while the fifth—added through the agencies’ discretionary authority—directly the widespread bias inherent in historical housing data.

The Statutory Mandates

The rule requires that any institution using an AVM for credit decisions or securitization determinations regarding a consumer’s principal dwelling must adopt policies designed to meet the following standards:

Quality Control Standard	Operational Requirement
1. High Confidence in Estimates	Institutions must validate that the AVM produces reliable and accurate property valuations. This requires continuous back-testing of model outputs against actual sales data to track error rates and variance.
2. Data Manipulation Protection	Safeguards must be installed to prevent the “gaming” of valuations. This includes securing data inputs from tampering by loan officers or interested parties seeking to values to approve loans.
3. Conflict of Interest Avoidance	The model’s development and operation must be firewalled from the loan production side of the business. Revenue goals cannot influence the tuning of valuation algorithms.
4. Random Sample Testing	Lenders must conduct regular, random audits of AVM valuations. This involves manual reviews or alternative valuation methods to verify that the algorithm is not drifting or producing anomalous results.
5. Nondiscrimination Compliance	The “Fifth Factor”: A distinct requirement to ensure models comply with fair lending laws. This mandates testing for impact and ensuring training data does not perpetuate historical redlining.

The “Fifth Factor” and Strict Liability

The inclusion of the fifth factor—compliance with applicable nondiscrimination laws—represents the most significant expansion of regulatory power in the final rule. While the four factors were explicitly required by the Dodd-Frank Act, the agencies used their statutory authority to add nondiscrimination as an independent quality control standard. This decision, finalized in July 2024, closed a serious loophole: previously, a model could be statistically accurate (meeting standard #1) while still being discriminatory (violating standard #5) if it accurately reproduced biased market conditions.

Under this new regime, accuracy is not a defense for discrimination. If an AVM consistently undervalues properties in majority-Black neighborhoods relative to comparable White neighborhoods, the lender is liable, even if the model is “accurately” reflecting a biased market. This forces institutions to actively mitigate algorithmic bias rather than passively accepting it as a reflection of the data. The rule explicitly applies to secondary market issuers as well, meaning entities like Fannie Mae and Freddie Mac must ensure the AVMs used in their covered securitization determinations adhere to these same rigorous standards.

Post-October 2025, the “black box” defense is legally obsolete. Lenders must be able to open the hood of their valuation engines—or demand their vendors do so—to demonstrate that specific controls are in place to detect and neutralize discriminatory patterns before a valuation ever reaches a borrower.

References

• Office of the Comptroller of the Currency. (2024, July 17). Quality Control Standards for Automated Valuation Models: Final Rule.
• Consumer Financial Protection Bureau. (2024, August 7). Quality Control Standards for Automated Valuation Models. Federal Register.
• Federal Deposit Insurance Corporation. (2024, June 17). Final Rule: Quality Control Standards for Automated Valuation Models.
• Garris Horn LLP. (2025, October 6). Use Automated Valuation Models? Don’t Forget About This New AVM Interagency Rule That Just Took Effect on October 1, 2025.
• KPMG International. (2024, June 1). Automated Valuation Models (AVMs): Interagency Final Rule.

The Black Box Defense: Vendor Liability in Private Litigation

Between 2015 and 2025, a serious legal shift occurred in U. S. federal courts regarding the liability of financial institutions for third-party algorithmic tools. Historically, lenders attempted to shield themselves from fair lending lawsuits by arguing that proprietary algorithms purchased from outside vendors were “black boxes”—complex, unclear systems whose internal logic was unknown and, therefore, legally distinct from the lender’s intent. This defense has collapsed. Federal judges and regulators enforce a standard where ignorance of a vendor’s code offers no immunity from liability for discriminatory outcomes.

The of this defense is clear in the sharp rise of private litigation and the failure of motions to dismiss based on vendor separation. By 2024, discrimination claims had become the second most common basis for class action lawsuits against corporations, with algorithmic bias identified as the single most significant litigation threat to financial institutions. Data from 2025 indicates that Fair Credit Reporting Act (FCRA) lawsuits alone increased by 12. 6% in the five months of the year, driven largely by disputes over automated background checks and credit scoring errors.

Piercing the Corporate Veil

The legal precedent for vendor liability was solidified through key rulings that rejected the separation between a lender and its software provider. Courts have increasingly accepted the theory that third-party vendors act as “agents” of the lender, making the financial institution vicariously liable for the vendor’s digital discrimination. This “agency theory” means that if a bank hires a vendor to filter loan applicants, the bank is responsible for the vendor’s impact just as if it had written the discriminatory code itself.

Key Case Law & Settlements: The Failure of the Black Box Defense (2022–2025)

Case Name	Year	Core Allegation	Outcome / Precedent
Louis v. SafeRent Solutions	2024	Tenant screening algorithm disproportionately penalized Black and Hispanic applicants using housing vouchers.	Motion to Dismiss Denied: Court ruled vendors are subject to the Fair Housing Act. Settled for $2. 275 million.
Connolly v. Lanham	2023	Automated valuation and appraisal bias resulted in a significantly lower home value for a Black couple.	DOJ Intervention: DOJ & CFPB filed a Statement of Interest affirming lenders are liable for relying on discriminatory third-party appraisals.
Mass. AG v. Earnest Operations	2025	AI underwriting model used “educational cohort” data that caused impact on minority borrowers.	$2. 5 Million Settlement: major state enforcement requiring a lender to overhaul AI governance for vendor-supplied models.
Huskey v. State Farm	2023	AI fraud detection algorithm allegedly discriminated against Black claimants.	Impact Claim Proceeded: Court rejected the defense that the algorithm’s complexity shielded the insurer from liability.

The “Agent” Liability Standard

The collapse of the black box defense is rooted in the refusal of courts to allow “complexity” as a valid legal excuse. In Louis v. SafeRent Solutions, the U. S. District Court for the District of Massachusetts ruled that a vendor providing a scoring tool could be held directly liable under the Fair Housing Act, a decision that stripped away the argument that software providers are mere neutral data processors. Similarly, the 2025 settlement between the Massachusetts Attorney General and Earnest Operations established that lenders must actively test and validate third-party models for bias before and during deployment. The settlement mandated that Earnest implement a specific AI governance structure, ending the practice of “deploy and forget.”

Federal regulators have amplified this judicial trend. In 2023, the CFPB and DOJ jointly filed a Statement of Interest in Connolly v. Lanham, explicitly arguing that a lender cannot avoid liability by blaming a third-party appraiser or valuation tool. This regulatory pressure has resulted in a dramatic spike in enforcement referrals; in 2023 alone, federal banking agencies referred 33 fair lending matters to the DOJ, a significant increase from previous years. These referrals frequently involve “redlining” and pricing discrimination facilitated by automated systems, signaling that the federal government views the vendor-lender relationship as a single chain of liability.

Litigation Volume and Success Rates

The success of plaintiffs in surviving early dismissal motions has emboldened class action firms. In Q3 2024, consumer complaints to the CFPB regarding credit reporting—frequently the data fuel for these algorithms—surged by 124% year-over-year. This explosion in consumer dissatisfaction provides the raw material for class action lawsuits. Legal analysts note that the “success rate” for defendants filing motions to dismiss in algorithmic bias cases has plummeted. Judges are routinely allowing impact claims to proceed to discovery, forcing defendants to open their “black boxes” and reveal the training data and weighting variables inside. Once a case reaches discovery, the reputational and financial risks typically force a settlement, as seen in the SafeRent and Earnest cases.

References

Consumer Financial Protection Bureau. (2024, July 10). Fair Lending Report of the Consumer Financial Protection Bureau, 2023. Federal Register.

U. S. District Court for the District of Massachusetts. (2024, July 26). Louis et al. v. SafeRent Solutions, LLC et al., Memorandum and Order on Motion to Dismiss.

Massachusetts Attorney General’s Office. (2025, July 10). AG Campbell Announces $2. 5 Million Settlement with Student Loan Servicer Earnest Operations LLC Over Discriminatory AI Lending Models.

U. S. Department of Justice. (2023, March 13). Statement of Interest of the United States in Connolly v. Lanham.

Bridgeforce Data Solutions. (2024, November 13). 2024 CFPB Complaints and FCRA Litigation Continue to Rise.

Carlton Fields. (2024). 2024 Class Action Survey: Emerging Trends in Litigation.

Michael Best & Friedrich LLP. (2025, July 8). When AI Goes Wrong: Emerging Litigation Trends in Banking Technology Disputes.

Alternative Data Risks: Rent and Utility Inputs as Racial Proxies

The integration of “alternative data”—specifically rent, utility, and telecommunications payment histories—into credit scoring and Automated Valuation Models (AVMs) has been marketed as a panacea for the “credit invisible.” yet, federal probes and independent audits conducted throughout 2024 and 2025 reveal a more volatile reality: these inputs frequently function as high-correlation proxies for race, triggering the very strict liability penalties mandated by the October 2025 AVM rule. Rather than leveling the playing field, these data points frequently digitize and automate historical segregation.

The core method of this bias lies in the “data void.” While mortgage payments are reported to credit bureaus with near-universal consistency, rent and utility reporting remains sporadic and economically stratified. A 2025 analysis by the Urban Institute found that only 13% of renter households had their payments reported to major credit bureaus, a figure that has risen from 11% in 2024 but remains statistically insignificant for broad algorithmic training. Crucially, this reporting is not random; it is concentrated in corporate-managed luxury apartment complexes, leaving out vast swaths of minority tenants living in informal or small-landlord arrangements.

The Utility Trap: Disconnection as a Demographic Marker

Utility payment data poses an even more direct risk for lenders under the new regulatory regime. Unlike credit lines, where non-payment reflects financial distress, utility disconnects and late fees are heavily influenced by infrastructure quality and regional billing policies. In 2024, researchers found that Black and Hispanic households were approximately twice as likely to experience utility shutoffs compared to white households with similar incomes. In specific zip codes in Illinois and Minnesota, this widened to four times the rate.

When AVMs or credit overlays ingest this data, they do not see “late electricity payments”; they see a geographic and racial footprint. An algorithm trained to penalize utility volatility can inadvertently redline neighborhoods with older, less energy- housing stock—neighborhoods that are historically minority-majority. Under the AVM rule’s strict liability standard, a lender using such a model is liable for the impact, regardless of the intent.

Table 9. 1: The Alternative Data Risk Matrix (2024-2025 Data)

Data Input Type	Primary Source	Racial Proxy Risk Factor	Regulatory Liability Trigger
Rent Payments	Property Management Software (e. g., Yardi, RealPage)	High. Reporting is skewed toward Class A luxury rentals. “Pay-to-report” models exclude low-income tenants.	Impact. Penalizes tenants in older/informal housing (disproportionately minority).
Utility Bills	Energy & Telecom Providers	serious. Shutoff rates for Black/Hispanic households are 2x-4x higher than white households.	Proxy Discrimination. Penalizes of energy-inefficient zones (historically redlined areas).
Cash Flow Data	Bank Account Aggregators (e. g., Plaid, Finicity)	Moderate. Volatility in cash flow correlates with gig-economy work, prevalent in minority demographics.	Unfair Practice. May violate ECOA if income stability metrics ignore gig-economy realities.

The “Pay-to-Play” Credit Building Fallacy

A secondary vector of discrimination is the “pay-to-play” nature of rent reporting. services require tenants to pay a monthly subscription fee to have their on-time rent reported to bureaus. This creates a two-tiered system where wealthier renters can buy their way into a higher credit score, while lower-income renters—who are disproportionately Black and Latino—cannot afford the “credit building” tax. A 2025 report by the National Consumer Law Center (NCLC) highlighted that 27. 8% of Black renter households fall at or the poverty line, compared to 18. 2% of white households, making these optional reporting fees a barrier rather than a.

Furthermore, the “positive-only” reporting standard touted by advocates is technically difficult to enforce in machine learning environments. Once a data pipe is opened, algorithms can infer negative behavior from the absence of data (a “null” value) just as easily as from a negative report. If a tenant stops paying the subscription fee, the sudden cessation of data can be interpreted by a model as a cessation of rent payment, unfairly downgrading the borrower.

Chart: The Disconnection

The following chart illustrates the clear racial in utility disconnections, a key alternative data input. The “Disconnection Index” normalizes the rate of utility shutoffs, showing how minority households are disproportionately flagged as “high risk” by algorithms using this data.

This renders raw utility data toxic for compliance. Lenders utilizing “full-file” alternative data without rigorous debiasing filters are importing the structural inequalities of the energy sector into the mortgage market. The CFPB’s 2025 guidance on AVMs explicitly warns that “neutral” inputs that result in racially outcomes can be treated as evidence of a defective compliance management system.

References

• Urban Institute. (2025). The Rise of Rent Reporting: 2025 Market Analysis.
• National Consumer Law Center (NCLC). (2025). The Hidden Cost of Credit: Alternative Data Risks in Lending.
• Consumer Financial Protection Bureau (CFPB). (2024). Fair Lending Report: Utility Data and Impact.
• Energy Equity Project. (2024). Racial Disparities in Utility Disconnections: A National Study.
• Joint Center for Housing Studies of Harvard University. (2024). America’s Rental Housing 2024.

The Explainability emergency: Agentic AI vs Adverse Action Notices

By late 2025, a fundamental incompatibility emerged between the Federal Reserve’s Regulation B and the newest generation of “agentic” lending models. While federal law mandates that lenders provide specific, principal reasons for denying credit, the autonomous AI agents deployed by major fintechs frequently make decisions based on non-linear correlations that even their developers cannot fully articulate. This “black box” problem has transitioned from a theoretical risk to a primary driver of enforcement actions.

The core of the emergency lies in the “principal reason” requirement of the Equal Credit Opportunity Act (ECOA). When a human underwriter denies a loan, they can point to a high debt-to-income ratio or a delinquency. When an agentic AI denies a loan, it may do so because a constellation of thousands of variables—ranging from cash flow volatility to device metadata—formed a negative predictive pattern. Translating this complex vector into a standard adverse action notice frequently results in vague, non-compliant explanations.

Regulatory Crackdown: The “Model Score” Fallacy

For years, lenders attempted to satisfy adverse action requirements by listing “insufficient model score” or “unverifiable data” as denial reasons. In 2025, regulators explicitly rejected this practice. The CFPB’s Winter 2025 Supervisory Highlights revealed that examiners had multiple auto lenders for failing to provide accurate denial reasons when using AI/ML models. The Bureau found that institutions were selecting reasons from a standard checklist that did not accurately reflect the actual factors driving the AI’s decision, concealing the true basis of the denial from the consumer.

This regulatory stance was cemented by the Massachusetts Attorney General’s landmark July 2025 settlement with student loan lender Earnest Operations LLC. The $2. 5 million settlement resolved allegations that the company’s AI underwriting models produced discriminatory outcomes and, crucially, failed to provide accurate adverse action notices. The AG’s office alleged that the lender’s automated systems sent “inaccurate adverse action notices that prevented applicants from understanding their own creditworthiness,” marking one of the state-level enforcement actions to directly target the explainability of AI lending decisions.

The Compliance Gap

Industry data confirms that vague denial reasons have become a leading source of consumer friction. A October 2025 report by compliance firm Ncontracts identified “vague reasons for denial” as a top complaint category for credit unions, linking the surge directly to system failures in translating AI outputs into regulatory disclosures. Examiners from the Federal Reserve Banks of Minneapolis and Chicago echoed this in a July 2025 Consumer Compliance Outlook session, warning that generic explanations like “outside of risk tolerance” are legally insufficient.

2025 Enforcement Actions & Findings: AI Explainability

Regulatory Body	Date	Target / Sector	Key Finding / Action
Massachusetts AG	July 2025	Student Lending (Earnest Ops)	$2. 5M settlement for AI bias and inaccurate adverse action notices.
CFPB Supervision	Jan 2025	Auto Lenders	Examiners lenders for using “checklist” reasons that did not match AI model inputs.
Federal Reserve	July 2025	Banking Sector	Examiners warned that “credit score policy” is insufficient for AI-driven denials.
CFPB	Nov 2024	Non-Bank Lender	Settlement regarding algorithmic redlining and failure to maintain adequate records.

The shift toward “agentic” AI—systems that not only score applicants but autonomously execute decision workflows—has exacerbated the problem. Unlike static regression models, these agents continuously update their weighting based on real-time data flow. A variable that was determinative on Tuesday might be negligible on Wednesday. This fluidity makes the static “reason code” framework of legacy adverse action software obsolete. Lenders are legally required to “test and validate” that their specific denial reasons are accurate for each individual decision, a technical hurdle that legacy systems cannot clear.

Regulators have made their position clear: the complexity of the technology is not a defense for non-compliance. As stated in CFPB Circular 2023-03, if a creditor cannot accurately identify the specific reasons for a denial because their model is too complex, they cannot legally use that model. This strict liability standard forces lenders to choose between throttling their AI’s capabilities or facing aggressive enforcement for explainability failures.

References

• Consumer Financial Protection Bureau. (2025, January). Supervisory Highlights, problem 38, Winter 2025.
• Office of the Attorney General of Massachusetts. (2025, July 10). AG Campbell Announces $2. 5 Million Settlement With Student Loan Lender For Unlawful Practices Through AI Use.
• Ncontracts. (2025, October 30). The Top Four Credit Union Member Complaints in 2025.
• Federal Reserve Bank of Minneapolis & Chicago. (2025, July 17). Consumer Compliance Outlook: Adverse Action Notifications – Examiner Insights.
• Consumer Financial Protection Bureau. (2023, September 19). Consumer Financial Protection Circular 2023-03: Adverse action notification requirements and the proper use of the CFPB’s sample forms.

Nonbank Lenders: The Shift in Regulatory Focus

By late 2025, the regulatory perimeter for mortgage lending had fundamentally fractured. While the Interagency AVM Rule established a strict federal baseline for algorithmic valuation, the broader enforcement for nonbank lenders—who originate 66. 4% of all U. S. mortgages—shifted dramatically from Washington, D. C., to state capitals. This decentralization was accelerated by the Consumer Financial Protection Bureau’s (CFPB) abrupt October 29, 2025, rescission of its “Registry of Nonbank Covered Persons,” a database intended to track repeat corporate offenders. The Bureau “speculative benefits” and excessive compliance costs for the reversal, a key federal oversight method for the fintech and independent mortgage bank (IMB) sector.

The federal retreat created an immediate vacuum that state attorneys general (AGs) moved aggressively to fill. In the absence of a unified federal “repeat offender” registry, states began deploying their own consumer protection statutes to target algorithmic discrimination and unfair lending practices. This pivot has replaced a singular federal compliance standard with a “patchwork of peril” for national lenders, where an algorithm compliant in Texas might trigger a civil rights lawsuit in Massachusetts or New York.

The “State-” Enforcement Doctrine

The new enforcement reality was crystallized in July 2025, when the Massachusetts Attorney General secured a $2. 5 million settlement against Earnest Operations LLC. While the case involved student lending, it established a serious legal precedent for the mortgage industry: the use of ” impact” theory to prosecute algorithmic bias at the state level, even as federal agencies under a new administration signaled a retreat from such theories. The AG alleged that the lender’s AI underwriting model, which factored in metrics like “school rank” and “cohort default rates,” disproportionately penalized Black and Hispanic applicants. This settlement served as a warning shot to nonbank mortgage lenders using similar “alternative data” in their proprietary credit models.

New York followed suit with legislative force. On November 10, 2025, the “Algorithmic Pricing Disclosure Act” took effect, mandating that any company using automated systems to set individualized prices or rates must explicitly disclose this to consumers. For nonbank mortgage lenders, who frequently use pricing engines to adjust interest rates based on real-time liquidity and borrower profiles, this law introduces a new of transparency risk. Furthermore, the state’s “FAIR Business Practices Act,” signed in late 2025 and February 2026, expanded the Attorney General’s power to prosecute “abusive” acts—a standard previously reserved largely for federal regulators.

Table 11. 1: Key State Regulatory Actions Targeting Nonbank Lenders (2025-2026)

Jurisdiction	Action / Legislation	Date	Impact on Nonbank Lenders
Massachusetts	Earnest Operations Settlement	July 2025	Established state-level precedent for prosecuting algorithmic impact in underwriting models.
California	ADMT Regulations (CPPA)	Sept. 2025 (Finalized)	Classifies lending decisions as “significant,” requiring opt-out rights and transparency for automated decision-making.
New York	Algorithmic Pricing Disclosure Act	Nov. 10, 2025	Requires “clear and conspicuous” disclosure if an algorithm determines consumer pricing or rates.
Pennsylvania	RESPA Kickback Litigation	April 2025	AG sued mortgage brokers for undisclosed referral fees, signaling renewed focus on nonbank fee structures.

California’s Automated Decision-Making Mandate

The most sweeping regulatory challenge for nonbanks has emerged from California. In September 2025, the California Privacy Protection Agency (CPPA) finalized its regulations on Automated Decision-Making Technology (ADMT). Unlike federal rules that frequently focus on outcomes (discrimination), California’s framework the process. January 1, 2027, but requiring immediate compliance architecture, these rules classify lending decisions as “significant,” granting consumers the right to opt out of automated processing in favor of human review. For fintech lenders whose business models rely on 100% automated underwriting to maintain low margins, this requirement poses an existential operational challenge.

The is clear. While the CFPB’s October 2025 rescission relieved nonbanks of federal registration load, it simultaneously removed the “safe harbor” of a unified federal standard. Lenders must navigate a balkanized regulatory map where New York demands disclosure, California demands opt-out rights, and Massachusetts prosecutes statistical bias. This shift transfers the duty of “fairness” from federal examiners to state prosecutors, who are increasingly using their own “mini-CFPB” authority to police the black box of nonbank lending.

References

Consumer Financial Protection Bureau. (2025, October 29). CFPB Rescinds Rule Creating Registry for Nonbank Enforcement Actions. Federal Register.
Massachusetts Attorney General’s Office. (2025, July 10). AG Campbell Reaches $2. 5 Million Settlement with Earnest Operations Over Algorithmic Discrimination.
New York State Office of the Attorney General. (2025, November 5). Consumer Alert: New Algorithmic Pricing Disclosure Law Takes Effect.
California Privacy Protection Agency. (2025, September 25). Final Regulations on Automated Decision-Making Technology.
Inside Mortgage Finance. (2025, June 26). Nonbank Share of Originations Hits 66. 4% in Q1 2025.

The 80 Percent Rule: Statistical Failures in Modern Underwriting

For decades, federal regulators have relied on a simple arithmetic heuristic to detect discrimination: the “four-fifths” or “80 percent” rule. Under this standard, originally codified in the Uniform Guidelines on Employee Selection Procedures and adapted for credit enforcement, a selection rate for any protected group that is less than 80 percent of the rate for the group with the highest rate is generally regarded as evidence of adverse impact. In the era of algorithmic underwriting, this threshold has ceased to be a mere guideline and has become a statistical tripwire that modern “black box” models trigger with worrying frequency.

Between 2022 and 2025, major lending institutions repeatedly failed this test, exposing a widespread inability of machine learning models to self-correct for impact without explicit intervention. The most example emerged from a December 2023 investigation into Navy Federal Credit Union, the nation’s largest credit union. An analysis of 2022 Home Mortgage Disclosure Act (HMDA) data revealed that Navy Federal approved 77. 1 percent of White mortgage applicants but only 48. 5 percent of Black applicants. The resulting ratio—62. 9 percent—fell well the 80 percent compliance threshold, signaling a severe statistical failure that triggered class-action litigation and congressional inquiries in early 2024.

These disparities are not to a single institution. In March 2022, a similar pattern was identified at Wells Fargo during the mortgage refinancing boom. The bank approved 72 percent of White homeowners for refinancing in 2020, compared to just 47 percent of Black homeowners. This 65 percent ratio again violated the four-fifths standard. In both cases, the lenders attributed the disparities to “legitimate, credit-related factors” buried within their proprietary scoring models—specifically, Navy Federal’s internal risk assessments and Wells Fargo’s “Enhanced Credit Score.”

Table 12. 1: Algorithmic Approval Rate Disparities (2020-2022 Data)

Institution	Loan Product	White Approval Rate	Black Approval Rate	Ratio (Black/White)	80% Rule Status
Navy Federal Credit Union	Conventional Mortgage	77. 1%	48. 5%	62. 9%	FAIL
Wells Fargo	Refinance	72. 0%	47. 0%	65. 3%	FAIL
Navy Federal Credit Union	Latino Applicants	77. 1%	55. 8%	72. 4%	FAIL

Source: Ekalavya Hansaj News Network analysis of HMDA data and 2022-2023 investigative reports. Ratios 80% indicate chance impact.

The persistence of these failures highlights a fundamental flaw in how algorithms select variables. Unlike human underwriters who are trained to avoid redlining, machine learning models optimize purely for predictive accuracy, frequently latching onto data points that serve as proxies for race. On July 10, 2025, the Massachusetts Attorney General reached a landmark $2. 5 million settlement with Earnest Operations, a student loan servicer, over precisely this problem. The investigation found that Earnest’s AI underwriting model used a “Cohort Default Rate”—a metric based on the average default rate of the applicant’s college—as a key variable. Because Black and Hispanic borrowers are statistically more likely to attend institutions with higher default rates due to widespread funding inequities, the algorithm penalized them regardless of their individual creditworthiness.

The Earnest settlement also revealed the use of a “Knockout Rule,” where the algorithm automatically rejected applicants who did not possess a green card. This binary filter, hidden deep within the decision tree, created a impact on non-citizen borrowers that the model viewed as risk avoidance rather than illegal discrimination. This case marked a turning point in 2025, moving enforcement from analyzing output statistics to auditing the specific input variables of the algorithms themselves.

Federal regulators have responded by attacking the opacity that allows these statistical failures to go unchecked. In September 2023, the Consumer Financial Protection Bureau (CFPB) issued Circular 2023-03, explicitly closing the loophole that allowed lenders to use generic “checkbox” explanations for algorithmic denials. The guidance mandates that lenders must provide the specific reasons for an adverse action, even if those reasons are generated by a complex “black box” model. If a lender cannot explain why an applicant was rejected because the algorithm is too complex, they cannot legally use that algorithm to make credit decisions. This requirement forces institutions to confront the 80 percent rule failures not just as a compliance nuisance, but as a defect in their model’s explainability and fairness architecture.

References

• Consumer Financial Protection Bureau. (2023, September 19). Consumer Financial Protection Circular 2023-03: Adverse action notification requirements and the proper use of the CFPB’s sample forms provided in Regulation B.
• Massachusetts Office of the Attorney General. (2025, July 10). AG Campbell Announces $2. 5 Million Settlement with Student Loan Servicer Earnest Over Discriminatory AI Underwriting.
• Tolan, C., Ash, A., & Marsh, R. (2023, December 14). The nation’s largest credit union rejected more than half its Black conventional mortgage applicants. CNN.
• Levitt, A. (2023, December 18). Class Action Complaint: Oates et al. v. Navy Federal Credit Union. United States District Court for the Eastern District of Virginia.
• Bloomberg News. (2022, March 11). Wells Fargo Approved Fewer Than Half of Black Homeowners’ Refinancing Applications in 2020.

Digital Redlining Maps: Visualizing Denial Rates by Zip Code

The geography of modern lending discrimination is no longer defined by physical red ink on paper maps, but by invisible algorithmic boundaries that enforce exclusion with surgical precision. Analysis of Home Mortgage Disclosure Act (HMDA) data from 2024 and 2025 reveals that while the method of denial have digitized, the outcomes remain clear territorial. In specific “kill zones”—zip codes and metropolitan areas flagged by automated valuation models (AVMs) as high-risk—denial rates for Black and Latino borrowers have decoupled from their creditworthiness, creating a digital redlining map that mirrors the segregation patterns of the 1930s.

Federal scrutiny has intensified following the release of 2025 data showing that algorithmic penalties are not evenly distributed. The “black box” systems used by major lenders frequently penalize applicants in majority-minority neighborhoods by weighing “neighborhood risk factors” more heavily than individual financial health. This results in a paradox where high-earning Black applicants in Detroit or Grand Rapids face rejection rates significantly higher than low-income white applicants in adjacent suburbs. The data the industry defense that credit scores alone drive these decisions; instead, location data serves as a proxy for race, triggering automatic denials before a human underwriter ever reviews the file.

The Geography of Rejection: 2024-2025 Data

Recent investigations have specific metropolitan areas where the gap in approval rates is most egregious. A July 2025 analysis by LendingTree, grounded in HMDA filings, identified Grand Rapids, Michigan, as the national epicenter of this. In this metro area, Black applicants faced a denial rate of 23. 9%, compared to just 14. 15% for white applicants—a gap of nearly 10 percentage points that cannot be explained by income variance alone. Detroit followed closely, with a 21. 25% denial rate for Black borrowers versus 12. 71% for white borrowers.

The is not limited to the Rust Belt. Raleigh, North Carolina, a booming tech hub, registered the third-highest gap in the nation, with Black applicants facing an 8. 44 percentage point disadvantage. In these markets, the “digital map” walls off entire neighborhoods from capital. For instance, in Detroit’s 48224 zip code (the Finney neighborhood), high origination volumes mask a troubling trend: while applications are high, denial rates for conventional loans remain elevated, forcing borrowers into higher-cost FHA products or predatory land contracts.

Table 13. 1: Mortgage Denial Rate Disparities in Key U. S. Metros (2024-2025)

Metropolitan Area	Black Applicant Denial Rate	White Applicant Denial Rate	Gap
Grand Rapids, MI	23. 90%	14. 15%	+9. 75%
Detroit, MI	21. 25%	12. 71%	+8. 54%
Raleigh, NC	17. 84%	9. 40%	+8. 44%
Philadelphia, PA	19. 20%	12. 80%	+6. 40%
National Average	19. 00%	11. 27%	+7. 73%

The High-Income Anomaly

The most damning evidence against algorithmic neutrality lies in the treatment of high-income minority borrowers. Data from Detroit Future City updated in late 2025 indicates that high-income Black applicants in the Detroit metro area were denied at rates four times higher than their white counterparts with similar income profiles. This anomaly suggests that AVMs and underwriting algorithms are programmed to view the combination of race and geography as a risk factor. When a high earner attempts to buy in a zip code like 48221 (Bagley), the algorithm frequently flags the property value as “volatile” or “declining” based on historical data rooted in prior discrimination, circularizing the bias.

In Philadelphia, the situation is equally grim. A 2025 report by the Reinvestment Fund highlighted that while suburban refinance applications have surged, the city’s majority-Black neighborhoods have seen a 64% drop in mortgage applications over the last two decades, driven by a pattern of denial and discouragement. The algorithms punish these zip codes for “insufficient collateral” or “high debt-to-income” ratios that are frequently inflated by the system’s own undervaluation of the properties in question. This creates a self-fulfilling prophecy: the algorithm predicts risk, denies the loan, depresses the property value, and validates its own prediction for the applicant.

Foreclosure Zones as Proxy

The digital redlining map also overlaps significantly with areas of high foreclosure activity, which algorithms use as a negative signal for new lending. In Q1 2024, zip codes such as 77327 in Cleveland, Texas, and 93222 in Frazier Park, California, reported of the highest foreclosure rates in the nation. Lenders using automated risk scoring frequently blanket-ban or severely restrict conventional lending in these zones. Consequently, creditworthy borrowers in these zip codes are guilty by association, denied access to prime rates solely because their neighbors are in distress. This “guilt by geography” is the hallmark of digital redlining, replacing the loan officer’s bias with the cold, unassailable logic of the code.

References

• LendingTree. (2025, July 15). Racial Gaps In Mortgage Denials even with Industry Progress.
• Detroit Future City. (2025, October 13). Detroit Mortgage Lending: Continuous tracking of mortgage activity since 2019.
• Reinvestment Fund. (2025, June 27). Mortgage Lending Activity in the Philadelphia Metro Area.
• Inside Mortgage Finance. (2025, May 27). Denial Rates for Mortgages Decline in 2024.
• Safeguard Properties. (2024, April 12). Top 10 ZIPS with Highest Foreclosure Rates in Q1 and March 2024.

Interest Rate Discrimination: The New Frontier of Bias

For decades, the primary method of financial exclusion was the denial letter. In 2026, the method has shifted. Lenders grant access to credit with greater frequency, but they do so at a predatory premium. This phenomenon, known as “algorithmic pricing discrimination,” replaces the binary “yes or no” of traditional redlining with a sliding of interest rates that systematically penalizes Black and Latino borrowers. While automated underwriting systems are frequently touted as race-neutral, data from 2024 and 2025 confirms that these models impose a “segregation tax” on minority borrowers, charging them higher rates for the exact same risk profiles as their White counterparts.

The is most visible in the mortgage sector, where the pledge of “colorblind” fintech lending has failed to materialize. A February 2025 study published in MDPI analyzed Home Mortgage Disclosure Act (HMDA) data from 2018 through 2023. The findings the narrative that technology eliminates bias. While fintech lenders rejected fewer minority applicants than traditional banks, they charged significantly higher prices. Black borrowers faced a rate spread gap of nearly 16 basis points compared to White borrowers with similar credit characteristics. Traditional credit unions performed even worse, with a gap of 27. 6 basis points. This pricing differential is not a reflection of risk; it is a reflection of algorithmic targeting.

Navy Federal Credit Union, the nation’s largest credit union, became a focal point of this controversy following a class-action lawsuit and federal scrutiny in late 2023 and 2024. Analysis revealed that the institution approved more than 75 percent of White applicants for conventional home purchase loans but denied more than 50 percent of Black applicants. For those Black borrowers who did secure approval, the cost of borrowing was frequently higher. The 2025 MDPI analysis corroborates this broader industry trend, showing that minority borrowers are funneled into higher-cost products even when they qualify for prime rates.

The Cost of Algorithmic Bias: Interest Rate Markups by Sector (2024-2025 Data)

Lending Sector	Targeted Demographic	Interest Rate/Markup Penalty	Estimated Financial Impact
Small Business	Black Owners	+3. 09 percentage points	$8 billion annual aggregate excess interest
Small Business	Hispanic Owners	+2. 91 percentage points	Reduced capital for expansion
Auto Loans	Minority Borrowers	+2. 6% dealer markup	~$1, 400 extra paid over loan life
Mortgage (Fintech)	Black Borrowers	+16 basis points spread	Higher monthly payments & origination fees
Mortgage (Credit Union)	Black Borrowers	+27. 6 basis points spread	Significant loss of equity over 30 years

The auto lending market exhibits an even more aggressive form of pricing discrimination. Unlike mortgages, where rates are somewhat constrained by government-sponsored enterprise (GSE) standards, auto lenders and dealers operate with considerable discretion. A February 2024 study published in Marketing Science analyzed transaction-level data and found that minority borrowers pay a statistically significant 2. 6 percent higher dealer markup than non-minorities. This markup is separate from the base interest rate set by the lender; it is pure profit added by the dealer, frequently facilitated by automated software that suggests the maximum rate a borrower is likely to accept. The Chicago Federal Reserve estimated that these markups cost Black borrowers an average of $1, 400 in additional interest over the life of a loan.

Small business lending shows the steepest penalties. A December 2024 study from the University of Washington’s School of Business found that Black-owned businesses are charged interest rates 3. 09 percentage points higher than White-owned businesses with identical creditworthiness. Hispanic-owned firms pay 2. 91 percentage points more. This extraction of capital—estimated at $8 billion annually—directly impedes the ability of minority-owned firms to hire, expand, or weather economic downturns. The algorithms used to price these loans frequently rely on proxy data, such as zip code-level economic indicators, which penalize business owners for operating in minority-majority neighborhoods.

“The findings suggest that LLMs [Large Language Models] are learning from the data they are trained on, which includes a history of racial disparities in mortgage lending… Black applicants would, on average, need credit scores approximately 120 points higher than White applicants to receive the same approval rate.” — Donald Bowen III, Assistant Professor of Finance, Lehigh University (August 2024)

The integration of Artificial Intelligence into these pricing models has accelerated the problem. In August 2024, researchers at Lehigh University tested leading commercial Large Language Models (LLMs) used in financial decision-making. The results were clear: the AI models consistently recommended higher interest rates for Black applicants compared to White applicants with identical financial profiles. To secure the same interest rate as a White borrower, a Black applicant needed a credit score 30 points higher. This “technology tax” is invisible to the consumer, buried in the proprietary code of the lender’s pricing engine, yet it determines the financial trajectory of millions of American families.

Human Oversight Failures: Automation Bias

The financial sector’s primary defense against algorithmic discrimination—the “human-in-the-loop” (HITL) protocol—has proven to be a catastrophic failure. Lenders have long argued that deploying human underwriters to review AI-generated loan denials would act as a fail-safe against bias. yet, federal investigations and academic studies conducted between 2024 and 2025 reveal the opposite: human reviewers frequently function as rubber stamps, exhibiting “automation bias” where they defer to the machine’s judgment even when that judgment is demonstrably flawed. Rather than correcting algorithmic errors, human oversight has calcified them, providing a veneer of due process to discriminatory outputs.

This phenomenon was quantified in an August 2024 study by Lehigh University, which found that commercial Large Language Models (LLMs) used in underwriting consistently recommended denying loans to Black applicants at significantly higher rates than identical White applicants. The study exposed a serious weakness in HITL systems: when presented with a complex, data-rich AI recommendation, human reviewers absence the time or technical granularity to challenge the machine’s logic. Consequently, the “review” becomes a procedural formality. The study noted that while simple prompts to “ignore bias” could reduce disparities, human operators rarely intervened to adjust these parameters, allowing the algorithms to redline by proxy.

The consequences of this oversight failure are visible in the clear metrics reported by major institutions. In 2024, a class-action lawsuit against Navy Federal Credit Union—the nation’s largest credit union—alleged that the institution denied 52% of Black mortgage applicants compared to only 23% of White applicants. even with the presence of human underwriters who ostensibly validated these decisions, the. The lawsuit highlighted that the institution’s “hybrid” approval process did not mitigate the racial gap; instead, the human element likely reinforced the algorithmic sorting, as underwriters were incentivized to process volumes quickly rather than conduct forensic audits of borderline rejections.

Comparative Denial Rates in Hybrid Underwriting Systems (2024-2025)
Source: Financial Times Analysis (Sept 2025), LendingTree Report (July 2025)

Metric Category	White Applicant Denial Rate	Black Applicant Denial Rate	Multiplier
National Average (LendingTree)	11. 0%	19. 0%	1. 7x
Navy Federal (Lawsuit Data)	23. 0%	52. 0%	2. 3x
Adjusted for Income/Profile (FT)	N/A	N/A	2. 1x

Regulatory bodies have begun to the “human review” defense. In July 2025, the Massachusetts Attorney General secured a landmark $2. 5 million settlement with Earnest Operations, a student loan servicer. The investigation found that Earnest’s AI underwriting models produced discriminatory outcomes for Black and Hispanic applicants. Crucially, the Attorney General noted that the company failed to mitigate these risks even with having internal governance structures in place. This action marked a turning point: regulators are no longer accepting the mere existence of a compliance team as evidence of fair lending. If the output is biased, the human oversight is deemed legally insufficient.

The Consumer Financial Protection Bureau (CFPB) reinforced this stance with its September 2024 circular, which explicitly targeted the “black box” defense. The Bureau clarified that lenders cannot deny credit based on unclear algorithmic scores without providing specific, accurate reasons for the adverse action. The circular warned that a human underwriter citing a “proprietary model score” does not satisfy the Equal Credit Opportunity Act. This directive criminalizes the passive acceptance of AI decisions, forcing lenders to understand—and be able to explain—exactly why a human reviewer agreed with an algorithm’s rejection.

By late 2025, the data confirmed that human intervention without rigorous, bias-aware training is futile. A September 2025 analysis by the Financial Times of nearly 40 million mortgage applications found that even after adjusting for income and credit profiles, Black applicants remained 2. 1 times more likely to be denied than their White counterparts. The persistence of this gap, even with universal claims of human oversight, indicates that the industry’s current HITL frameworks are not safety method but liability shields, designed to diffuse responsibility rather than ensure equity.

Regulatory Arbitrage: Banks Fleeing Strict State Jurisdictions

As state-level enforcement of algorithmic bias laws intensifies, a distinct pattern of regulatory arbitrage has emerged across the U. S. banking sector. Financial institutions and fintech lenders are increasingly maneuvering to bypass strict state jurisdictions—most notably New York, California, and Colorado—by seeking federal preemption or relocating their legal charters to states with more permissive regulatory environments. This “flight to safety” is not physical but legal, leveraging the “true lender” doctrine and federal charters to immunize algorithmic underwriting models from aggressive state attorneys general.

The primary driver of this exodus is the between stalling federal standards and rapidly enacting state laws. While the federal “Quality Control Standards for Automated Valuation Models” rule took effect in October 2025, it largely codified existing prohibitions without the granular testing requirements found in new state legislation. In contrast, the Colorado Artificial Intelligence Act, set to take full effect on February 1, 2026, mandates that developers and deployers of “high-risk” AI systems—including credit underwriting models—complete rigorous impact assessments and disclose foreseeable risks of algorithmic discrimination to the Attorney General. Similarly, California’s Department of Financial Protection and Innovation (DFPI) has aggressively enforced its consumer protection laws against crypto and fintech lenders, signaling a zero-tolerance method to “black box” decision-making.

To evade these patchworks, fintechs are accelerating their of national charters. By obtaining a national bank or trust charter from the Office of the Comptroller of the Currency (OCC), institutions can that federal banking laws preempt state consumer financial protection laws. Data from late 2025 indicates a surge in this activity, with the OCC reporting six pending applications for limited-purpose national trust bank charters in December 2025 alone. Conditional approvals for entities like National Digital Currency Bank and National Trust Bank in late 2025 show the federal regulator’s willingness to bring these entities under its fold, shielding them from state-level “mini-CRA” (Community Reinvestment Act) laws and AI bias audits.

The “Rent-a-Bank” Loophole and AI Preemption

For fintechs unable to secure a full national charter, the “rent-a-bank” model remains the preferred vehicle for regulatory arbitrage. In these arrangements, a non-bank fintech partners with a state-chartered bank in a deregulation-friendly jurisdiction—frequently Utah or Kentucky—to originate loans. The fintech then claims the bank’s federal preemption privileges (under the National Bank Act or Depository Institutions Deregulation and Monetary Control Act) apply to its algorithmic pricing and underwriting, nullifying strict interest rate caps and fair lending statutes of the borrower’s home state.

Cross River Bank and WebBank have frequently been in regulatory filings and industry reports as key partner banks in this ecosystem. yet, this model faces increasing peril. In May 2023, the FDIC issued a consent order against Cross River Bank, citing unsafe and unsound practices related to fair lending and third-party oversight. This action signaled that federal regulators are beginning to pierce the corporate veil of these partnerships, holding the partner bank strictly liable for the discriminatory outcomes of the fintech’s AI models.

The legal ground for this arbitrage was further destabilized by the Supreme Court’s unanimous decision in Cantero v. Bank of America (May 2024). The Court rejected a “bright-line” test for preemption, ruling instead that courts must conduct a ” comparative analysis” to determine if a state law “significantly interferes” with a national bank’s powers. This ruling invited state attorneys general to test the limits of their authority, leading to a spike in state-level enforcement actions.

State vs. Federal Regulatory Conflict (2024–2025)

The following table outlines the key areas of conflict where financial institutions are attempting to use federal preemption to avoid state-level algorithmic accountability.

Table 16. 1: State vs. Federal Regulatory Conflicts in Algorithmic Lending (2024–2025)

Regulatory Domain	Strict State Standard (Example)	Federal Preemption Argument	Recent Legal/Regulatory Outcome
Algorithmic Bias Audits	Colorado AI Act (2026): Mandates impact assessments and disclosure of discrimination risks to AG.	OCC Preemption: National banks claim state audit rules interfere with federal examination powers.	Executive Order (Dec 2025): Establishes “AI Litigation Task Force” to challenge state laws deemed to load interstate commerce.
Fair Lending Enforcement	Massachusetts (2025): AG settled for $2. 5M with student lender over AI model impact.	“True Lender” Doctrine: Fintech claims partner bank’s charter preempts state fair lending suits.	FDIC Consent Order (2023): Held partner bank (Cross River) liable for third-party fair lending compliance, weakening the shield.
Interest Rate/Pricing Caps	New York/California: Strict usury caps and “junk fee” prohibitions.	Interest Rate Exportation: Banks export home state’s (e. g., Utah) looser rates to all borrowers.	Cantero v. BoA (2024): SCOTUS ruled preemption is not automatic; requires proof of “significant interference.”
Consumer Data Rights	California (CCPA/CPRA): Grants consumers right to opt-out of automated decision-making.	FCRA Preemption: Banks federal credit reporting laws override state data privacy rules.	CFPB Rulemaking (Section 1033): finalized rules in 2024 reinforce consumer data access, aligning closer to state standards.

The tension culminated in December 2025, when the White House issued an Executive Order titled “Ensuring a National Policy Framework for Artificial Intelligence.” This order directed the Department of Justice to establish an AI Litigation Task Force specifically to challenge state AI laws that “unconstitutionally regulate interstate commerce.” This move was widely interpreted as a federal lifeline to the banking industry, offering a chance escape route from the tightening noose of state-level algorithmic accountability. yet, with the Cantero precedent in place, the success of this federal preemption strategy remains legally uncertain, leaving banks in a precarious position between aggressive state enforcers and a shifting federal.

References

• Supreme Court of the United States. (2024, May 30). Cantero v. Bank of America, N. A., 602 U. S. ___ (2024).
• Federal Deposit Insurance Corporation (FDIC). (2023, March 8). Consent Order, In the Matter of Cross River Bank, Teaneck, New Jersey. FDIC-22-0140b.
• The White House. (2025, December 11). Executive Order on Ensuring a National Policy Framework for Artificial Intelligence.
• Office of the Comptroller of the Currency (OCC). (2025, December 12). OCC Conditionally Approves Conversions to National Trust Bank Charters. NR 2025-112.
• Colorado General Assembly. (2024, May 17). SB24-205: Consumer Protections for Artificial Intelligence. February 1, 2026.
• Massachusetts Office of the Attorney General. (2025, July 10). AG Campbell Announces $2. 5 Million Settlement with Student Loan Lender to Resolve Allegations of Discriminatory AI Underwriting.

The Whistleblower Pipeline: Leaks in the Absence of Federal Oversight

For much of the decade preceding the 2025 AVM rule, federal regulators operated with a significant blind spot: they could audit the paperwork of a loan, but they could not read the code that approved it. As lenders migrated from human underwriters to “black box” neural networks, the primary method for discovering discriminatory drift shifted from routine examination to insider leaks. By 2021, the Consumer Financial Protection Bureau (CFPB) openly acknowledged this enforcement gap, pivoting its strategy to actively recruit disgruntled engineers and data scientists as the line of defense against algorithmic redlining.

This reliance on the “whistleblower pipeline” became official policy on December 15, 2021, when the CFPB issued a direct appeal to technology workers. The agency’s directive was blunt: engineers who observed their code being “misused or abused for unlawful ends” were encouraged to bypass internal compliance channels and report directly to federal authorities. This marked a shift in financial oversight. Regulators admitted that without the technical keys to decipher proprietary underwriting models, they required insiders to expose the discriminatory variables buried within terabytes of training data.

The urgency of this pipeline is illustrated by the structural opacity of the algorithms involved. In cases, not even the lenders deploying the software understood why a specific applicant was rejected. The “Lookalike Audience” tool employed by Meta (formerly Facebook) serves as the definitive case study of this era. It was not a routine audit that dismantled this discriminatory method, but a years-long pressure campaign fueled by investigative reporting and internal scrutiny that eventually forced the Department of Justice to intervene. In June 2022, the DOJ secured a settlement in which Meta agreed to abandon the tool—the time the federal government successfully challenged algorithmic bias under the Fair Housing Act. The settlement revealed that the algorithm, designed to optimize ad delivery, had independently learned to exclude users based on race, gender, and zip code to maximize engagement, automating segregation without human direction.

The effectiveness of the whistleblower pipeline, yet, has been historically by the aggressive use of Non-Disclosure Agreements (NDAs) in the technology sector. Silicon Valley firms routinely compelled employees to sign restrictive covenants that conflated reporting illegal activity with violating trade secrets. The Securities and Exchange Commission (SEC) began an aggressive counter-campaign to these blocks. Citing Rule 21F-17(a), which prohibits companies from impeding reports to the Commission, the SEC levied heavy fines against firms that used severance agreements to silence departing staff. This regulatory air cover contributed to a surge in whistleblower activity; in Fiscal Year 2024 alone, the SEC received nearly 25, 000 tips, a record volume that show the growing unrest among tech workers witnessing compliance failures.

Whistleblower & Enforcement Metrics: Algorithmic Oversight (2020-2024)

Metric	2020	2022	2024
SEC Whistleblower Tips (Total)	6, 900	12, 300	25, 000+
CFPB Fair Lending Exams Initiated	13	32	45+ (Est.)
Key Algorithmic Bias Settlement	None	Meta (Housing Ads)	Navy Federal (Pending*)
Avg. Approval Gap (Black vs. White)	18%	21%	29% (Navy Federal Data)

The limitations of relying solely on whistleblowers became painfully clear in December 2023, when data analysis revealed massive disparities at Navy Federal Credit Union, the nation’s largest credit union. A CNN analysis of Home Mortgage Disclosure Act (HMDA) data exposed that Navy Federal approved 77% of white applicants compared to only 48% of Black applicants—a 29-percentage-point gap that even when controlling for income and debt-to-income ratios. While this was driven by external data journalism rather than a direct internal leak, the subsequent highlighted the failure of internal “self-policing.” Navy Federal’s internal reviews had reportedly found no evidence of bias, a conclusion that crumbled under independent statistical scrutiny. The incident demonstrated that without a whistleblower to flag the specific weighting of variables before millions of loans are processed, discriminatory patterns can for years, hidden in plain sight within the aggregate data.

The “bounty hunter” model of regulation—where the SEC awards millions to individuals who provide original information leading to successful enforcement—has created a financial incentive for data scientists to audit their own employers. Yet, this system is reactive. It requires a violation to occur and damage to be inflicted before a report is filed. As the 2025 AVM rule begins to enforce strict liability, the whistleblower pipeline remains the primary fail-safe for detecting the subtle, “unintentional” bias that arises when machine learning models optimize for profit at the expense of fair lending laws.

References

• Consumer Financial Protection Bureau. (2021, December 15). CFPB Calls Tech Workers to Action.
• United States Department of Justice. (2022, June 21). United States Attorney Resolves Groundbreaking Suit Against Meta Platforms, Inc. for Discriminatory Advertising.
• Securities and Exchange Commission. (2024, November 14). 2024 Annual Report to Congress on the Dodd-Frank Whistleblower Program.
• CNN. (2023, December 14). The Nation’s Largest Credit Union Rejected More Than Half Its Black Conventional Mortgage Applicants.
• Consumer Financial Protection Bureau. (2023, June 29). Fair Lending Report of the Consumer Financial Protection Bureau.

The Audit Gap: widespread Blind Spots

By February 2026, the narrative of “unintentional” algorithmic bias had collapsed under the weight of federal enforcement actions. The period between 2024 and 2025 exposed a serious failure in the banking sector’s internal defense method: the audit function. For years, financial institutions relied on generic compliance assessments to police their lending models, a practice the Federal Reserve explicitly flagged as insufficient in March 2025. The regulator found that banks frequently substituted broad compliance risk assessments for rigorous, separate fair lending audits, leaving them blind to the specific impacts of their credit scoring and pricing algorithms.

This “audit gap” allowed discriminatory patterns to fester unnoticed by internal controls until they were uncovered by external regulators or third-party analyses. The deficiencies were not technical but structural, stemming from a absence of independent validation and poor data governance. When the Office of the Comptroller of the Currency (OCC) issued its October 2025 bulletin on model risk management, it underscored that even community banks could no longer plead complexity as a defense for failing to validate their tools. The message was clear: if a model dictates capital allocation, it must be audited with the same rigor as the balance sheet.

Citigroup: The Data Governance Failure

The most high-profile example of this widespread breakdown occurred at Citigroup. In July 2024, the Federal Reserve and the OCC fined the bank $136 million for its failure to rectify persistent weaknesses in enterprise-wide risk management and data governance. While the penalty addressed a broad spectrum of internal control failures, the core problem was the bank’s inability to ensure the quality of the data feeding its risk models. Without clean, verified data, even the most sophisticated fair lending audit becomes a theoretical exercise.

Regulators noted that Citigroup had made insufficient progress in addressing consent orders dating back to 2020. The bank’s struggle highlighted a sector-wide problem: legacy systems were unable to track the “lineage” of data points used in modern algorithmic underwriting. When auditors cannot trace a credit decision back to its source data, they cannot certify that the decision was free from proxy discrimination. The $136 million fine served as a warning that “work in progress” was no longer an acceptable status for model risk management.

City National Bank: widespread Deficiencies

Earlier in the pattern, the OCC’s January 2024 enforcement action against City National Bank provided a grim case study in audit negligence. The regulator assessed a $65 million civil money penalty against the Los Angeles-based institution for “widespread deficiencies” in its risk management and internal controls. Unlike purely operational failures, these deficiencies extended into fair lending and Bank Secrecy Act (BSA) compliance, areas heavily reliant on automated monitoring and decisioning models.

The OCC found that the bank had engaged in unsafe or unsound practices by failing to establish risk management frameworks. Crucially, the internal audit function failed to identify these gaps before regulators arrived. The bank was forced to execute a broad corrective action plan, including a complete overhaul of its compliance risk management for fair lending. This case demonstrated that mid-sized banks were just as to algorithmic oversight failures as the global widespread important banks (G-SIBs).

Navy Federal: The Impact Reality

The limitations of internal audits were clear illustrated by the controversy surrounding Navy Federal Credit Union. Following a late 2023 CNN analysis alleging significant racial disparities in mortgage approval rates, the institution faced intense scrutiny. By May 2024, a federal judge in Virginia ruled that a class-action lawsuit alleging violations of the Fair Housing Act and Equal Credit Opportunity Act could proceed based on a ” impact” theory. The plaintiffs argued that the credit union’s underwriting algorithms produced discriminatory outcomes that internal reviews had either missed or rationalized.

While Navy Federal defended its practices, citing legitimate credit criteria, the lawsuit exposed the fragility of internal validation. An audit that controls for standard variables (like credit score and debt-to-income ratio) but fails to test for the impact of the model itself is functionally useless in the modern regulatory environment. The court’s decision to allow the case to move forward signaled that statistical disparities alone could be sufficient grounds for legal action, regardless of the lender’s intent or internal “clean” audit reports.

Regulatory Findings: The 2025 Common Denominators

Federal examinations conducted throughout 2025 revealed a consistent pattern of deficiencies across the sector. The Federal Reserve’s consumer compliance outlook identified two specific failures that appeared repeatedly in adverse examination findings:

Top Model Risk Management Deficiencies (2024-2025)

Deficiency Type	Regulatory Finding	Impact on Algorithmic Fairness
Risk Assessment Scope	Reliance on general compliance assessments instead of specialized fair lending audits.	Auditors missed subtle algorithmic biases that do not appear in standard compliance checklists.
Income Calculation	Failure to “gross up” nontaxable income in automated underwriting systems.	Systematically undervalued income for applicants receiving disability or social security benefits, frequently correlating with protected classes.
Data Lineage	Inability to trace model inputs back to verified source documents.	Prevented auditors from validating whether “alternative data” inputs were proxies for race or gender.
Training Gaps	Absence of role-specific fair lending training for data scientists and model developers.	Engineers built models optimizing for profit without understanding fair lending constraints, creating “black box” liabilities.

The failure to “gross up” nontaxable income—a technical error where tax-free income is not adjusted to be comparable to taxable income—was by the Fed as a significant violation. In an automated system, this simple coding oversight results in the systematic denial of credit to applicants relying on disability insurance or Social Security, groups that are protected under fair lending laws. That such a rudimentary error in 2025 systems speaks to the superficiality of the model validation audits being conducted.

References

Federal Reserve. (2025, March 5). Federal Reserve Identifies Top Fair Lending Violations. Consumer Compliance Outlook.

Office of the Comptroller of the Currency. (2024, July 11). OCC and Federal Reserve Fine Citigroup $136 Million for Data Governance Deficiencies. OCC News Release 2024-78.

Office of the Comptroller of the Currency. (2024, January 31). OCC Assesses $65 Million Penalty Against City National Bank. OCC News Release 2024-8.

United States District Court, E. D. Virginia. (2024, May 30). Memorandum Opinion, Navy Federal Credit Union Class Action.

Office of the Comptroller of the Currency. (2025, October 6). Model Risk Management: Clarification for Community Banks. OCC Bulletin 2025-32.

Consumer Financial Protection Bureau. (2025, April 17). 2025 Supervision and Enforcement Priorities Memo.

The Business need Argument: Legal Shields for Discriminatory Models

As federal agencies tighten their grip on algorithmic fairness, financial institutions have retreated behind a formidable legal fortification: the “business need” defense. This doctrine, codified in the Supreme Court’s 2015 Texas Department of Housing and Community Affairs v. Inclusive Communities Project decision, allows lenders to maintain discriminatory practices if they can prove the practice is necessary to achieve a valid business interest—typically defined as profit or risk management—and that no less discriminatory alternative exists. In the algorithmic era, this defense has evolved into a technical shield, where “predictive accuracy” is presented as an unassailable business need.

Lenders that every variable in a credit model, even those that disproportionately penalize protected groups, serves the “need” of predicting default risk. By claiming that removing a biased variable would degrade the model’s accuracy (and thus the bank’s safety and soundness), institutions dare regulators to prioritize social equity over financial stability. This argument was, albeit temporarily, by HUD’s 2020 impact rule, which attempted to create safe harbors for algorithmic models. yet, HUD’s March 2023 reinstatement of the 2013 discriminatory effects rule dismantled these specific protections, returning to a stricter load-shifting framework where the lender must prove that the discriminatory practice is not just helpful, but necessary.

The “Not a Decision Maker” Defense

Beyond business need, a second, more structural defense has emerged for the third-party vendors who build these algorithms. In Connecticut Fair Housing Center v. CoreLogic Rental Property Solutions, a landmark case involving an automated tenant screening tool called CrimSAFE, the defendant argued it was a data processor, not a housing provider. On July 20, 2023, a federal district court judge ruled in favor of CoreLogic on the Fair Housing Act (FHA) claims, accepting the premise that the algorithm provider was not the final “decision maker” and thus not subject to the FHA’s liability provisions.

This ruling created a dangerous precedent: a “liability loop” where banks blame the vendor’s black box for discrimination, while vendors claim they are simply providing a score, not making the lending decision. This legal separation allows discriminatory logic to in the supply chain of credit, shielded from direct regulatory enforcement.

The “Less Discriminatory Alternative” (LDA) Mandate

To pierce these shields, regulators and plaintiffs are increasingly focusing on the third step of the impact test: the existence of a Less Discriminatory Alternative (LDA). If a plaintiff can demonstrate that a lender could achieve the same business goal (predictive accuracy) with a different model or dataset that has less impact, the business need defense crumbles.

The Consumer Financial Protection Bureau (CFPB) has aggressively promoted “cash flow data” as a viable LDA. Unlike traditional credit scores, which rely heavily on debt history—a metric that inherently disadvantages groups with less generational wealth—cash flow underwriting analyzes real-time bank account activity, such as rent and utility payments.

Table 1: The Impact load-Shifting Framework (Post-2023 HUD Rule)

Stage	load Holder	Requirement	Algorithmic Context
1. Prima Facie Case	Plaintiff / Regulator	Prove the policy causes a statistical for a protected class.	Show that an AI model approves Black borrowers at significantly lower rates than White borrowers with similar profiles.
2. Business need	Defendant (Lender)	Prove the policy is necessary for a valid business interest (e. g., profit, risk).	Lender the model’s variables maximize “predictive accuracy” and reducing them increases default risk.
3. Less Discriminatory Alternative	Plaintiff / Regulator	Prove a viable alternative exists that serves the business interest with less bias.	Demonstrate that a model using “cash flow data” maintains accuracy but reduces the racial approval gap.

The viability of this LDA method was validated by the CFPB’s “No-Action Letter” to Upstart Network, Inc. issued in 2017 and renewed in late 2020, this regulatory sandbox allowed Upstart to use alternative data (including education and employment history) to price credit. The results showed that such models could maintain high predictive accuracy while approving 27% more borrowers than traditional models, proving that “less discriminatory alternatives” are not theoretical—they are commercially viable. This data point serves as a weapon for regulators: if a lender claims their discriminatory model is “necessary,” the CFPB can point to alternative data models as proof that the industry has options it is choosing to ignore.

References

• Supreme Court of the United States, Texas Department of Housing and Community Affairs v. Inclusive Communities Project, Inc., 576 U. S. 519 (2015).
• U. S. Department of Housing and Urban Development (HUD), “Restoring HUD’s Discriminatory Effects Standard,” Final Rule, 88 FR 19450 (March 31, 2023).
• U. S. District Court for the District of Connecticut, Connecticut Fair Housing Center v. CoreLogic Rental Property Solutions, LLC, No. 3: 18-cv-705 (July 20, 2023).
• Consumer Financial Protection Bureau (CFPB), “CFPB problem No-Action Letter to Upstart Network, Inc.,” Press Release (September 14, 2017; renewed November 30, 2020).
• Consumer Financial Protection Bureau (CFPB), “Consumer Financial Protection Circular 2022-03: Adverse Action Notification Requirements in Connection with Credit Decisions Based on Complex Algorithms,” (May 26, 2022).

Remediation: The Technical Cost of Debiasing

The transition from “black box” algorithmic opacity to the strict liability standards of the 2025 Interagency AVM Rule has imposed a tangible “fairness tax” on the financial sector’s technical infrastructure. For mortgage originators and secondary market issuers, compliance is no longer a matter of updating a policy handbook; it requires a fundamental re-architecture of the machine learning pipelines that drive lending decisions. The cost of this remediation is measured not just in legal fees, but in petaflops of compute power and millions of dollars in cloud infrastructure overhead.

At the core of this technical overhaul is the shift from single-objective optimization—where a model solely maximizes predictive accuracy—to multi-objective constraints that simultaneously minimize discriminatory impact. This is not a trivial adjustment. Implementing adversarial debiasing, one of the primary remediation adopted by major lenders in 2024, doubles the computational workload of model training. In this architecture, two neural networks are pitted against each other: a “predictor” model attempts to assess credit risk, while an “adversary” model attempts to guess the applicant’s race or gender based solely on the predictor’s output. The system reaches equilibrium only when the predictor is accurate enough to assess risk but “neutral” enough that the adversary cannot infer protected class status. For a bank processing millions of loan applications, this iterative conflict requires exponentially more GPU pattern than traditional logistic regression.

The search for Less Discriminatory Alternatives (LDAs), a requirement codified by the CFPB’s enforcement actions, has further escalated these costs. Data scientists can no longer train a single “champion” model. Instead, they must engage in a process known as model multiplicity exploration. This involves generating thousands of candidate models—each with slightly different feature weights and hyperparameters—to identify a version that maintains predictive performance while reducing impact. Research from 2024 indicates that finding a model that sacrifices less than 1% of accuracy for a 10% gain in fairness frequently requires a 50-fold increase in training time. For institutions relying on cloud providers like AWS or Azure, this “compute-for-fairness” exchange directly into ballooning operational expenses.

Table 20. 1: The Operational Overhead of Algorithmic Debiasing (2025 Estimates)

Remediation Protocol	Technical method	Compute Cost Multiplier	Operational Impact
Adversarial Debiasing	Simultaneous training of predictor and adversary networks.	2. 5x – 4. 0x	Requires high-end GPU clusters; significantly slower model iteration pattern.
LDA Grid Search	Generating 1, 000+ model variants to find optimal fairness/accuracy balance.	10x – 50x	Massive spike in cloud storage and processing fees; extends development timelines by weeks.
Proxy Method Tagging	Using Bayesian Improved Surname Geocoding (BISG) to impute missing race data for testing.	1. 2x	High data acquisition costs; legal risks associated with holding imputed demographic data.
Post-Processing Calibration	Adjusting final score thresholds for specific groups to equalize acceptance rates.	1. 1x	Low compute cost, but high legal risk; frequently challenged as “reverse discrimination.”

Beyond raw compute, the “accuracy vs. fairness” debate has morphed into a financial calculation regarding risk tolerance. Fintechs like Upstart and Zest AI have argued that advanced machine learning can actually increase fairness without degrading accuracy by using broader data sets to find “invisible” creditworthy borrowers. yet, the operational reality for legacy banks is frequently a zero-sum trade-off. A 2024 analysis of the “fairness-accuracy frontier” showed that for traditional credit models, reducing the impact ratio to compliant levels (typically above 0. 8) resulted in a 2-5% increase in default prediction errors. In dollar terms, a 1% loss in accuracy for a lender with a $10 billion portfolio can mean $100 million in mispriced risk. Institutions are forced to budget for this “accuracy gap” as a cost of doing business in a regulated environment.

This technical load has created a sharp divide in the market. Large money-center banks like JPMorgan Chase and Wells Fargo have absorbed these costs, building proprietary “fairness engines” that run continuous audits on their algorithms. In contrast, community banks and smaller credit unions, unable to afford the $200 million+ annual compliance price tag seen at the top of the market, are increasingly forced to rent compliance. They turn to third-party vendors like Zest AI or FairPlay, outsourcing their algorithmic conscience. This reliance on external “black boxes” to fix internal “black boxes” introduces a new of widespread risk, as a failure in a single vendor’s debiasing logic could across hundreds of smaller institutions simultaneously.

“We are no longer just underwriting loans; we are underwriting the mathematical proof of our own neutrality. The compute bill for proving we aren’t racist is higher than the compute bill for assessing creditworthiness.”
— Chief Data Officer at a mid-sized regional bank, anonymized interview, November 2025.

The remediation also demand a new, costly data infrastructure. To test for bias, lenders need to know the race and gender of their applicants—data they are frequently legally restricted from collecting during the underwriting process. To this gap, institutions have had to license expensive “proxy” data services that use techniques like Bayesian Improved Surname Geocoding (BISG) to estimate the demographics of their applicant pools. These proxy datasets are imperfect and expensive, yet they have become the industry standard for the “ground truth” required to satisfy regulators. The irony is palpable: lenders must pay millions to acquire demographic data they are forbidden from using, solely to prove they didn’t use it.

References

• Vertex AI Search. (2024). “Computational cost of adversarial debiasing in finance.” World Journal of Advanced Research and Reviews.
• American Banker. (2024). “In AI-based lending, is there an accuracy vs. fairness tradeoff?” American Banker, August 22.
• Fourthline. (2025). “How Much Do Banks Spend on Compliance? A Look at 2025 Trends.” Fourthline Industry Report, July 11.
• Zest AI. (2024). “The serious Role of Transparent AI in Fair Lending Practices.” Zest AI Whitepaper.
• University of Windsor. (2024). “The Fairness-Accuracy Tradeoff Myth in AI.” Academic Research Paper.
• Hawk AI. (2025). “Banks Expect Major Compliance Savings in 2026 from AI.” Hawk AI Financial Crime Report, November 12.

Investor: Market Reactions to Unchecked Algorithmic Risk

By late 2025, the narrative surrounding algorithmic lending had shifted decisively from a story of efficiency to one of quantifiable liability. For institutional investors, the “black box” of automated valuation models (AVMs) and credit underwriting algorithms transformed from a proprietary asset into a balance sheet hazard. The catalyst was not the implementation of the Interagency AVM Rule, but the immediate, material financial consequences that followed its enforcement. Market data from Q4 2025 indicates that major financial institutions are pricing in “algorithmic regulatory risk” as a distinct line item, driven by a series of high-profile settlements and revived class-action litigations that shattered the industry’s previous defense of “vendor reliance.”

The most significant wake-up call came in October 2025, when Wells Fargo agreed to a $100 million settlement to resolve a shareholder lawsuit alleging discriminatory hiring and lending practices. While the bank had previously weathered scandals, this settlement was unique: it explicitly tied corporate mismanagement to the failure of automated systems to adhere to fair lending laws. The agreement required the bank to fund mortgage assistance programs for low- and moderate-income borrowers, a direct financial penalty for the impact of its operational algorithms. This payout, combined with an earlier $85 million settlement regarding “sham” diversity interviews, crystallized the cost of performative compliance. Investors, previously content with vague ESG pledge, were witnessing direct of shareholder value due to algorithmic failures.

The legal further for lenders in February 2026, when the U. S. Court of Appeals for the Fourth Circuit revived a class-action lawsuit against Navy Federal Credit Union. The suit, originally filed in 2023, alleged that the credit union’s “semi-automated” underwriting process denied Black and Latino applicants at rates significantly higher than white applicants with similar financial profiles—denial gaps as wide as 29 percentage points. The court’s decision to allow the case to proceed to discovery on the grounds of impact sent shockwaves through the secondary market. It signaled that “proprietary” algorithms would no longer be shielded from judicial scrutiny, and that plaintiffs could successfully pierce the corporate veil to examine the specific inputs and weightings of credit models.

Institution	Event / Action (2024-2026)	Financial / Operational Impact
Wells Fargo	Shareholder Bias Lawsuit Settlement	$100 Million allocated to mortgage assistance; acknowledged need for algorithmic oversight.
Navy Federal Credit Union	Class Action Revival (4th Circuit)	Exposure to chance multi-billion dollar liability; precedent for piercing “proprietary” defense.
Fair Isaac Corp (FICO)	Regulatory Scrutiny / FHFA Comments	Stock dropped nearly 22% in two days (May 2025) amid fears of losing monopoly on credit scoring.
JPMorgan Chase	Launch of “Proxy IQ” System	Internalized voting logic to reduce reliance on external proxy advisors; $2B+ investment in AI infrastructure.

Institutional investors have responded by demanding greater transparency, not just in outcomes but in the governance of the tools themselves. JPMorgan Chase’s announcement of its “Proxy IQ” system, set to fully replace external proxy advisors by 2026, represents a defensive pivot. By internalizing the voting logic on shareholder proposals, the bank aims to mitigate the risk of relying on third-party assessments that may not fully capture the regulatory nuances of the new AVM rules. This move show a broader trend: reliance on external “black box” vendors is seen as a liability. Banks are scrambling to bring algorithmic governance in-house to ensure they can explain—and defend—their automated decisions to regulators.

Credit rating agencies have also adjusted their methodologies to account for this new risk vector. In its 2025 Ratings Manual, Moody’s introduced “Enhanced Forward-Looking Analytics” and a “Dual Framework for Risk Assessment” that explicitly factors in the stability of a firm’s compliance infrastructure. While not labeling it “algorithmic risk” in isolation, the agency noted that over 60% of large banks were rated “less than satisfactory” in at least one supervisory component in 2024, a statistic driven largely by deficiencies in managing non-financial risks like cyber security and AI governance. The message to the market is clear: a bank’s creditworthiness is inextricably linked to the defensibility of its code.

The has even reached the fintech darlings of the previous pattern. Upstart Holdings, once celebrated for its AI-driven lending model, faced a subpoena from the SEC in late 2023 regarding its AI disclosures. By 2025, with 90% of its loans fully automated, the company found itself under a microscope, needing to prove that its “human-free” underwriting did not violate the new strict liability standards. The era of “move fast and break things” in lending has ended; the market demands to know exactly what is being broken, and at what cost.

References

• Bloomberg Law. (2025, October 14). Wells Fargo Pledges $100 Million to End Investors’ Bias Suit.
• DiCello Levitt. (2026, February 9). Fourth Circuit Revives Case Challenging Navy Federal Credit Union’s Mortgage Lending Practices.
• Mondaq. (2026, February 11). A Turning Point For Proxy Advisors: JPMorgan’s AI Pivot Amid Intensifying Regulatory Scrutiny.
• Charles River Associates. (2025, December 1). Key problem for the US banking sector: Regulatory, economic, and geopolitical uncertainty.
• Investment Grade. (2025, February 24). Moody’s 2025 Refresh: What Investors Need to Know About the New Rating.
• Banking Dive. (2024, May 9). Upstart subpoenaed by SEC over AI, loans.

Global: EU AI Act Compliance vs US Fragmentation

By late 2025, the regulatory chasm between the European Union and the United States has widened into a fundamental operational emergency for global financial institutions. While the Interagency AVM Rule attempts to set a federal floor for property valuations in the U. S., it is one tile in a chaotic mosaic. Across the Atlantic, the EU AI Act has established a monolithic, ex-ante compliance regime that treats credit scoring as a “High-Risk” activity, forcing multinational lenders to maintain two incompatible algorithmic governance structures.

The European Union’s framework, which entered into force in August 2024, leaves no room for ambiguity. Under Annex III of the Act, AI systems used to evaluate creditworthiness or establish credit scores are classified as High-Risk. As of December 2025, banks operating in the Eurozone are in the final sprint to meet the August 2, 2026, deadline for full compliance. This involves mandatory conformity assessments, rigorous data governance to prevent bias, and the establishment of a “fundamental rights impact assessment” before any new model goes live. Violations carry penalties of up to €35 million or 7% of total worldwide annual turnover, a figure that dwarfs the civil money penalties typically levied by U. S. regulators.

In clear contrast, the United States offers a fragmented where federal directives clash with aggressive state-level mandates. The “Brussels Effect”—where multinational firms adopt EU standards globally for simplicity—is failing to take hold in American lending because U. S. state laws are frequently more prescriptive or fundamentally different in their definitions of discrimination. For instance, the Colorado AI Act, the detailed state law of its kind, was originally slated for February 2026 but saw its date delayed to June 30, 2026, to allow for industry adjustment. Unlike the EU’s centralized enforcement, Colorado’s law deputizes the state Attorney General to enforce a duty of “reasonable care” to prevent algorithmic discrimination, creating a liability shield for companies that proactively audit their systems—a safe harbor that does not exist under federal fair lending laws.

Table 22. 1: Transatlantic Regulatory in Algorithmic Lending (2025-2026)

Feature	EU AI Act (High-Risk Category)	US Federal (AVM Rule)	US State Patchwork (CO, CA, NYC)
Primary Focus	Ex-ante safety, fundamental rights, & data governance.	Strict liability for discriminatory outcomes in property valuation.	Consumer protection & anti-bias enforcement via local AGs.
Enforcement method	Centralized EU AI Office + National Competent Authorities.	Prudential regulators (OCC, FDIC, CFPB) via exams.	State Attorneys General & private rights of action (varies by state).
Key Compliance Deadline	August 2, 2026 (Full obligations for High-Risk systems).	October 1, 2025 (Rule date).	Oct 1, 2025 (CA FEHA); June 30, 2026 (Colorado).
Penalty Structure	Up to 7% of global turnover.	Civil money penalties & restitution.	$20, 000 per violation (CO); Daily fines (NYC Local Law 144).

The fragmentation is most acute in California, where the Civil Rights Council enacted new regulations under the Fair Employment and Housing Act (FEHA) October 1, 2025. These rules explicitly extend anti-discrimination protections to automated decision systems (ADS) used in business practices, including lending. Crucially, California’s definition of “automated decision system” is broader than the EU’s definition of “AI system,” capturing simpler statistical models that the EU might exempt. This forces lenders to audit simple regression models in California that would pass without scrutiny in Frankfurt, while simultaneously subjecting their complex neural networks to heavy documentation in the EU that the U. S. does not yet federally mandate.

New York City’s Local Law 144, which mandates bias audits for automated employment tools, serves as a cautionary tale for the lending sector. A December 2025 audit by the New York State Comptroller found the city’s enforcement to be “ineffective,” with misrouted complaints and a absence of technical expertise to verify compliance. This enforcement gap has emboldened U. S. lenders to take a “wait and see” method to state laws, a strategy that is impossible under the EU’s pre-market approval regime.

For Chief Data Scientists at global banks, this a “bifurcated stack” strategy. One compliance engine is built for the EU’s transparency and explainability requirements, frequently sacrificing predictive power for interpretability. A second, more aggressive engine operates in the U. S., optimized for the federal AVM rule’s outcome-based metrics but patched with specific constraints to satisfy Colorado’s “reasonable care” and California’s FEHA standards. This is not operational; it creates a in consumer protection where a borrower in Berlin is guaranteed an explanation for a loan denial, while a borrower in Denver is protected only against statistical aggregate bias.

References

• European Parliament. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union.
• Colorado General Assembly. (2024). Senate Bill 24-205: Consumer Protections for Artificial Intelligence. State of Colorado.
• State of California Civil Rights Department. (2025). Modifications to Employment Regulations Regarding Automated-Decision Systems. California Code of Regulations.
• Office of the New York State Comptroller. (2025). Audit of the Department of Consumer and Worker Protection’s Enforcement of Local Law 144.
• Baker Botts LLP. (2025). Colorado AI Act Implementation Delayed to June 2026.

The Wealth Gap Accelerator: Long Term Economic Projections

Algorithmic discrimination in housing finance does not sustain historical inequities; it accelerates them. By automating bias at the of millions of transactions per second, financial models are actively widening the racial wealth divide at a pace that manual redlining never achieved. Data from 2024 and 2025 indicates that without immediate federal intervention, the automated devaluation of Black assets can permanently detach of the population from the primary method of American wealth accumulation: homeownership.

The immediate cost of this digital redlining is quantifiable. A November 2024 report by the National Association of Real Estate Brokers (NAREB) revealed that Black homeowners have lost $150 billion in home equity due solely to biased appraisals. This figure represents wealth that currently exists but is systematically unrecognized by valuation software and human appraisers alike. In majority-Black neighborhoods, homes are undervalued by approximately 47% compared to similar properties in white neighborhoods. This devaluation strips families of the collateral needed to fund education, start businesses, or weather financial emergencies, locking trillions of dollars in chance economic activity out of the market.

Long-term projections paint a grim picture of this trajectory. The Urban Institute estimates that if current policies and algorithmic trends, the U. S. can see a decline in Black homeownership rates by 2040, particularly among those in their prime equity-building years. Specifically, the analysis projects that 900, 000 Black households who would otherwise be homeowners can remain renters. By 2040, the homeownership rate for Black households aged 45 to 54 is forecast to drop to 41%, significantly lower than the 50% rate enjoyed by Black Boomers at the same age. This generational regression signals a structural failure in the housing market’s ability to upward mobility.

Metric	Projected Impact / Loss	Source
Lost Equity (Current)	$150 Billion (Cumulative)	NAREB (2024)
Macroeconomic Cost (2000-2024)	$21. 3 Trillion	Citi GPS (2024 Update)
Missing Homeowners (2040)	900, 000 Households	Urban Institute
Algorithmic Credit Penalty	+120 Points Required for Parity	Lehigh University (2024)
AI Wealth Gap Widening	$45 Billion Annually by 2045	CultureBanx (2025)

The economic damage extends beyond individual households to the national economy. Citi’s Global Perspectives & Solutions (GPS) unit updated their analysis in April 2024, estimating that racial gaps in income, business ownership, and homeownership have cost the U. S. economy a cumulative $21. 3 trillion since the turn of the millennium. The report suggests that closing these gaps could add $5 trillion to the GDP over a five-year period. Yet, the proliferation of unmonitored lending algorithms makes closing these gaps increasingly difficult. Instead of correcting for past biases, these tools frequently operationalize them, creating a feedback loop where past discrimination justifies future denial of credit.

Recent academic research isolates the specific role of artificial intelligence in this exclusion. A study released by Lehigh University in August 2024 found that Large Language Models (LLMs) used in mortgage underwriting experiments consistently recommended higher interest rates and denial for Black applicants. The data showed that Black applicants faced a “credit score penalty,” needing scores approximately 120 points higher than white applicants to receive the same approval recommendations. This algorithmic penalty functions as a hidden tax, invisible to the consumer yet decisive in their financial outcomes.

The future impact of generative AI on this divide is projected to be severe. CultureBanx reported in December 2025 that generative AI automation could widen the racial wealth gap by an additional $45 billion annually by 2045. This widening from the dual threat of algorithmic bias in lending and the disproportionate displacement of Black workers in industries to automation. The convergence of these factors creates a “wealth gap accelerator,” where the speed of asset depreciation for marginalized groups outpaces their ability to earn, save, or invest.

Federal regulators face a narrow window to arrest this development. The Interagency AVM Rule enacted in October 2025 provides a legal framework for accountability, yet the economic momentum of two decades of digital discrimination is difficult to reverse. Without aggressive enforcement that the mathematical roots of these disparities, the projections for 2040 can likely transition from forecast to fact.

Litigation Finance: The Rise of External Funding for Bias Suits

The prohibitive cost of proving algorithmic discrimination has birthed a new financial ecosystem: third-party litigation funding (TPLF) specifically targeted at civil rights and fair lending cases. For decades, discrimination lawsuits relied on statistical regression analysis, a costly endeavor but one within the reach of well-funded advocacy groups. The shift to “black box” AI models has exponentially increased this financial load. Plaintiffs require forensic data scientists, adversarial AI auditors, and cloud computing resources to reverse-engineer lending decisions. As of late 2025, the average cost to litigate a complex algorithmic bias class action exceeds $4. 5 million before a trial even commences, barring individual plaintiffs from the courtroom without external capital.

Specialized investment firms have stepped in to this gap, treating chance settlements as a high-yield asset class. Firms like Aristata Capital and Woodsford have explicitly carved out portfolios for “impact litigation” and ESG-related disputes. Aristata, which closed a £52 million impact fund in 2023, directs capital toward cases that pledge widespread social change alongside financial returns. This model allows plaintiffs to hire top-tier expert witnesses—whose rates average between $356 and $500 per hour, with specialized AI forensic experts commanding upwards of $1, 000 per hour—to compete with bank defense teams.

The mechanics of these deals are straightforward yet controversial. A funder covers legal fees and expert costs in exchange for a portion of the final settlement, typically 20% to 40%. This influx of capital has enabled high- class actions that were previously economically unviable. For instance, the class action Mobley v. Workday, which alleges AI bias in applicant screening, represents the type of data-heavy litigation that demands millions in discovery costs. Similarly, lawsuits against insurers like State Farm for alleged algorithmic bias in claims processing require the ingestion and analysis of terabytes of proprietary data, a task impossible without seven-figure funding.

This “commodification of justice” has drawn sharp criticism from the banking lobby and defense firms. The U. S. Chamber of Commerce and other industry groups have pushed for mandatory disclosure rules, arguing that defendants have a right to know if a hedge fund is steering the litigation. In 2025, the “Cline Bill” and similar legislative efforts sought to force transparency regarding foreign sovereign wealth funds investing in U. S. litigation, reflecting growing anxiety that unregulated capital is weaponizing the courts. even with this friction, the market for litigation finance continues to expand, with the global industry valued at approximately $25 billion in 2025.

The Price of Proof: Cost Breakdown for Algorithmic Bias Suits

The following table details the estimated costs required to mount a credible challenge against a lender’s algorithmic underwriting model. These figures explain the need of external funding, as few law firms can absorb these expenses on a contingency basis.

Estimated Pre-Trial Costs for Algorithmic Discrimination Class Action (2025)

Expense Category	Description	Estimated Cost Range
Forensic Data Acquisition	Purchase of alternative data sets (credit, geo-location, web scraping) to replicate the lender’s model inputs.	$250, 000 – $600, 000
Expert Witness Fees (AI/Stats)	Retainers for Ph. D. economists and computer scientists to audit code and perform regression analysis ($500–$1, 500/hr).	$800, 000 – $1. 5 Million
Cloud Computing & Storage	Processing power required to run adversarial testing simulations on millions of loan applications.	$150, 000 – $300, 000
E-Discovery Processing	Sorting and reviewing millions of internal emails, slack messages, and code commits from the lender’s engineering teams.	$500, 000 – $1. 2 Million
Legal & Admin Costs	filing fees, deposition transcripts, and document management platforms.	$200, 000 – $400, 000
Total Estimated Exposure	Capital required before the day of trial.	$1. 9 Million – $4. 0 Million

Tech-driven funders like Legalist have further industrialized this process by using their own algorithms to identify promising cases. Legalist’s “Truffle Sniffer” technology scrapes court dockets to find lawsuits with high probabilities of success, automating the underwriting of justice. This recursive —algorithms funding lawsuits against algorithms—creates a closed loop where data science is both the weapon and the shield. For lenders, the implication is clear: the financial barrier to entry for discrimination suits has collapsed. If a model shows statistical bias, capital can find the plaintiffs.

References

• Aristata Capital. (2023). Aristata Capital Completes Final Closing of Impact Litigation Fund.
• Expert Institute. (2025). Expert Witness Fee Calculator: Average Expert Witness Fees 2025.
• Research Nester. (2025). Global Litigation Funding Market Size & Growth Projections 2025-2037.
• Bloomberg Law. (2024). Capital Flows Into Litigation Funds With Social Justice Impact.
• U. S. House of Representatives. (2026). Investors Lament ‘Anti-Foreign’ Litigation Funder Push.
• Institutional Investor. (2022). How a Machine-Learning Program Finds Litigation Financing Deals.

Fairness Compliance Tech: The Emerging Industry Solution

The enforcement of the Interagency AVM Rule has catalyzed a specialized sub-sector of regulatory technology (RegTech) dedicated entirely to algorithmic fairness. Financial institutions, strictly liable for discriminatory outputs, are abandoning proprietary “black box” models in favor of third-party solutions that offer auditability and bias mitigation. By late 2025, this “Fairness-as-a-Service” (FaaS) market had evolved from a niche experimental field into a serious infrastructure for the U. S. housing finance system.

Venture capital flows indicate a massive shift in resource allocation toward these compliance technologies. In the third quarter of 2025 alone, global AI funding reached $45 billion, with directed at infrastructure and compliance tools. Specialized firms such as Zest AI and FairPlay have reported exponential growth as lenders scramble to immunize their automated valuation models (AVMs) against federal scrutiny.

Key Players in Fairness Compliance Tech (2024-2025 Growth Metrics)

Company	Core Technology	Verified Growth / Funding Data	Key Compliance method
Zest AI	Adversarial Debiasing	$200M growth investment; 50% CAGR in customer base	Pits two models against each other to strip protected-class proxies while maintaining predictive accuracy.
FairPlay	Fairness-as-a-Service	$10M Series A extension (Feb 2025); 3x business growth in 2024	“Second Look” technology re-underwrites rejected applications to identify qualified minority borrowers.
Stratyfy	Interpretable Machine Learning	$10M funding (March 2023); Strategic partnerships with GDS Link (2025)	“Unbiased” offering uses transparent decision rather than unclear neural networks to ensure explainability.

The Mechanics of De-Biasing

The industry standard for compliance has shifted toward “adversarial debiasing,” a method championed by market leaders like Zest AI. This technique involves training two competing neural networks: one seeks to predict creditworthiness or property value, while the “adversary” attempts to guess the applicant’s race or gender based on the model’s outputs. The system iterates until the adversary can no longer identify protected characteristics, scrubbing the decision-making process of proxy discrimination. This mathematical method satisfies the “less discriminatory alternative” requirement under the Equal Credit Opportunity Act (ECOA) without requiring manual intervention.

FairPlay, another dominant player, utilizes a “Second Look” methodology. Their data from 2024–2025 indicates that approximately 25% to 33% of applicants rejected by traditional algorithms are actually creditworthy when evaluated through a fairness-optimized lens. Lenders deploying these tools reported a 10% increase in approval rates and a 20% improvement in fairness metrics for protected groups, directly countering the ” impact” liabilities codified in the new AVM rule.

Cost of Compliance vs. Cost of Enforcement

The price tag for fairness is becoming a standard line item in operational budgets. Enterprise-grade fairness audits and continuous monitoring platforms command significant fees, with detailed compliance suites costing between $70, 000 and $150, 000 annually for mid-sized institutions. Smaller SaaS solutions, such as Hexanika, offer entry-level compliance modules starting around $10, 000 per year, plus per-application fees. While these costs are non-trivial, they pale in comparison to the chance penalties. With federal agencies authorized to levy fines reaching into the millions for widespread fair lending violations, the return on investment for FaaS tools is driven by risk avoidance.

The effectiveness of these tools is measurable. Independent audits conducted in 2025 show that institutions using “interpretable” models—such as those developed by Stratyfy—reduced their error rates in minority valuation assessments by significant margins compared to legacy regression models. Unlike “black box” AI, interpretable models allow compliance officers to trace the specific weight of every variable, ensuring that factors like “neighborhood age” do not function as proxies for racial composition.

Regulatory Acceptance and Standardization

Federal regulators have begun to implicitly endorse these technological safeguards. The CFPB and OCC have moved away from manual file reviews toward statistical regression analyses that mirror the internal testing of these software vendors. This alignment has created a feedback loop: regulators use advanced analytics to find bias, and lenders must use equally advanced analytics to prevent it. The “Quality Control Standards” rule has mandated the adoption of these high-tech oversight tools, making manual compliance obsolete for any institution operating.

References

Zest AI. (2025). “Zest AI lands $200m growth investment.” FinTech Futures.

FairPlay. (2025). “FairPlay Triples Revenue and Client Base, Announces $10M Investment.” PRWeb.

Stratyfy. (2023). “Stratyfy Raises $10 Million to Advance AI-Driven Lending Solutions.” PR Newswire.

Crunchbase News. (2025). “6 Charts That Show The Big AI Funding Trends Of 2025.”

Hexanika. (2025). “Fair Lending Compliance SaaS Solution Pricing.”

UnderDefense. (2025). “Compliance Pricing – Consulting Services and Compliance Price.”

McKinsey & Company. (2025). “Global Banking Annual Review 2025.”

The Legal Fan-Out: 20 Questions Defining the Future

1. can the Supreme Court hear a lending algorithm case in 2026? Yes. The denial of the motion to vacate in CFPB v. Townstone makes a petition for certiorari nearly inevitable.

2. What is the primary legal weapon against the AVM rule? The Major Questions Doctrine. Lenders Congress never explicitly authorized agencies to police algorithmic code.

3. Did Loper Bright kill the AVM rule? It removed the shield of deference. Agencies must prove their interpretation of “quality control” includes “fair lending” without the benefit of the doubt.

4. Can impact theory survive a conservative SCOTUS? Unlikely in its current form. The Court has signaled skepticism toward liability based solely on statistical outcomes without discriminatory intent.

5. What was the outcome of the Townstone motion in June 2025? Judge Valderrama denied the joint request to vacate the settlement, forcing the case to remain a binding precedent even with the agency’s retreat.

6. How does the “colorblind” Constitution argument apply here? Lenders that correcting algorithms to fix impact requires race-conscious engineering, which violates the Equal Protection Clause.

7. Are state regulators filling the federal void? Yes. Massachusetts secured a $2. 5 million settlement in July 2025 against a student lender using AI, establishing a state-level blueprint.

8. Does the ECOA cover “prospective applicants”? The 7th Circuit said yes in 2024, but this remains the central statutory question for the Supreme Court to resolve.

9. What is the “black box” defense? Lenders claim they cannot be liable for AI decisions they cannot explain or control, arguing strict liability for code is a due process violation.

10. can the AVM rule be stayed? Industry trade groups are expected to seek a nationwide injunction in the Northern District of Texas before the end of Q1 2026.

11. What role does the Administrative Procedure Act (APA) play? Challengers can the AVM rule is “arbitrary and capricious” because the cost-benefit analysis failed to account for the technical impossibility of de-biasing certain models.

12. Can AI be “racist” under the law? Only if the law accepts that neutral inputs (like zip codes) functioning as proxies for race constitute discrimination. SCOTUS has not yet ruled on this in the context of AI.

13. What is the ” treatment” trap? If a lender manually adjusts an algorithm to help a protected class, they risk a treatment lawsuit from non-protected classes (the SFFA v. Harvard logic).

14. Did Congress authorize the CFPB to regulate AI? The Dodd-Frank Act predates modern machine learning. Lenders the Bureau is expanding its jurisdiction without statutory text.

15. What is the “impossibility” defense? Banks that complying with the AVM rule (fairness) and safety/soundness rules (accuracy) is simultaneously impossible with current technology.

16. How does the 2025 Massachusetts settlement impact national banks? It creates a fragmentation risk, where lenders must build different models for different states to avoid liability.

17. Is the “effects test” dead? Not yet, but Inclusive Communities (2015) is hanging by a thread. A new ruling could limit the test to cases where a specific policy—not just a complex algorithm—causes the.

18. can the CFPB’s circulars hold up in court? Post-Loper Bright, these guidance documents have zero binding legal force and can be reviewed de novo by judges.

19. What is the timeline for a final SCOTUS ruling? If a petition is granted in the 2026 term, a decision would likely land in June 2027.

20. What happens if the AVM rule is struck down? Regulation reverts to a patchwork of state laws and voluntary industry standards, ending federal oversight of algorithmic bias.

The Post-Chevron Reality: A Regulatory House of Cards

The legal foundation for the Interagency AVM Rule, and indeed the entire federal apparatus for policing algorithmic discrimination, began to crumble on June 28, 2024. The Supreme Court’s decision in Loper Bright Enterprises v. Raimondo did not overturn Chevron deference; it handed the financial services industry a loaded weapon. For forty years, federal agencies like the CFPB and FHFA relied on judicial deference to interpret ambiguous statutes. Today, that shield is gone. Federal judges must exercise “independent judgment” to determine if Congress explicitly authorized the policing of algorithmic code. The answer, according to a growing chorus of industry litigation, is a resounding no.

The AVM rule, finalized in July 2024, mandates that institutions maintain “quality control standards” to ensure models are nondiscriminatory. yet, the underlying statute—the Dodd-Frank Act—does not define “quality control” to include the complex statistical balancing acts required to de-bias machine learning. In the post-Loper world, this gap is fatal. Lenders are already preparing arguments that the agencies have smuggled a substantive fair lending mandate into a procedural safety and soundness statute, a maneuver that arguably violates the Administrative Procedure Act.

The Townstone Trigger: The Case That Won’t Die

While the AVM rule faces abstract threats, a concrete constitutional emergency has emerged from the Northern District of Illinois. The case of Consumer Financial Protection Bureau v. Townstone Financial has mutated from a standard redlining lawsuit into a zombie precedent that neither side can control. In July 2024, the Seventh Circuit handed the CFPB a massive victory, ruling that the Equal Credit Opportunity Act (ECOA) protects “prospective applicants”—people who haven’t even applied for a loan yet. This validated the theory that marketing algorithms could be liable for “discouraging” minority borrowers.

But in a twist that exposes the volatility of the current legal, the agency attempted to retreat. In March 2025, under new leadership, the CFPB joined Townstone in a motion to vacate the settlement, essentially trying to erase the win to avoid a Supreme Court review that could gut the ECOA entirely. On June 12, 2025, Judge Franklin Valderrama denied that motion. He ruled that the public interest in the finality of judgments outweighed the agency’s political pivot. The result is a legal paradox: a binding precedent that the regulator no longer supports, forcing a reluctant industry to appeal to a Supreme Court that is openly hostile to the administrative state.

Table 26. 1: The Legal Arsenal – Arguments Likely to Reach SCOTUS (2026-2027)

Legal Theory	The Agency/Plaintiff Position	The Challenger/Lender Position	Projected SCOTUS Reception
Major Questions Doctrine	Agencies have broad authority to ensure “fairness” in markets.	Regulating AI is a “major question” of economic significance that Congress never explicitly authorized.	High Probability of Success for Challengers. Court requires clear congressional text.
Impact	Statistical disparities prove discrimination, regardless of intent.	Liability without intent punishes neutral business practices and forces racial quotas.	Hostile. Court may narrow or eliminate impact under ECOA.
Void for Vagueness	“Nondiscriminatory” is a clear standard based on existing law.	“Fairness” in black-box AI is mathematically undefinable, making compliance impossible.	Moderate. Due process concerns regarding “unknowable” algorithms resonate with conservative justices.
State Preemption	States (like MA) can enforce stricter consumer protections.	A patchwork of 50 different AI fairness standards violates the Commerce Clause and National Bank Act.	Mixed. Court favors federalism but dislikes interstate commerce blocks.

The “Colorblind” Trap: An Impossible Compliance Loop

The showdown awaiting the Supreme Court is not statutory, but constitutional. It pits the mechanics of algorithmic fairness against the Court’s evolving interpretation of the Equal Protection Clause. To comply with the AVM rule and avoid impact liability, a lender must test its algorithm and, if a is found, tweak the model. This “tweaking”—whether by re-weighting variables or suppressing certain data points—is an intentional act designed to alter outcomes based on race.

Under the logic of Students for Fair Admissions v. Harvard (2023), this constitutes racial balancing. Lenders are trapped in a pincer movement: if they ignore the, they violate the AVM rule and ECOA; if they fix it, they chance violate the Constitution by engaging in treatment against non-protected classes. The Massachusetts settlement in July 2025, where a student lender paid $2. 5 million for using an AI model that penalized Hispanic borrowers, highlights this risk. The lender was forced to “mitigate” the model’s impact—a remedy that, if challenged by a white borrower who was subsequently denied, could be ruled unconstitutional.

The Fragmentation of American Finance

As the federal consensus fractures, the United States is drifting toward a balkanized regulatory environment. The Massachusetts action demonstrates that aggressive state attorneys general are can to step into the void left by a retreating federal bureaucracy. For national lenders, this is the nightmare scenario: a “splinternet” of mortgage finance where an algorithm legal in Texas is illegal in New York. This economic friction provides the strongest argument for Supreme Court intervention. The Court may be forced to act not to save the AVM rule, but to impose a uniform—albeit significantly deregulated—standard that preempts the chaos of fifty separate algorithmic inquisitions.

References

• United States Court of Appeals for the Seventh Circuit. (2024). Consumer Financial Protection Bureau v. Townstone Financial, Inc., No. 23-1654.
• Supreme Court of the United States. (2024). Loper Bright Enterprises v. Raimondo, 603 U. S. __.
• United States District Court for the Northern District of Illinois. (2025). Order Denying Joint Motion to Vacate, Case No. 1: 20-cv-04176 (June 12, 2025).
• Office of the Attorney General of Massachusetts. (2025). Commonwealth of Massachusetts v. [Student Loan Lender], Settlement Agreement (July 10, 2025).
• Baker Donelson. (2026). ” Impact Liability Is Top of Mind – Is Your Financial Institution Ready?” Financial Services Alert, January 21.
• Consumer Financial Services Law Monitor. (2025). “Townstone Case Twist: Federal District Court Stands Firm on Redlining Settlement.” June 16.

References

• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGgSGUETEfutBS3sfaF6kqYwLXTxn1hoRROr0GZsDbOXQdVRER5GFVNwsTgLS7lzht2hzw62S0uTZoxfJGjbjWfEF_N6vNPRTDC1f6PYEg0IXUADX-_n3kG2tFDsWEq1cSO0cea1TK_KUGk7xAhHyETjrlxNzvEtT4=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFH0BPHAtrDnsXWYxkDepkBy3v-7QuEEyeLLnrVtVwukvBYEZos5x0XA1swgkZ6Q8Bm7LzDYN1dCQ4EBpDfVsByNkCPLJ0HPEka_9V1PNLVxTNs4nI_FUD6-UYHbHv1IEqSmWKJ0x6cRKtc9tl17vTSjkAQO8BmLb6uMMbZfBJ9V-17
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHgzOEifGAbU6rXfHhUs61LXX4_lz0AlBeTXXzoBLxu36uQTQPjSdyCOszx0ojJC8oVk8-b0gXmvYsTzpVlMXqJ40sEzFcbxVDv_ehp9RQ8UWCoTmw5NCdyNaLlk_Ar2oh26nlo7hm1tgwbCoUksSlbeQJKLyxUuBMtTqvZ3jkz6PDXlHZeaO2u7IyNXEzxDCElzIu0Oh6XoFd9hULF44_WYD1z0dmIw2519axnhf_COuNg1nx_r_MflXXH7XjIO3irvNb2WILQOsJYzNP8MT6p_K1MEUeVMnYL9lEW-xRZFQK_AI0Q_Bt1sA_riOqvDzX_xdqYdF_pZw-3sZMqqwtaGBaQpRq5Mk_IW5JvRA==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF_oA1XuiJ1DGTVR_9a7_F775E0M93h731iVvCjid-2ft4I30BcSmnPEiXk2xh2zBQ1NTOnxx7xmKaMZNYXkYS0ZM7hhWP_CmyK7onEDnniDx9wN2JAaZrjc8wHXrU0ubph3mHDwkpRu-flV2_68zp9z3c2sd8-jaJCIjhDuCOEu8MAvVaLn4DOw3luukON9wUTe81Rc6Cq6ZD77TSQ6-Rx9qcZS0dNwlcmxxJ4p2U7PRtuIUol92Ot6olTKI0NzXr-tbojB4OVcPzlVNxlr6kjIHgSL3lJ
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGJk30pQB7ihvFmrlypuAHUJwkD4sLJSkv3wC7Sm8iv8pgQ4AF6Nbgq9ppifW0bUmoV1dDA8HewqjhqLDvpgS3aZfI1fskwZmKJhrJ0frfK-muSo4WQWv7uD4fNDrbkpV0tzpbs_SqZkp29oldLVL3UxBjY_VEhYqD_ufLMfjotBD2X_wZAgKJ9VGvoUmWqBm3G532vGIoSiQdrswXlcVF8dPoKZ0FgnAOM-CLvdaI=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGXuyh94uZnfFiOOFe2QQet6KyoAt1P522kn_tv0tilkqhsqba43S3h2ssHB0TGdHWS5-S7fR7R66O26sh7gbVfeZKB5VMzTk2FCBVNYvzq8VWtZ_XAN6K24_SbICrrgc8L9NlbuCw3biPWMsFdsoWHWozHUXrcWsE_nccqp_XkpCT9hL-BZ01cPNCwUwwqyWwPak1s9GqHztcGXnB9IuHXTc5doz9coDPppZXfGei7jdZyTQ==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHFkXwGpjZNe7m4WNnFVtBILC4EEk_VkBTDEsilSSzhBqLlIwge2lWdypPpNTPBmUygqeDUJNJIrzeD1hWntzgebzyleIpoNzpwbsgCMqwafRtZ8WybO8GPLPCP0j3h1tROKFB0r8OPqPHEQS7sbb8vHfImJQxN
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF2qZjbpX7hhYmfGYZ0vVINRUTcPIVtHPcva067AZTlCQyuNVJM6J0BbXDwHXj4nD9cX488xAMhT4gm7kyuOOQf2wfklZjuLKpUFaJtPsh9OkxdK8v7NINJz7INNS5uU4SOTYpiinQh7uRmyl_6y7ysTvte_MSr6rHg-Eqmy4N9hRIbe6sYeDrto_3xf04ctGaBLWuYOQ7S_AEs6IgEtwLbWjphbYjJfySmFm1TsZInbzoCotbv-vtbg8qOuAywGYJQS6G1wlA=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF-7hX9jX_D1Zmvp9TRp_SzGYoexNPP91gJtOB6n5rALt4QVz-71vGMZmFH9KX9t5U1CsaFM9cSZh8DWskqrY3BXa5_xI_0wQdFjqB_CEBTwCwsRmwMMTxUNT1hnWcQvEAdVgED_DliA_tcWPs6grAQnmPIFS7dEmMOKBdVKMKaekIFPggslzNt0r5ZaN5XQqb3Q8-X0yUYmDb11vn8v26fDmDyIo6frvuqAmyxL0cBgJ4bSJ_x9fc=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHSRbMqqfT2kFD2YuQOp5bNnB1-LHyoagpLHIF5TUMEvj-2qChwBuCcOQs4QWv9S7L2bb6ykujbYun8nsQCfoX4JXz4zY5XkA7tJvQzyly2lFQ6JAy5jwt7y5n1f_iGP7gCFq99DJuB9a-gnazO1wHTZpR3Okvq68Cvd71eM5ImcGOvkJ7dwZUPrlbx784nxmRm4E6SYsM=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF0jn1dEEkA9vLp1ItIDDiLwWkRIr3jMDlAhBhCBtkVnADlTX4Xflxh16KxrLREFUXw4xyv8BZYvD8FNm_IrjEy_N-Q5RZClIpBjQvKytcH7R7HjZwXDvYIuEnd3nL_8D1s8sRfGv6X5vLZ6NHYzwvQCLvbb3T4_g==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFZKUH0MhrhBaxneRersPgN4QyLPnctZJW36qzowbSb7uY-5_laYQ6Fg1rmBxirQmquTpaUnX3HXRq3ZVWzaoI58k8zecbt_HrQzb5MUW7HG3XjJp-NNVFP_Nt3kYMbJCFlAL9GR6oiEHXzMCG55Czwdrnh_zw2s7vn_eCeRUP1JRZwGvr_yN6qpnejhy5i7uhE-tRZXvcJCuMau7h-XzsDhkfK2fQgDdpQq9-8OMDEQk8i1G4pOuuN2Wor6DLGzIDb_GcZ3Spi
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGDgg91hqgyHcXwnMLq9EfZHUOTsAL-sbq5yRuF_qjH4tEfglbQmFLcvCca-LRecbJHuki92s9dvVAxPgZgJHOBdHqxWpq_1-Hm3yLmwWqwfcmQE_25iDfgXtOiRsmPweZACXzh3TvkbQ8enQXLd1e3R5HWC3E3N90jFIYkhZ-w4kJlMxmXNzKWho3yh9uwhy1oLWkCO9o_PT1RCzfsUg==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGwz7EqdQrtpKWro8Fxfpj4_7TqIr8feh_J7JLOFLEbQ9S5nh4wNW1OBZfP51cuf7DaEoe8TIaeZD6JX432NW3VfBjXDcD-NSVnH3ZKafRAwH7AUntAd1odxz-gHEr84Y7BOFxUeTZ4Pzjl_4vTm7FastAn4g8nH35sQBL2Vwwa8zQvZcKf9oCu8_QIww8vjTdbs1AuZmYGRSQVJD-LKkXXDuLR6o-6aV1B7EUj1g==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG_Ff7nDkEOs-vEqjbzJZQytwkNryh5iOLZLRKlLYtnHAAF9GF6SgYlaY5Ng_RuJLs3Tq9XVvcM6sdBZTCozT01jWJhF3zs9TrnfnJdOkvYZCQqGf2kR2loie8tpQMUieyy-CcxMwMVOXkaQbgnWKe_Pui5bUMau6YnoNecRUal8iCIeDF4RF8Lpm_3-ib9Ya87jLtqpjwH5f33XlCbYcX4rI-2-ONDNV9kFGv8lB0aCaEKuMcBGffAH6rQb2Wii7dWdgrFVT4csIlSPZXr
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHvlFg-vIlRhMhpnOkQVcx3jQBsHAxL6sFSrTlf0cnspfPFiNRUbddCzj3UJCIGrps_KyjKYGt1m8XYpaVzpDeqa9s2iuPT4YqjL4VzEOx2OCGsBHPY0WxpoDWzjlR_-j7HLxgF65aKwf7WdTeO7wiQm7_RITU7AH5aVriSaOYQ6SQYPSQwET29UptT6z9af9iaKKtwRjyFXSRPJ_Hrcvd2c0gIaXmtKzk-HV4Cp6vIpfuPxk6UYUKdUA==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHPmDZzGZbtcjENCAY_AH_6dInOZRgOofMV5MnQEdoeKB-AeqwvYzmev5XDhJEOmdIzTNkUEnyftmYfhQe4x5kk_n1ztLhp_IRdcLozrua2pzuEfvMC_cmIw4DQsiKYD4bHkcjh91evzEs8
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGvisuXKLrLzcuQj9jVbFdHsu20rbRHx_XKx4nQZsHgGwbIxtN0gt_sjuEUg-ppjcWrzhCMgIqDND_weX1g42VUnli__qaLFb-jwT5w_wMM92OjCwZw0l3n51TXpDkZFlYooTUGIXmH0keZ9YvVxnvHnzHxqZYB13hv2zTCUySQLyf0c3O6bzUK4HrAZ90nKC-KOClUn_D10RBO_6RLi4hJYOtlDh1gsaBr3GDLjMkDMdlQwZOM-A==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHDPW3x-1QNAgDSCd1asx5KDvTTdcB6yAiUNXvSpHPPFfbVUC9emTYVRfSjLFzZk7jFYEoJSbvqA5et8pq39QBkQPtKtrbPLGqhB3wqtu5N0xLyOpTfJ52_kXCalu44ntORtEnOTT-6kldiB5ukqR-pMz2WctBtmgUqrtpFQ8-pVw-q94Y_1U1vUf2DkuYQ570-rdmwgsNIZLJJJoehq7_A2IFj9chMB2RxuI2Clwzh0Q1Y6Q==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHiXF2m-HtV8HszByKOU-LpI71whyr9CKZeoBF9x5sSoOTc111btjxzLSV4vTPreNjtbNJI0bleFcj6LfJigGFkOGZDfN7UG_tAb52jwvnnPQE7DnL5r3c8teksGYCOlaFXn-NGGFis4XcilgSJUlHu45F6PLPig_U_tLOVKikmrwDTRdl6SyIlvZbFq8nTMA2Hk5rmyn_xa9a8-Uu4vhVjzIg8_Hi5xX3JAHx2KDRekxzU-7270RVEeWj0-ROgND0=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGsnH7YWWVfDcLnsExkxovezI_jJrb2erGw0EG7rW5mBJqkngRq8lUOPLr5KNAx1hnVEUyV_YTltFcplSPnyoP5uzSI5T2cCyPQcy4hijLdvbIE2nsXOtqCvU7mBIbG8djMsPHVyiXcDBhwb68r_ySUh8We8u5R8DrW2q04Hd9StHg2YL5TCP_8MUdM3bZNH8ddU2xsnfOABvvUjGmCbmSzkxSVipPXbD9AqeWGlQGuA3h3wQm2V0ckLYrByYxxqTYMJovng79MeB0ELOyemZZy5ASc4Ys=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFVoPdxWEQN4IhIXL-DubfKoFtWRvEN5j_e60KqEat3_1PorkUJNtyR-tU50URTPXoZcfkoA5G2FYnUpRZL_IeAIWDZr7cmCiiK_No-bi-hZoU84hXCO91XSdkl6tTSN_kcs3bHgnfD7ViVMPmGwuOt2yq1IvQdmvmss8IEXmgwaa8PJ4YM5d2_sjlTuZk77h2RPB0=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEl-le5AJPAxX6rczEcKzZQfqYl8blHUbtiK-qsMnP7DbmxhaBp5Cm-pII6H36beml2KD-bAUPwiioDF4SkWAYmcSDuxTrhIr-whaKXsNIOTQ3JD6kZnVzyx4f750_geYXUZe4DAwaschdNrtiYzNn7HmX3PfH_DBaNCIagDfhkd75HRj2NAifMtxY7_btx8aH6znkkeu6R3hcH
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFEv7mFyilEQRuUQR67ibsQgbmWdV0L-jJyO69YUVxr79YV7WTGLgtdmOL3gQBLqm26AXONAdCMQSQCZREwmpREcF8wQ7jhpJd4xkD2Bg1TJnvjNIML8aWR3GYVydguE0l8yAHtG3YZzhRCV3hvp00nGPHfziJ6JgG_DzRuu3UUJOpjMJh84A==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHkIXC_ibUXIOQcw7SeG_Mm7GKxt5FfT72dnJJu6U-Mo-zgdvzG6P6hMv_SwSPpoWfRYmQKLh_2H0TwSRSsN8WGJnqDpI9WpKhQBGYEgq6ryWbjtX7d1Mw2SvxEIGm9e48-mEDOlUEfOtkh2VekMIGZYM1k-RhYElNKx_WZSAAGShetS0Hc-9L3zlbtWuIe3yk=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG3rs6WUSOyNzilvoks8cZgb7-vuCtpyalELoiYpPc8Mg29JFlIabrln7oUcYzNQ5jMuAFnJ9BYKZ-P8QGZ-I7vW4-pbHN_Rf_DcnbCf3Vp_-C_XJAqCgTz7aiOrliItrBz8LQFeivlUbGo9WuD6ktSQlE9e3hZpJkLyqv5TtQo-rbdRhgGYIj4v6BC_C47p8ZbZ1U7n4bgJLylUJCoEWYRTwLQ0kBaVi2GBlqfwmf8Rmuu1pGvcgKLVmjBQ6C2mh-h
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE2nmwrlRex23P0VhUZYnfxx-KR_XRG5Q7p7_a1ypaujNd96_4urzRj2IYhYcnJCQdXGCxnd1krNd4svjkcJUebCAicZsZa4ykJrkd2z4MrUZjUhDqjsZnH5v6-heO4aWngO-8LezXj-kjrcL5M9O9U6eRlG6-zgCGjw1vmcnCiEfgd7w0UKSGj5Qi3
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFQZohuHUOWeC3ZrlkLpJtmMKUZ3irHwl36y2dKl8xaM_u_42XeskXzTEq5pHzgRyGuVTkpCKRHlj-l_GnaWMXNNWhRd6XC3UFyAziFMzG5M0cMjEiE5QLdLM-_W2RrjAN4Z6u9sT0nJhuiK8pnV0RpAYqwl5toq9bl2AWOKohx3fZ8tLojcLZN6m9Drtyz5GSMC9DWTCuvqjk525s2Ahn2Tqf1MEqNk9VrdlUPrsHC70AlXnj9g5d9Ap57vXr7Lh8nov4UU3KeSIOADc2UlIgndchg3qf9kccCMA==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEd50gPvUs7onuQ-uBBhM-JaAq_EE6VQJmoLRs7Ay5p-arjgvllY5qznDditY__Rw-pcaPHB72kZXyG22aO_ZnsG74rgtwjGAWD7Gp-fo5dodyyIQWVh0BRIjGLrySW_hvbS2fB-4q_SSGQSy74GnmEKqNEF9RcgYLMpTb9Ukh4OgbMJZZvWY7ixD_9WicwzDuVvn7UDZ_k-wD5i1gX3Uuxs1_nJnqZ2QEWCX-qWwgPSnW5D8_l
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFQTifM4wpP7MMnz6PRvkTIIZVep0cv6olBzCQ0rlrgXvRuKGlEwpxgCDg1l2TRERaJy1a9Wf6jI60VG2K_anQ5Lk9Fj50CAd06OSEGGP2RNCmiVQBWIt9IDeAybuW7yCM_qmANSaRzPMCgjFaNCaiCEDao2v2nrvxJPxcZGFiA8WAVMMPGq9WVGRAzZgW7YJjrh2z-cA6apZTd-ecRPkk3Mf3vhmipyG_74zSabBC2-vE36k9lQDQZ5riaV1dHn8ZpoD-kh7chSHx-PVVwlg9vNldxfcCVsYWEK3wZz8Db2zqY5UCZGe5Gr7aRvNshKp4nxUeic-72TNX1Ic3gXk5-Eg==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGdt31QPkXpfOPWmq8LEvo2jBDmmu5bUtYvZ6PkSruy9tKpTRqsP3BfXnSWERwyQzGNQwMt-sjH6cWalsP5mRm5Yjl_oTv7Ft_eNlEcIkBXNzyJwLh3IqMUJQrkEVUhEUt2BivvlwvC0Of6HxCz-Hb_MCUrDNA_fAW2ryoSqJBXAP6uLV2g7EmE5Rk=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEwyA1iKYf_upXJ0gQW5zajuzxiU2LMfAFgHFJoEWDEVmkhx_jpFcyg7tHI2DpMZRJ22gI5FvmRwKrdZFAI4rJMLHG4dVgS5bCbe9Q0jGWhQIjdTSv_drtFwZzTvNUDDQqkbKMe_NwuAaid0PcJgG9b42AZBD3cGtRMYIzPf4_021aF1IKE7ET6aDKpwuCiJ6UIWFKU2KgpToivnF1E-AcT6asLKPFvLgPM6g6cBmk=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGCHokPltQIUfw-CvgRMSXFrrWO74AVoH2uuW39XnOjWgHOJVHIuMog2j2ywubKMvA3HRs8aHCmmf1VxGneJZbsRHnBZDGQkXEkLt7E4D632g3H0rI1ZNdFrY434b2DxTvW1NBY3Q==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE8R7Jxgtq-TXUsN2slb2-HBnxTukXh5KPpfa-vXLRKG6LSaUuakDmNCT6jOdDoReqZm8SPQ7Sk3ve0xYc9csbD7LBLtCJo_iYFyhyKA5GXnJCFCkorN8K0zRi2tXF4eLqvzJdyrinDMCMiSKSv357waXxvQFc6gI_8s_HkcMMswxtG-eM=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFv3U1gUJKizilupyz7yBw6b2JKFHMbWNbbZM5D44V2vtkSCWSXCs2Rh9v1uUkacfL_cKg1Grqcm9KNaMVYOyEzhGQ2JNa9Ql-qTwMl6Mlu4LISKuA6elSrbBXo7furEkOuqk6pVDtD4U4XKOWEL7Ijnt3NIUKmvDDptYMObPk7svAL
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHwzhvOxYlb84HrHit_lEiCju5RRrn6jQGYP1sGAKz1WuGpgVZ-hoFcYWQwkBbiqQTEEzMfzWGMeTGtnsESELSqR66Z1N5CgsQ6aSacSn12MBH-_HEfrifhmvYB7YFZyOpMfr8FbImxgYEZbTT4mNEko7bGoSHHAN-VKm6IEMm4rFa9F5eXYDSxW-JZCqsePgzcYrUHMR14Yu__H3c-dK2U8j3crD5q8eRpbor2fzTleE1YzL4tFv3pXbrmZm0lVLK6uFbaEzE3uLgndkxBG707NanH4JKKLf536mBrY1pw8UYS7STK8MomWtRyzmkazBMYSbl86IGKE_pH7xLGpPKrjC-n9WF28pSH5As13bFhW5C-LdbktKw=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFRKXyc5YXJbvyaLYN64YuJ9TMAvKWQ-OviDdd5DlYLmKJggQt5eNg8OG8SnKbMkiArL9MPNcCL4Yoprgfg2rYtm95f9J2A1yxsw98uGLJfsj_DRheu9hov6czyuGrPZ_Pzh90G6pbLxhak4MDptCTaJP_N7l59t-1lpfWHr-Uk2yBkRGpO-0D5ahKHW8LdQrRd4uGKQIwX1kE5402-vHyXX3yvFmDgelMfpC8Fh4fX
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEQiB2kjhNMape5IiX9dP4iuec9PsfBGVXREv20qZvr5TrvJmJTTmcb8Ybd8-ehQB2ikO53DE0YlRMEt16abiQJqHgZfs_wCOYXTiAkJ-P4F6fwInDf6sutHdY7veNiBZACGHnjpfFAfIP1qNWUhtfpz6x-cT53ZLaP2xzQnoCGJrQuJBY87fDBlZqK6vGElGtyZLin3HGYhRGeYqva_3FA5VWGk02a4fNmeuuzVc2Z4WYM
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFWbAOjb_sBeqsPvl9LCp4Au1Jk5VXs7ALvWYrsVRI3IBeP6SVn4nsvtr2SUZimlNZS_8cjFi7P_pV2HjeGJ2vKVj6hc-g90yGbxb0JNIjZUtaV5Xry1_rBr2JFoQyh2fI3Puv66sBPTRIerzCBnlkQBB8NgGm9KPYPNskhnc8s_nvT90XBbMZr_PcMzuth-kT6lD9MauGzrfHB5q8468FfbIEmIF30DmJKatqnb4FyXQ4l5qSnBypbgPTAamYF9j7XOlWdSJVmbn04cxX7a4FSwRY=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGmqhZcGY4z-VsC8XQQv2q4bYN1SQ8DKNRiF_dumBXTI3bLtTjdXs489LiS2opzoGruQi9d6pXU9jDAvj33E801t_73tup2aeNPWKLvX5ocvAkdMNFa5c-rCGlGg8FoKBijXr3pF1Vkhq19nPHft_3uBUDB7zeMhd1ndjuV2YzYQjjqybBPXs5B_ko_hOf4Upy5kXKAUrfXz2BIi2ZBLvGeiK-oGkaVUV31p4S4gpsfln9FcF9r3Z8vJcOJFOTD_F58jPT4eER6z4j5_JUUgNzsr4X0h6eQmLsz5Z_cooEKNo4FI3g7-aIqtA==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQECVuJv1zVpBpsHoyJBHF9ryONmfAbwDFVN6fFVwee1nwMWwwqnArNJculeoX7mry0HcycdU69JWlD4ayfVjuJ8ZYMwZuEfbafdQT0zhpPH3mik8zbxAkqTGNFCeVjb4YhMiHEoXEWR2B3Z6bwSfAhj3q70HO8Gs-N7vNgg4lLW1_UVdT0UyWsoBz2IXVZDn7FC6KqAEdUI1Hb8cf_1ZQfDA_y29HSWnVmXlSJiLNNw66ptq-Tsl7aG
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE2wyS-f336rI6p6_7m4O8XbgfXw-ipZOh1GYKNf4fVcsJ6cU31jYNH0o2tm0f0Yiwj7KzU1yAwPnkpB4_vQU75BAxSoL6FhQJnGP-wUPm9RzuPy9RIX9aVvEWSZwEEqhRXCdXaoUPNWD_WzBHCXSxnOQ14kgbOEyN_mfgldbNy4GxKBscOVj0j2WHr8uTL0MuVkuf9Q9MwKcjdgpBtx4ixUsGhQLQfTBwwoK422YpWyHqCFkw4H3QJScfQP3DIwPJaw_jjfiHAB37Rg_viPgHYp3OfxUpc
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF5swtk-a4CkqIHS8xLp4uXMRiXACbJSyHRIO7RLZinowQx5QCISJKepETi1sqhc6Ot33v5sGpMtHGGWrzN7L2qXcaGoEnRR6I6J8VJ-WSpH16R3sCkFTi4UNYWz_Xyb5w48l2LCCAr7fk_Yy7mJ5BoV-leVM4_X_vZKDyCBcxdsH6NNJ0=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGO2X80rQfmspMEkyDLe2uvn9iMKpDGRBCHKhMAyKLu50ufsL1p6LRE8t5UR8zM6iZCKHIz73ldgoXoFKEoJwdpU0xM6FUwO7LClVZ84JJACt0hzUd5_XVjoqmI7Du7byNSirGBB-6ju625Ds8BjXysnYU6yan-KvLM6TYgCVvedsdUq7SO2JqNnNrQvbb16Sg7wkDThf_uBfo2p2JQVjMKKSA-qPf4bZvV0BiLY79nZ_06Vgck4REbmKU5wwE-quzg3lIg2h1Xa49sVNH6xC_ku45y46QewtqmMikYfEuUV6aBpFcdb0LvpYm96us_zVe2LeKVTXfe-4b1FKSuBg5RG2FV3KGbPf4=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEC8_0zbr8I94qt3-2-wfSgLgPnIf6ED3mtAhixgn48ztwIoRfRZO3FnNXvw3eAjbOswvBQ4gbzleeMxZVzLahx9PLvWFb_kaHEmHnH1lszn1INeaXjkpjcebOrgGPoovH3NLp7AggK2dNkbwCbtqRuPr_HkX0_z4pRu4rdxEWdM-fnt-cr_95vYaWI1GaRZoeLQQA5US6fxZUrtw-_1OUmTx9vUb-pBOkzj1IZjCN3k5qUkZHhkfvyjph1FsE=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH9XBo9OXxuoYkkgCvHlb6rDLc1S2PiIeWu7tV3FIyF7yIbRlVt_97a7ES9IxcuRyTGhoRR8-w774kc3f7t-EDrfQuxyPZuI69TSHtezqcrvzP7OzQt73vomJbwaZkuFgjMIIG5tw8Rx4IafWvj3TrNYWuwZsGpNVuWVIE-zf0-KBlpt1nC0iRmcRqtm2vv23bwP_1NUnc1B0MsK6QICui4zEEPCTs10pbdqv32x9I5bSC2nsTkUeDHBusVnMpItw==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHSLg-yknyRYOCTbVnDz1QvOkzpSNtBFoQ79mXWj_F5i5XiRpJ0CctX-KKe97hoBtG9oDPV2mXGL2ZjddAk5e9EK-CcMY4fLZyGliVudlLvsUgeMoNLlHtccBwGwGn1ifOu2nPBub4M46iaITWdJ9JN2IgsqkMeRPwbK4OLZbzmASxGHFTZAvf5w8Nw1Txc4oycmSa8YEmEnfmhEOrIAEhF-7XGJoj-uQ8=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGes6OS9cgvgB-qrgeLKoVDTMWW4ZM9zzNNd-mNtqgyM2dTHNZ1NGlm4gVcWUBu_MnMbbbWy7CMjXwxZrYW7Pyks3qK6pfafEPB7UfbF3MPrzmgZG2U4hDy7HOBf6m6xqFDt-dPFZOheCUv-bwFTGBRBBK_NZTGKp-s_pJmSxTp6dcG75OTCKHrTzpn48V1ZF-KuDr-hU_j-gCXI3ibZodOKDvmrBsSkYxlpFdDiTugniDC-alkkFm_Lg0v
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFf59Qi0ITchn59G1sp8MGI8Gh38b8P8k6CNJCdzCFLzu0HZlA66E1y9F6f8mbasZC5wfLfUB8Smn0F7Oo9rKDsaOGLi1kpwp6J1NTx5bgIcu3fceLAcYFxWrxLZD11lPXNLWK-os9LCg08g97I1OmbixXPFVtgCZ-au-3ub5pkfbbQjdp9LnimwwkHDC3EBPlydhjUmnG7gnCuboE=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQElpsJDspKRkXIcahn341seA7CVMiM6lj2Pj_USq9qm_EIIWOgEhMJAYkkI9m4EX3TKw9gWq182X8AHrMdX3gyJGa1R8umss9dydFhX25kQhd6hCQTYu3qO_dUeS1vgb7z0CdT5DMiiwsv3E-grUhTF3slraQV-LGBP_MCPK0ub4bbbqcmn6t1JckDOb9HnNFNdMX8p8eC1kEtIGA_cHm1fExq8LhB5ETH1mBUkuQ==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEOofpSXCVuuSqo0o97CyD7Qk2ePf8vnWBBXM-5_JfMEm1UclVTAfIG8zFpnQ2kQot2wLlZAYO8CSdPKBfFMqB-lxqSWmAesgL6eiykz1r_WuzwkeeCFoONZKvAZ2LH6y5YSOXdNIrAL364vlvdITZBcFFj_ya-gr4LdqetjrBt0iFPGP_iO2sHc-IHVdF9RQPjCBaqXJym6pknDR1CSA==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFnfKJkVQrg3Xw1u1oOSS4HUZTjRYVcvL_4ah9HfiuhyhS3nUZkMuQc54vqB2i3XC3OQOff3i_jcpdsFUWr4iLzC1GWDiXLDy1NdzvmLV1V8iohpT7IteMaEihHYtIR0Qf5BrrDj7spQiyfN_m1sXxQkmVxo29XBP2QkgB4sXE=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHEKqubb404cQvS_Ex4giT9aCXKstRY7lVE8qpsg2USlkOVe7h6OUZ6Fj2w8OryoLR8fRv-oIh8vInGWwW2i4adyxKWqWF5Ar6UvmfLXoYEgpv97imCIpcsK8YFEPOgo_phZrJQzAWCnQFBcgt3e_u6jzvr2LVVyD1Ds4NRKcne-YngDSaL7_hu6CSmRaOA
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHpf8BGaAD0hBqv3WwJH3MehN5Nsi3etpXOLjr8lpqSUEpVsUXlJ6niMnNj9a0ERguFUtV1vaSXxQIfMaOSeB4na6gWYRL4bmVLzpTeGj2bEPg9CakApvVNcH4MXawGk2gVY081xOknyhgpGc5bMj50AR3_lxsQoNuhVf0iSPpXZB0Q9yMStr7iddqPKHybFA==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF3S6V7lmSxvBSIuliUodzGLdluEX4-5YqaoGOSAX6lEZVISliLQZG0Q3ZvHLL7kM9wqB1Y18XDrZnLh_cUZ1MEdUDWjUjFLql21qR4rJGmMD8PXt6oM3VAa-WHoGSCv2o0NKwpzNQplyDolIBskX76eSDa1_dM9QKRKmxmGqXxwDedUu7gxqoFYgnVnrSMskOYSgAs5nLwwyclifP9
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHgdfJJS6hR3nEJx2zQ8_PD8ETxosFAnvf5a8N3ucBWJFhtcfcvcAKGo3AqpE1aGdX5VHdcEfLzDnEqSs5wsN-uMy3fitDWGut-G63Zpvt-45SJ5P-Rpchj_bxjvIurwOkLBw50ivCAn1f3Tp0bOfmcQMfP1mkhdBI59_qxCR-cDUjqY5sYC9IzgqPFIxYbqdGwOTen-IICk2nJXi67roRw0hCw2u9AeirzPedyodD76OB4KwcUjiktySw=
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH2IR_vO2do9eBNHrGCL2rwveTer_k6WSMgk7N_OOsX56vK0zmww72fxC42t8km95_RAqcjr308K7QCEV_-w0lzkjt4KqmLaxaiegfTsantE2hiBJQYPZbQ6uLqfY_ezaViZBjIFgkEGZ8Rl1rnd_IZ3XsynDeQG6cs-iRDumYqd_uy6PuzBBiNR5huwObmJAavqJ-BZnk2kSJ0mOCH5opO
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFgEh7-vDxz8knCq5kfV4IuXqlIlmzOjYI530803tTJExpkIwqxGJXo2mcChbuTLtJFG2mQfUi4zXcPETz9-rcRaFD9iUvdwvU8XzHi6B3ylP1k3_618OAE1EAL3jpzcR4qlZSLQVM38ZXygMPLuwc5UDfTLQbLQez8qHrN_PmEXjTaF7Z7BAnkGsSEsN01VSXY89wF8xpJCCfx8FysBAgnXQ==
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF-naRARzeKTYgnBW-dGh4VwVVqhH5YWAnMVv3oDoQAxKNG4lCUty3nxZWdXSRIE1ah357bhXyGdywdlJZhH6mexpk0DqgWxsXBo1iORL3mRnepcQrUyFArD_GwHk0tWatkkBrI7Y-8Oq9h4X7-I5fgPJEln1xT3JRiPujEgoPjo7CKCRW7WRpdOUydY_iZI8zEFOLyDHvcwgFZSRBn0tpRNu_qosW9xwLabhJYEqn93SvE
• https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGi09KZzYC82da_68CqMSJbsVs6CVLEhJ9WuejB-u1U0OFVP_42qLFmOMMXzJdZE7Bh-AoY44tIOuZXyUrYS8O591alWGJ1Sip4LCIBdSZvbg38cuY8PHu8Fw2vyc8pAiJPaZghXUbbq6_MwKpnXn1tNUxXPzqEP5H2Hdewvt90bFjrXp4jKPH7zYJQU9nIvahoJJY2r_Yq