AI Regulation Tracker / Enforcement shift

CFPB Examiners Are Stress-Testing Your AI Underwriting Models

The CFPB has made clear there is no AI exception to fair lending law. Examiners now test whether your underwriting and pricing models can prove adverse-action accuracy, fair-lending testing, and less discriminatory alternatives.

By Anthony Guerriero · Reviewed by The Leveraged Years Editorial Desk · Published July 4, 2026 · Last updated July 4, 2026

CFPB Examiners Are Stress-Testing Your AI Underwriting Models regulation briefing

The Leveraged Years AI Regulation Tracker

Regulatory Briefing

Bindingness: Enforcement Shift — supervisory expectations for AI credit models remain active even where guidance was withdrawn.
Scope: Mortgage lenders and brokers, consumer lenders (personal, auto, credit card), hard-money shops with consumer-purpose credit, and compliance officers signing fair-lending memos.
What Changed: CFPB examiners treat opaque ML and third-party scores as a compliance problem, not a technology excuse. January 2025 Supervisory Highlights documented underwriting and pricing disparities, weak CMS, and adverse-action processes not validated against model outputs.
Who This Affects: Any creditor using machine learning, complex algorithms, or vendor scores in credit decisions — solo shops included.
Current Status: Active 2026 supervisory posture; fair-lending testing and adverse-action accuracy are exam priorities.
Effective Date: Circular 2022-03 (May 2022) and Supervisory Highlights special edition (January 17, 2025) define current exam expectations.
Primary Source: CFPB Circular 2022-03 — adverse action and complex algorithms
Secondary Analysis: CFPB Circular 2023-03 (sample adverse-action forms) and Supervisory Highlights: Advanced Technologies in Credit Scoring (January 2025).
Workflow Impact: Black box is not a defense. If you cannot explain principal reasons for an adverse action, you cannot lawfully use the model.
Operational Next Step: Run adverse-action reason testing this quarter — compare denial notices to model principal reason codes before the examiner asks.
Last Verified: July 4, 2026

The honest read

HUD fair-housing screening is a different lane. This is consumer lending under ECOA and Regulation B, and the Bureau is treating opaque models as a compliance problem, not a technology excuse. If you use machine learning or third-party scores in consumer credit, your next exam is likely to ask for model documentation, fair-lending testing, and adverse-action notices that name the real reasons for a denial — not a Regulation B checkbox.

The rule in plain English

Federal fair lending law already applies to every credit decision, including decisions made with AI, machine learning, or complex third-party scores. The CFPB has said repeatedly — in binding circulars and supervisory findings — that black box is not a defense, sample adverse-action checklists are not a shortcut, fair-lending testing must match the model, and complex models raise proxy risk. In January 2025, the Bureau published a Supervisory Highlights special edition focused on institutions using advanced computation in credit scoring. Reported examination findings included underwriting and pricing disparities for Black and Hispanic applicants, weak compliance management systems, and adverse-action processes that had not been validated against model outputs.

Who it hits

Mortgage lenders and brokers using AVM-adjacent pricing or automated underwriting; consumer lenders in personal, auto, and credit-card models; real estate investors and hard-money shops on consumer-purpose credit; financial advisors at broker-dealers with lending affiliates; and CPAs signing the fair-lending testing memo the examiner asks for. Solo shops are not exempt — a thin compliance team makes documentation gaps more visible, not less.

What to do this week

1. Inventory every model that touches a credit decision — internal, vendor, or bureau — including re-scores and pricing tiers. Name the owner and the last validation date. 2. Run adverse-action reason testing this quarter. For a sample of denials, compare the notice the applicant received to the model's principal reason codes. 3. Document fair-lending testing commensurate with risk, including disparate-impact analysis for underwriting and pricing outcomes. 4. Search for less discriminatory alternatives (LDAs). Write down what you considered and what you rejected. 5. Review alternative-data inputs. If a variable is not intuitively related to creditworthiness, your adverse-action disclosure and fair-lending monitoring must still catch it. 6. Align vendor contracts so third-party scores require reason-code access, validation support, and audit cooperation. 7. Brief the board or partners in the next compliance memo: which models are in scope, last fair-lending test date, and open remediation items.

Related tracker rows

[AI tenant screening and Fair Housing disparate impact](/ai-regulation-news/ai-tenant-screening-fair-housing-disparate-impact-2026)
[FINRA agentic AI and deepfake controls](/ai-regulation-news/finra-agentic-ai-deepfake-controls)
[SEC RIA AI governance exam files](/ai-regulation-news/ria-ai-governance-exam-files)
[Federal AVM rule for AI home valuation](/ai-regulation-news/avm-rule-ai-home-valuation-real-estate-2026)
[CFPB fair-lending rollback — what still binds lenders](/ai-regulation-news/cfpb-ai-fair-lending-rollback-what-still-binds-lenders-2026)

Source File

Primary Source: CFPB Circular 2022-03 — adverse action and complex algorithms
Url: https://www.consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms/
Date: May 26, 2022
Primary Source: CFPB Circular 2023-03 — adverse action sample forms (PDF)
Url: https://files.consumerfinance.gov/f/documents/cfpb_adverse_action_notice_circular_2023-09.pdf
Date: September 2023
Primary Source: CFPB Supervisory Highlights — Advanced Technologies in Credit Scoring
Url: https://www.consumerfinance.gov/compliance/supervisory-highlights/
Date: January 17, 2025 (special edition cited in industry compliance summaries; confirm current PDF on catalog)
Statutory Authority: ECOA section 701(d), 15 U.S.C. 1691(d), and Regulation B, 12 CFR 1002.9
Last Checked Against Primary Source: July 4, 2026
Methodology: Researched from primary sources, drafted with AI assistance, fact-checked and edited by the editorial desk. Informational only, not legal, tax, medical, or investment advice.

Frequently Asked Questions

Did the CFPB withdraw its AI lending guidance?

Yes — Circular 2022-03 and 2023-03 were withdrawn May 12, 2025. That does not repeal ECOA or Regulation B. Examiners still expect specific adverse-action reasons and fair-lending testing commensurate with model risk.

What will examiners ask for on AI underwriting models?

Model documentation, fair-lending testing (including disparate impact), adverse-action reason validation against model outputs, and evidence you searched for less discriminatory alternatives where appropriate.

Does this apply to mortgage lenders only?

No. It covers consumer credit broadly — cards, auto, personal loans, and mortgage — wherever ML or complex vendor scores drive underwriting or pricing.