The GSE AI Governance Wave: How Fannie Mae, Freddie Mac, and Federal Regulators Built a Multi-Layered Oversight System for AI in Mortgage Lending

The GSE AI Governance Wave: How Fannie Mae, Freddie Mac, and Federal Regulators Built a Multi-Layered Oversight System for AI in Mortgage Lending

From underwriting to servicing, AI is now embedded across the mortgage lifecycle — and a cascade of GSE frameworks, federal rules, and state enforcement actions has created a new compliance baseline that no lender can ignore.


AI as Mortgage Infrastructure

Artificial intelligence is no longer experimental in residential mortgage lending. It is embedded across the loan lifecycle — from borrower intake and underwriting to pricing, fraud detection, document analysis, servicing, loss mitigation, and quality control. As one legal analysis put it, AI now functions as “essential production infrastructure” rather than a pilot initiative, touching nearly every technology-enabled function in the mortgage pipeline.[1]

What has changed is not whether lenders can use AI — they can — but whether they can defend it. In 2025 and 2026, a cascade of governance frameworks from Fannie Mae, Freddie Mac, the Federal Housing Finance Agency (FHFA), the Consumer Financial Protection Bureau (CFPB), and state attorneys general has translated longstanding fair lending expectations into concrete, auditable AI governance requirements. The result is a multi-layered oversight system in which a lender’s ability to sell loans into the secondary market now depends on its ability to document, test, monitor, and explain every AI system it deploys.


The GSE Frameworks: Fannie Mae and Freddie Mac Set the Baseline

For mortgage lenders and servicers, including nonbank lenders, compliance with Fannie Mae and Freddie Mac requirements is foundational — these government-sponsored enterprises (GSEs) are the primary buyers of residential mortgages in the secondary market. In late 2025 and early 2026, both GSEs issued formal AI governance frameworks that directly condition a lender’s ability to sell loans.[1]

Fannie Mae: Principles-Based Governance

In April 2026, Fannie Mae issued Lender Letter LL-2026-04, establishing a governance framework for any approved seller or servicer using artificial intelligence or machine learning (AI/ML) in connection with the origination or servicing of loans sold to or guaranteed by Fannie Mae.[2]

The letter applies broadly — it does not distinguish between internally developed tools and vendor-provided systems, nor does it limit its scope to underwriting. Instead, it captures any AI or ML system used in origination or servicing activities. Its effective date is August 6, 2026.[1]

Under Fannie Mae’s framework, seller-servicers must:

  • Maintain written policies and procedures covering the full lifecycle of any AI/ML system — development, implementation, use, maintenance, and retirement — reviewed and updated at least annually. The policies must be communicated to relevant staff, grounded in applicable legal and regulatory requirements, calibrated to the institution’s own risk tolerance, and assigned to a designated owner.[2]
  • Comply with information security obligations, including Fannie Mae’s Information Security and Business Resiliency Supplement.[2]
  • Manage vendor and subcontractor AI risk — applying governance standards to third-party AI use that are “no less protective” than the controls applied internally.[2]

The logic here is straightforward: if a lender cannot identify an AI system, it cannot test, monitor, restrict, or explain it. Fannie Mae’s principles-based approach provides the structural bones of a governance program — defining what must exist — while leaving the operational details to the lender. Upon request, Fannie Mae reserves the right to ask sellers and servicers why AI/ML is being used, what purposes it serves, and what safeguards are in place.[1]

Freddie Mac: An Operational Checklist

Freddie Mac acted earlier and more prescriptively. Through Guide Bulletin 2025-16, issued December 3, 2025, Freddie Mac amended its Single-Family Seller/Servicer Guide to impose detailed governance, audit, and security requirements for AI and ML systems, codified principally in Section 1302.8 of the Guide.[3]

This bulletin took effect on March 3, 2026 — months before Fannie Mae’s framework.[1] Where Fannie Mae provides principles, Freddie Mac delivers an operational checklist. Sellers and servicers must:

  • Actively assess AI/ML systems for specific attack vectors, including model inversion, data poisoning, and prompt injection.[1]
  • Conduct regular internal and external audits measured against named industry standards — specifically NIST 800-53 and ISO 27001.[1]
  • Maintain ongoing monitoring for performance degradation and bias over time.[3]
  • Implement segregation of duties with documented accountability structures and lines of communication.[1]
  • Obtain senior management approval of AI/ML policies, including executives such as the chief information officer, chief technology officer, chief information security officer, or chief risk officer.[1]

Unlike the Fannie Mae letter, the Freddie Mac guidance includes a broad indemnification obligation — because it is part of the Seller/Servicer Guide, it requires sellers and servicers to hold Freddie Mac harmless from losses or liability arising from their use of AI/ML.[1] This is a significant enforcement lever: a lender that cannot demonstrate adequate AI governance could face not only loss of GSE eligibility but direct financial liability.

For dual-approved seller-servicers, the practical advice from legal practitioners is to build a single enterprise AI governance program that satisfies Freddie Mac’s stricter operational demands, then map each control to both GSEs’ obligations individually.[4]


The FHFA Foundation: Supervisory Expectations for the Enterprises

The GSE frameworks did not emerge in a vacuum. They build upon supervisory guidance from the FHFA, the independent agency responsible for regulating Fannie Mae, Freddie Mac, and the Federal Home Loan Bank System. FHFA’s Advisory Bulletin AB 2022-02, originally issued in February 2022 and revised in May 2025, provides guidance to the Enterprises on AI and ML risk management.[5]

The bulletin establishes five core ethical principles that should guide AI/ML use:[5]

  • Transparency — adequate clarity regarding how and why AI/ML is used, with sufficient interpretability and explainability for objective assessment and validation.
  • Accountability — assigning appropriate human responsibility for AI/ML outcomes with adequate explanation and justification throughout each lifecycle stage.
  • Fairness and Equity — implementing processes that drive fair and equitable AI/ML outcomes across different groups, evaluated in light of social, economic, political, or cultural biases.
  • Reliability — designing AI/ML capabilities to operate as intended throughout each lifecycle stage, accounting for purpose, values, accuracy, and safety.
  • Privacy and Security — respecting and protecting privacy rights and data used throughout each lifecycle stage using industry best practices.

The FHFA bulletin also requires Enterprises to maintain a comprehensive AI/ML inventory — capturing use cases across business lines, from proof-of-concept through production, including both internally developed tools and third-party vendor systems.[5] The FHFA explicitly warns that as AI/ML becomes more automated and integrated into business processes, “the interconnected nature of the risks can introduce more complexity in risk management” and that “[r]eliance on AI/ML without sufficient risk oversight and transparency can create heightened risks for the Enterprises.”[5]

The FHFA framework matters because it is the supervisory document that GSE examinations are measured against. Fannie Mae and Freddie Mac have now translated these supervisory expectations into binding requirements for the lenders in their networks — extending the oversight chain from regulator to GSE to seller-servicer.[6]


The CFPB Layer: Adverse Action Explainability and the Reg B Rollback

The CFPB has repeatedly made clear that AI does not exempt lenders from existing consumer protection obligations. The bureau’s position is succinct: “Creditors must be able to specifically explain their reasons for denial. There is no special exemption for artificial intelligence.”[7]

This principle has been reinforced through multiple guidance documents. The CFPB’s earlier circular clarified that the Equal Credit Opportunity Act (ECOA) requires creditors to explain the specific reasons for adverse actions, even when companies use complex algorithms and black-box credit models that make it difficult to identify those reasons.[7] On May 5, 2026, the CFPB issued Circular 2026-03, advising that lenders using complex algorithms — such as machine-learning underwriting models — remain fully responsible under ECOA and Regulation B for providing specific, accurate reasons for adverse action.[8]

The practical implication is significant. If a vendor’s model cannot clearly explain why a consumer was denied, offered different pricing, or assigned an alternate pathway, the lender faces a Regulation B issue — regardless of whether the lender built the model.[6] Lenders cannot simply claim that the black box “told us to do it”; proprietary or “uninterpretable” models do not excuse compliance.[8]

The Regulation B Disparate-Impact Removal

In a separate action, the CFPB removed disparate-impact (effects-test) provisions from Regulation B and took the position that ECOA does not recognize disparate-impact liability, in a final rule effective July 21, 2026.[4] This has led some to question whether AI governance programs still need to address unintentional bias.

The answer from legal practitioners is emphatic: the rollback does not eliminate fair lending risk for AI.[4] Disparate-impact theories remain available under multiple other frameworks:

  • The Fair Housing Act, which continues to recognize disparate-impact liability under the Supreme Court’s 2015 decision in Texas Department of Housing and Community Affairs v. Inclusive Communities Project.[1]
  • GSE contractual requirements, including the Fannie Mae and Freddie Mac seller/servicer standards described above.[4]
  • State fair lending regimes — California, New York, Colorado, Illinois, and New Jersey all maintain anti-discrimination frameworks that can reach algorithmic practices.[4]

Disparate-treatment theories under ECOA also remain fully intact, including claims based on the use of facially neutral criteria as proxies for protected characteristics with discriminatory intent.[4] The core takeaway: lenders that treat the Reg B rollback as a license to relax AI bias controls may find themselves rebuilding their frameworks under less favorable conditions later.[4]


The AVM Rule: Quality Control Standards for Automated Valuations

One of the most concrete federal AI-adjacent rules in mortgage lending targets automated valuation models (AVMs) — the computerized models used to estimate property values without a traditional human appraisal. In July 2024, six federal agencies — the OCC, Federal Reserve, FDIC, NCUA, CFPB, and FHFA — issued a joint final rule establishing quality control standards for AVMs, with an effective date of October 1, 2025.[9]

The rule, implementing Section 1125 of Title XI of the Dodd-Frank Act, requires institutions to adopt policies, procedures, and control systems that ensure AVMs meet quality-control standards.[9] Critically, the rule includes a nondiscrimination factor — requiring that AVMs adhere to standards designed to prevent biased valuations.[9] This connects AI appraisal technology directly to fair housing concerns, extending oversight from the lending decision itself to the property valuation that underpins it.


State Enforcement: The Massachusetts Earnest Settlement as a Warning

While federal agencies have been building frameworks, states have been pursuing enforcement. On July 10, 2025, Massachusetts Attorney General Andrea Joy Campbell announced a $2.5 million settlement with Earnest Operations LLC, a Delaware-based student loan company, resolving allegations that the company’s AI-driven lending practices violated consumer protection and fair lending laws.[10]

The case is notable as one of the first major state enforcement actions specifically targeting AI lending discrimination. According to the Attorney General’s Office, Earnest used algorithmic models to make lending decisions on loan eligibility, terms, and pricing, but failed to take reasonable measures to mitigate fair lending risks — including failing to test its models for disparate impact and training its models based on arbitrary, discretionary human decisions.[10]

The AGO alleged that Earnest’s use of a “Cohort Default Rate” variable — an average rate of loan defaults associated with a specific educational institution — resulted in disparate impact in approval rates and loan terms, with Black and Hispanic applicants more likely to be penalized than White applicants.[10] The company also used a “Knockout Rule” to automatically deny applications based on immigration status, and sent inaccurate adverse action notices that prevented applicants from understanding their own creditworthiness.[10]

As AG Campbell stated: “Earnest’s failure to comply with consumer protection and fair lending laws, including through its AI models, unfairly put historically marginalized student borrowers at risk of being denied loans or receiving unfavorable loan terms — impeding their chances of economic growth and opportunity.”[10]

Under the settlement, Earnest must implement a detailed corporate governance structure, develop written policies for responsible AI use, discontinue the Cohort Default Rate variable and the immigration-status Knockout Rule, and regularly report compliance to the AGO.[10] The settlement signals that state regulators will pursue AI-related fair lending enforcement under their own laws regardless of federal regulatory shifts — a point reinforced by New Jersey’s 2025 regulations under the Law Against Discrimination, which expressly codify disparate-impact standards for mortgage lending and AI-driven decision-making tools.[4]


The Operational Architecture: What Lenders Must Actually Build

Beneath the layering of GSE frameworks, CFPB guidance, and state enforcement lies a practical question: what does compliant AI governance actually look like? Legal practitioners analyzing the new requirements have identified a common operational architecture built on three pillars.[4]

1. The Model Scorecard

A model scorecard is more than a dashboard showing live metrics. It is a governance record that follows a model throughout its lifecycle — documenting what the model is intended to do, the populations it was developed and tested on, the data it relies upon, and the limitations placed on its use.[4] The scorecard should capture fairness testing, less discriminatory alternative analysis, explainability limitations, escalation thresholds, and where human review is required. Just as importantly, it should track how the model performs over time and identify when intervention is necessary.[4]

The core logic is that you cannot govern what you have not documented. A scorecard transforms AI governance from an abstract commitment into a concrete, auditable record that an examiner, investor, or plaintiff’s counsel can review.

2. Containment of Agentic AI

Agentic AI systems — which retrieve information dynamically, operate across multiple workflow stages, and generate recommendations beyond their narrow intended role — create risks different from traditional automation.[4] Problems can emerge when an agent accesses unauthorized data, recommends actions outside policy limits, or links otherwise acceptable steps in ways that create uneven outcomes across borrower populations.

Managing that risk requires three controls working together:[4]

  • Clear scope restrictions defining what the system can access and decide.
  • Human review triggers for defined exceptions.
  • A tested kill switch allowing the lender to disable the system when necessary.

More mature governance programs regularly test all three controls rather than assuming they will work when needed — a principle drawn from cybersecurity, where untested controls are assumed to fail.[4]

3. Vendor and Third-Party Oversight

One of the most consequential aspects of the new frameworks is their treatment of third-party AI systems. Fannie Mae requires lenders to manage vendor AI risk in a manner “no less protective” than internal controls.[1] For many lenders, this means revisiting vendor contracts and due diligence practices to ensure access to sufficient information about how AI systems function, how they are updated, and how risks are tested and mitigated.[1]

AI often enters through vendor upgrades, embedded platform functionality, marketing tools, productivity tools, and employee experimentation.[6] If the institution cannot identify the tool, it cannot test, monitor, restrict, or explain it — making a comprehensive AI inventory the starting point for any governance program.[6]


Why This Layered System Matters for AI Policy

The mortgage AI governance landscape illustrates a distinctive feature of American AI policy: rather than a single comprehensive federal statute, oversight is being built through the strategic layering of existing authorities. The FHFA supervises the GSEs, the GSEs impose requirements on seller-servicers, the CFPB enforces consumer protection law, six agencies jointly regulate AVMs, and state attorneys general enforce their own anti-discrimination statutes. No single actor controls the system, but each adds a constraint that lenders must satisfy.

This approach has strengths and weaknesses. Its strength is resilience — when the CFPB rolls back disparate-impact provisions under one statute, the Fair Housing Act, GSE contracts, and state laws continue to impose fair lending obligations on AI systems. Its weakness is complexity — a dual-approved seller-servicer must now navigate two differently structured GSE frameworks, CFPB adverse action guidance, an AVM quality control rule, and potentially dozens of state laws, all using somewhat different terminology and standards.

The GSE frameworks represent a particularly significant leverage point. Because Fannie Mae and Freddie Mac are the primary buyers of residential mortgages, their requirements function as de facto regulation for the entire mortgage industry — including nonbank lenders that might otherwise fall outside traditional banking supervision. By conditioning access to the secondary market on AI governance, the GSEs have created a compliance baseline that reaches far beyond what any single federal statute could achieve.

For AI policy more broadly, the mortgage sector offers a case study in how market structure can be leveraged for regulatory purposes. When a few large institutions control access to a critical market — as the GSEs do for mortgages — their governance requirements can extend oversight to thousands of downstream firms. Whether this model should be replicated in other domains, or whether it creates too much concentrated power in too few hands, is a question policymakers will increasingly face as AI governance frameworks proliferate across the economy.


References

[1] Harris Beach Murtha — [SECONDARY] Palfreyman, Brendan M. “Fannie Mae and Freddie Mac Set New AI Standards for Mortgage Lenders.” May 27, 2026. https://www.harrisbeachmurtha.com/insights/fannie-mae-and-freddie-mac-set-new-ai-standards-for-mortgage-lenders/

[2] Fannie Mae — [PRIMARY] Fannie Mae. “Lender Letter LL-2026-04: Governance Framework on Use of Artificial Intelligence and Machine Learning.” April 8, 2026. https://singlefamily.fanniemae.com/news-events/lender-letter-ll-2026-04-governance-framework-use-artificial-intelligence-and-machine-learning

[3] Freddie Mac — [PRIMARY] Freddie Mac. “Guide Bulletin 2025-16: AI/ML Governance Requirements for Seller/Servicers.” December 3, 2025. https://guide.freddiemac.com/app/guide/bulletin/2025-16

[4] National Mortgage Professional — [SECONDARY] Gapp, Ronald. “Governing AI In The Post-Reg B World: Scorecards, Guardrails, And The Operational Architecture That Actually Holds Up.” May 12, 2026. https://nationalmortgageprofessional.com/news/governing-ai-post-reg-b-world-scorecards-guardrails-and-operational-architecture-actually

[5] Federal Housing Finance Agency — [PRIMARY] FHFA. “Advisory Bulletin AB 2022-02: Artificial Intelligence/Machine Learning Risk Management.” February 10, 2022; revised May 2025. https://www.fhfa.gov/advisory-bulletin/ab-2022-02

[6] National Mortgage Professional — [SECONDARY] Brody, James. “AI Governance in Mortgage Banking: Moving From Policy Discussion To Regulatory Reality.” May 4, 2026. https://nationalmortgageprofessional.com/news/ai-governance-mortgage-banking-moving-policy-discussion-regulatory-reality

[7] Consumer Financial Protection Bureau — [PRIMARY] CFPB. “CFPB Issues Guidance on Credit Denials by Lenders Using Artificial Intelligence.” September 19, 2023. https://www.consumerfinance.gov/archive/newsroom/cfpb-issues-guidance-on-credit-denials-by-lenders-using-artificial-intelligence/

[8] National Mortgage Professional — [SECONDARY] Liput, Andrew. “CFPB Issues AI Underwriting Guidance On Adverse Action Notices.” July 9, 2026. https://nationalmortgageprofessional.com/news/cfpb-issues-ai-underwriting-guidance-adverse-action-notices

[9] Federal Register — [PRIMARY] OCC, Federal Reserve, FDIC, NCUA, CFPB, and FHFA. “Quality Control Standards for Automated Valuation Models.” 89 FR 64538. August 7, 2024. https://www.federalregister.gov/documents/2024/08/07/2024-16197/quality-control-standards-for-automated-valuation-models

[10] Massachusetts Office of the Attorney General — [PRIMARY] Massachusetts AGO. “AG Campbell Announces $2.5 Million Settlement With Student Loan Lender For Unlawful Practices Through AI Use, Other Consumer Protection Violations.” July 10, 2025. https://www.mass.gov/news/ag-campbell-announces-25-million-settlement-with-student-loan-lender-for-unlawful-practices-through-ai-use-other-consumer-protection-violations


AI automated and generated content. Human reviewed.

The Weekly AI Policy Brief

AI policy moves fast. Get the week’s developments, new trackers, and podcast episodes in one short email — free, ad-free, no spam.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *