Reference · Fair lending

ECOA and Regulation B: Adverse Action Notices for AI Credit Decisions

The specific-reasons rule, the 2025 and 2026 changes around it, and the records an examiner asks for

Last reviewed November 4, 2026

At a glance

Statute: ECOA section 701(d), 15 U.S.C. 1691(d)
Regulation: Regulation B, 12 CFR 1002.9 (notices) and 1002.12 (record retention)
AI-specific rule: None. Circulars 2022-03 and 2023-03 were withdrawn on 12 May 2025.
Latest amendment: 91 FR 21620, effective 21 July 2026. Removes disparate impact; leaves notices and retention unchanged.
Official text: 12 CFR 1002.9 (eCFR)

A creditor that denies credit, or takes other adverse action, owes the applicant a statement of the specific reasons. The requirement sits in section 701(d) of the Equal Credit Opportunity Act and in Regulation B, and it applies to every credit decision whatever system produced it: a scorecard, a machine learning model or an underwriter.15 U.S.C. 1691 · (d)(2), (d)(3)12 CFR 1002.9 · (a)(2), (b)(2)

The guidance around that requirement changed twice. On 12 May 2025 the CFPB withdrew its two adverse action circulars, including the one on complex algorithms, and on 21 July 2026 a Regulation B amendment took effect providing that the Act does not authorize disparate-impact liability.90 FR 20084 91 FR 21620 Neither amended 12 CFR 1002.9, and in July 2025 Massachusetts enforced it against an AI lender under the state's consumer protection law.Mass. AG release

Regulation B requires the specific principal reasons, whatever made the decision

Within 30 days of receiving a completed application, a creditor notifies the applicant of approval, a counteroffer or adverse action.12 CFR 1002.9 · (a)(1) A notice of adverse action is in writing and carries either a statement of specific reasons or a disclosure of the applicant's right to one, if requested within 60 days.12 CFR 1002.9 · (a)(2) The statute is direct: a statement of reasons meets its requirements "only if it contains the specific reasons for the adverse action taken."15 U.S.C. 1691 · (d)(3)

The reasons must be specific and indicate the principal reasons for the action. A statement that the applicant fell short of the creditor's internal standards, or failed to reach a qualifying score on its scoring system, is insufficient.12 CFR 1002.9 · (b)(2)

The official commentary ties each reason to the decision process. The reasons must relate to and accurately describe the factors actually considered or scored, and no factor that was a principal reason may be left off, even when its link to creditworthiness is unclear to the applicant.Reg. B commentary · 9(b)(2)-2, 9(b)(2)-4 When scoring and judgment are combined, the reasons come from the component the applicant failed.Reg. B commentary · 9(b)(2)-7 Regulation B mandates no method for selecting reasons from a scoring system; it describes two acceptable methods and accepts any other that produces substantially similar results.Reg. B commentary · 9(b)(2)-5 It sets no number of reasons, and says more than four is not likely to help the applicant.Reg. B commentary · 9(b)(2)-1

For a model with hundreds of inputs, comment 9(b)(2)-2 is the test that decides the notice: each reason must describe a factor the model actually considered or scored for that decision.

The CFPB withdrew its two AI adverse action circulars on 12 May 2025

Circular 2022-03, issued on 26 May 2022, answered whether creditors whose complex algorithms prevent them from identifying the specific reasons still have to give them. The answer was yes, because the notice requirements "apply equally to all credit decisions, regardless of the technology used to make them."Circular 2022-03 Circular 2023-03, released on 19 September 2023, said a creditor using AI or complex credit models may not rely on the checklist of reasons in the sample forms when those reasons do not specifically and accurately indicate the principal reasons.Circular 2023-03

Both were among 67 guidance documents the CFPB withdrew in a notice published on 12 May 2025.90 FR 20084 · Section III CFPB withdrawn guidance The Bureau said guidance is generally non-binding and does not create substantive rights, that the withdrawal is not necessarily final, and that it will deprioritize enforcement against parties whose conduct does not conform to the withdrawn guidance while its review continues.90 FR 20084 · Sections I, II

The notice withdrew interpretations of the law. Section 701(d) of the Act and 12 CFR 1002.9, which the circulars applied, were left as they were, and the commentary quoted above is part of the regulation's official interpretations.12 CFR 1002.9 Reg. B commentary

Regulation B no longer recognizes disparate impact from 21 July 2026

The CFPB's final rule, published on 22 April 2026, adopted its November 2025 proposal as proposed and took effect on 21 July 2026.91 FR 21620 · p. 21620 It amends three areas of Regulation B and leaves 1002.9 and 1002.12 unchanged.91 FR 21620 · pp. 21668-21670

Adverse action records are kept as before: the application, the notice of action taken and the statement of specific reasons for 25 months after the creditor notifies the applicant, generally 12 months for business credit, and beyond that once the creditor has actual notice of an investigation or enforcement proceeding.12 CFR 1002.12 · (b)(1), (b)(4) The rule itself notes that creditors remain subject to other antidiscrimination laws, such as the Fair Housing Act and state laws similar to ECOA.91 FR 21620 · p. 21661

Effects test. Section 1002.6(a) now says the Act "does not provide that the 'effects test' applies." The comment that explained disparate impact is replaced by one on disparate treatment, which reaches facially neutral criteria only where they function as proxies for protected characteristics, designed or applied with the intent to advantage or disadvantage.91 FR 21620 · 1002.6(a), comment 6(a)-2
Discouragement. A statement is prohibited when the creditor knows or should know it would cause a reasonable person to believe the creditor would deny the application, or grant it on less favorable terms, because of a prohibited basis characteristic.91 FR 21620 · 1002.4(b)
Special purpose credit programs. A for-profit program may not use race, color, national origin or sex as an eligibility characteristic, and may use another otherwise prohibited basis only with evidence, for each participant, that the participant would not receive the credit without the program because of that characteristic.91 FR 21620 · 1002.8(b)(3), (b)(4)

Before and after

Regulation B before and after 21 July 2026

UnchangedBefore 21 July 2026From 21 July 2026

Effects test (disparate impact): Recognized in comment 6(a)-2; →; Not provided for by the Act
Discouragement (1002.4(b)): Would discourage a reasonable person; →; Knows or should know; deny or worse terms
For-profit special purpose credit: Shared characteristics under a written plan; →; No race, color, national origin or sex
Adverse action notice (1002.9): Unchanged; →; Specific principal reasons
Record retention (1002.12): Unchanged; →; 25 months; 12 for business credit

From the final rule at 91 FR 21620, its discussion of the prior text and its amendatory text, and 12 CFR 1002.12.

States enforce adverse action accuracy against AI lenders under their own laws

On 10 July 2025 the Massachusetts Attorney General announced a $2.5 million settlement with Earnest Operations LLC, a student loan lender that uses algorithmic models to decide eligibility, loan terms and pricing.Mass. AG release The assurance of discontinuance, filed in Suffolk County Superior Court, alleged violations of the state's consumer protection statute, Chapter 93A, including inaccurate and non-specific adverse action notices in violation of ECOA and Regulation B. Earnest denied the allegations.Earnest AOD · paras. 5, 7

The notice allegations describe reasons that did not match the model. Employees chose decline reasons from a hard-coded menu that left out educational variables the models used as weighted inputs, and picked another available reason when the actual one was missing. When a model deployed to generate decline reasons failed to produce one, an employee had to work out why the model had declined the application without insight into it.Earnest AOD · para. 52

The settlement requires a written governance system of fair lending testing, internal controls and risk assessments for AI models; annual fair lending testing of the underwriting models; an annual inventory of each model's training data, parameters, dates in use and test results; documentation of every algorithmic underwriting decision, kept for four years; and adverse action safeguards that include interpretable models and validation of the accuracy of each reason on every notice.Earnest AOD · paras. 67, 75, 78

Massachusetts reached Regulation B through Chapter 93A. That route runs through the state's own statute and its attorney general, whatever the CFPB's guidance or enforcement priorities.

FCRA adds its own notice when a consumer report is used

When adverse action rests in whole or in part on information in a consumer report, section 615(a) of the Fair Credit Reporting Act requires a notice of the action; any numerical credit score used, with its key factors; the reporting agency's name and contact details; a statement that the agency did not make the decision and cannot give the specific reasons; and notice of the consumer's rights to a free copy of the report and to dispute it.15 U.S.C. 1681m · (a)

A model that draws on credit bureau data brings both notices into the same decision: the ECOA reasons from the creditor, and the FCRA disclosures about the report.

Examiners and attorneys general ask for the trace behind each reason

The rules above, and the terms Massachusetts imposed, point to the same set of records:

The notices sent and the reasons on each, kept for the Regulation B period.12 CFR 1002.12 · (b)(1)
For each decision, the model version, the inputs it scored and its output, so each reason can be traced to the factors actually considered or scored.Reg. B commentary · 9(b)(2)-2
The method used to select reasons from the model's output, and the evidence that it produces accurate reasons.Reg. B commentary · 9(b)(2)-5 Earnest AOD · para. 78
Fair lending testing results for each model and each automatic decline rule, with the dates each was in use.Earnest AOD · paras. 75, 76
Records of overrides and human review where judgment changed the outcome, since the reasons then come from the judgmental component.Reg. B commentary · 9(b)(2)-7
Records of any AI system that drafts or explains a notice, such as a language model that turns reason codes into letter text or answers an applicant's request for reasons: the prompt, the output and the model version, so the text the applicant received can be checked against the reasons the decision produced.

Control mapping

What a reviewer expects to be able to see.

Obligation	What the system must do	Evidence a reviewer expects
Specific reasons	State the principal reasons for each adverse action, specific to the applicant	A copy of each notice with the reasons given
Reason accuracy	Disclose reasons that describe the factors the model actually considered or scored	The reason-selection method, its validation, and the trace from model output to reasons
Timing	Notify within 30 days of a completed application	Application and notice dates for each decision
Model versions	Know which model version made each decision	Model inventory with versions and the dates each was in use
Fair lending testing	Test models and automatic decline rules on a schedule and after material changes	Testing reports and results for each model and rule
FCRA notice	Add the FCRA disclosures when a consumer report was used	Score, key factors and reporting agency details on each notice
Retention	Keep applications, notices and reasons for 25 months (12 for business credit), longer under investigation	Retention schedule and a hold on notice of an investigation
AI-drafted notices	Keep a record of AI systems that draft or explain a notice	Prompts, outputs and model versions tied to the decision

Key dates

Timeline, drawn to scale

The guidance around adverse action notices since 2022

Issued or announcedThe two federal changes

Dates from the CFPB withdrawn guidance list, 90 FR 20084, the Massachusetts Attorney General and 91 FR 21620. Status as of November 4, 2026.

26 May 2022CFPB issues Circular 2022-03 on adverse action notices for credit decisions based on complex algorithms.
19 September 2023CFPB releases Circular 2023-03 on reasons from the Regulation B sample forms in AI and complex model decisions.
12 May 2025CFPB withdraws 67 guidance documents, both adverse action circulars among them (90 FR 20084).
10 July 2025Massachusetts Attorney General announces a $2.5 million settlement with Earnest Operations over AI underwriting and adverse action notices.
13 November 2025CFPB proposes the Regulation B amendments (90 FR 50901).
22 April 2026CFPB publishes the final rule (91 FR 21620).
21 July 2026The final rule takes effect: no effects test, a narrower discouragement standard and new conditions on special purpose credit programs.

Sources cited

Common gaps

Where lenders using AI in credit decisions most often come up short on the rules above.

Reasons drawn from a fixed menu. A reason list that does not cover every input the model uses forces staff to pick the nearest available reason. Massachusetts alleged that practice produced inaccurate notices.Earnest AOD · para. 52
Withdrawn circulars read as a withdrawn requirement. The circulars applied 1002.9, and 1002.9 is still in force. A policy that retired its specific-reasons controls along with the circulars is out of step with the regulation.
Disparate impact testing dropped without a state law review. The effects test is gone from Regulation B. The final rule itself notes that state laws similar to ECOA still apply, and the Massachusetts settlement required annual fair lending testing of the lender's models.91 FR 21620 · p. 21661 Earnest AOD · para. 75
No model version recorded per decision. Models are retrained and replaced. Without the version behind each decision, a reason cannot be traced to the model that produced it.
Generated letter text never checked against the reasons. When a language model writes or explains the notice, its text can drift from the reasons the decision produced. Keeping the prompt and the output lets the two be compared.

In practice

How the Meilynx audit trail records the AI calls that draft or explain a decision →

Last reviewed November 4, 2026. This reference summarises publicly available regulatory guidance and is provided for general information. It is not legal advice. Obligations depend on an institution's charter, registration status, size, and activities. Verify against the primary sources cited above and consult counsel before relying on any summary here.