Adverse Action Notices in AI Lending: Massachusetts Rules After the Earnest Settlement

On July 10, 2025, Massachusetts Attorney General Andrea Joy Campbell filed an Assurance of Discontinuance in Suffolk County Superior Court requiring student-loan lender Earnest Operations LLC to pay $2.5 million and overhaul its algorithmic underwriting (Mass.gov, Jul 10, 2025). State investigators alleged that Earnest trained automated models on arbitrary human decisions without disparate-impact testing, scored a school-level Cohort Default Rate variable penalizing Black and Hispanic applicants in approvals and pricing, ran an automated immigration "Knockout Rule" barring applicants without a green card, and sent rejected borrowers inaccurate adverse-action notices masking true denial triggers. Most telling was the date: Massachusetts filed two months after the federal Consumer Financial Protection Bureau (CFPB) withdrew 67 guidance documents on May 12, 2025 (Federal Register, 90 FR 20084).

Assuming Washington's pullback shields automated underwriting in New England is an expensive mistake. Alongside the $2.5 million Earnest resolution, a Massachusetts federal class action over automated tenant-screening scores settled for $2.275 million (Cohen Milstein). Private claims under M.G.L. c. 93A carry up to treble damages plus fees (MA Legislature c. 93A), while Equal Credit Opportunity Act (ECOA) class actions expose creditors to punitive damages up to the lesser of $500,000 or a 1% net-worth threshold under 15 U.S.C. s. 1691e(b). Part of our safe use of AI tools at work hub and paired with our late-2026 state vs federal AI law scoreboard, this guide serves loan officers, credit union staff, fintech compliance leads, and Massachusetts borrowers. It provides general operational information, not legal advice.

What did the $2.5 million Earnest settlement actually find inside the underwriting pipeline?

The Suffolk Superior Court Assurance of Discontinuance against Earnest Operations LLC targeted four interlocking failures across model training, proxy variables, hard eligibility filters, and borrower disclosures (Mass.gov, Jul 10, 2025).

Think of a clinic triage nurse checking a patient for fever. Measuring the patient's own temperature tests individual risk; turning the patient away because the average temperature at their college was elevated three winters ago substitutes group history for individual vital signs, and writing "General wellness score" on the discharge slip hides the error. Where that physical analogy breaks is multi-variable interaction in machine-learning credit models. A thermometer reads one physical dial, whereas gradient-boosted trees or neural credit scorers blend dozens of correlated cash-flow, education, and bureau features whose combined weights must map accurately onto four plain-English denial lines.

The Massachusetts Attorney General's Office (AGO) alleged four unlawful practices under M.G.L. c. 93A, M.G.L. c. 151B, and ECOA:

1. Untested models trained on arbitrary human decisions: Earnest built underwriting models atop historical human choices without testing whether those labels encoded subjective bias or checking live outputs for disparate impact.
2. The Cohort Default Rate (CDR) proxy penalty: Earnest scored the federal institution-level Cohort Default Rate (the share of a school's borrowers defaulting within a tracking window). Because Black and Hispanic students disproportionately attend schools with higher aggregate CDR figures due to structural wealth gaps, scoring school-wide CDR penalized creditworthy Black and Hispanic applicants in approvals and interest rates.
3. The automated immigration "Knockout Rule": An automated filter summarily rejected applicants lacking permanent-resident green-card status before weighing individual creditworthiness, triggering national-origin and alienage scrutiny.
4. Inaccurate adverse-action reason codes: Rejected borrowers received generic or mismatched denial reasons that omitted the true variables (CDR or policy knockouts) driving the adverse decision.

Earnest paid $2.5 million, retired both the Cohort Default Rate variable and the immigration Knockout Rule, adopted written AI governance, instituted annual fair-lending disparate-impact testing, and accepted ongoing AGO compliance reporting.

Did the CFPB's 2025 guidance withdrawal and 2026 Regulation B rule erase AI lending rules in Massachusetts?

No: while the CFPB rescinded 67 informal guidance documents in May 2025 and stripped disparate-impact commentary from federal Regulation B effective July 21, 2026, specific adverse-action accuracy remains binding federal law nationwide and disparate-impact credit liability remains enforceable under Massachusetts statutes.

Separating sub-regulatory circulars, federal rules, and Massachusetts statutes clarifies what changed and what stayed intact:

• May 12, 2025 (90 FR 20084 guidance withdrawal): The CFPB rescinded 67 guidance items, including Circular 2022-03 (black-box complexity does not excuse ECOA specific-reason duties) and Circular 2023-03 (warning against checking approximate Regulation B sample form boxes for alternative data) (Federal Register; JD Supra). Rescinding a circular removes an agency explainer; it cannot repeal 15 U.S.C. s. 1691(d) or 12 CFR s. 1002.9(b)(2), which still require specific, accurate principal reasons on every adverse credit action.
• July 21, 2026 (CFPB Regulation B Final Rule): Following a November 13, 2025 proposal and April 22, 2026 Final Rule, the CFPB removed "effects test" (disparate-impact) text from federal Regulation B (12 CFR Part 1002), sparking federal Administrative Procedure Act suits by the National Consumer Law Center and civil-rights groups.
• Why Massachusetts enforcement stayed firm: Attorney General Campbell's April 16, 2024 AI Advisory confirmed that M.G.L. c. 151B, s. 4(14) (barring credit discrimination) and M.G.L. c. 93A, s. 2 (barring unfair or deceptive practices, including false AI notices and discriminatory models) govern AI systems (Orrick, Apr 2024), reinforced by Earnest ($2.5M) and Massachusetts tenant screening litigation after SafeRent.

Why do AI reason-code generators produce illegal adverse-action notices?

Under 12 CFR s. 1002.9(b)(2) and Official Interpretations (Supplement I, 1002.9(b)(2)), adverse-action notices must be specific and indicate the principal reason(s) for the credit decision; listing more than four reasons is unhelpful, and checking Appendix C Form C-1 sample boxes is unlawful if they misstate the true factors scored (eCFR 12 CFR 1002.9).

Vendor platforms break 1002.9(b)(2) via three shortcuts: (1) coarse bucket mapping, such as mapping late-night open-banking debit timing onto "Delinquent credit obligations" when the borrower has zero late payments (deceptive under c. 93A), (2) uncalibrated SHAP wrappers whose shifting baselines hand identical borrowers contradictory reasons, and (3) hidden upstream knockouts where a policy filter rejects the file before scoring while the letter prints bureau factors anyway.

How does a credit union or community lender audit reason codes on 50 declined files?

Small Massachusetts credit unions, community banks, and CDFIs can catch broken reason codes and proxy variables without a quant team by running a four-hour, 50-file pre-deployment shadow audit with one compliance officer and one senior underwriter. (See parallel oversight rules in how claims adjusters document human judgment alongside AI tools in Massachusetts.)

Worked 50-decline shadow audit (assumptions labelled): A 35-staff Worcester credit union tests a vendor AI personal-loan module ($18,000/year SaaS fee, 220 applications/month, 25% baseline decline rate = ~55 declines/month):

1. Pull 50 shadow-declined files (30 min). Export raw feature-contribution logs alongside draft adverse-action letters.
2. Check feature-to-letter match (2.5 hours, ~3 min/file): Confirm the top 1 to 4 negative model drivers match the printed lines. If File #14 lost 42 points on "Irregular deposit cadence" while the letter printed "Length of time accounts established", freeze launch until fixed.
3. Verify knockouts and disparity ratios (1 hour): Inspect every upstream filter (zero CDR, ZIP, or blanket visa knockouts) and review annual M.G.L. c. 151B adverse-impact ratios. Four hours (~$300 staff time at $75/hour) protects 660 declined Massachusetts members annually.

What can a rejected borrower in Massachusetts ask for within 60 days?

Rejected borrowers in Massachusetts hold four enforceable rights on strict statutory clocks:

1. Specific reasons within 30 days (12 CFR s. 1002.9(a)(1)(i)): Creditors must send adverse-action notice within 30 days of a completed application, listing specific principal reasons or a 60-day right to request them (15 U.S.C. s. 1691(d)(2)(B)).
2. Score and free FCRA report (15 U.S.C. s. 1681m(a) & 1681j): Any bureau or cash-flow aggregator used must be named with score, factors, and a 60-day free report right to dispute errors.
3. Uniform second-look human review: Credit unions offering manual reconsideration for union strikes, medical leave, or payroll errors must apply uniform written standards to prevent disparate treatment.
4. State enforcement complaints: Borrowers receiving deceptive or discriminatory AI notices can file with the Massachusetts Attorney General's Office and Division of Banks. (Community counselors can also connect via our ed.cafe learning circle start page.)

The ed.cafe Credit & Fintech AI Adverse-Action Reason-Code Audit & Applicant Rights Guide

Use the ed.cafe Credit & Fintech AI Adverse-Action Reason-Code Audit & Applicant Rights Guide at lending committees, member-service desks, and financial-coaching clinics. The ed.cafe Audit Guide pairs an eight-point creditor check with a Massachusetts borrower request template.

THE ED.CAFE CREDIT & FINTECH AI ADVERSE-ACTION AUDIT & APPLICANT RIGHTS GUIDE
Publisher: ed.cafe by No Worker Left Behind (Boston 501(c)(3), EIN 88-3268312)
Refs: 15 U.S.C. 1691(d); 12 CFR 1002.9; M.G.L. c. 151B & c. 93A; MA AG Earnest
Settlement (Jul 10, 2025). Operational reference; not legal advice.

PART 1: 8-POINT LENDER & CREDIT UNION AI ADVERSE-ACTION AUDIT
[ ] 1. BAN GROUP PROXIES: Confirm zero school Cohort Default Rate (CDR), college
    tier, or 3/5-digit ZIP variables in approval or pricing models.
[ ] 2. AUDIT HARD KNOCKOUTS: Inspect pre-model filters; strip blanket immigration
    knockouts untethered to valid work/contract status.
[ ] 3. RUN 50-DECLINE REASON CHECK: Verify top 1 to 4 model feature drivers on
    50 shadow-declined files match the exact lines printed on member notices.
[ ] 4. NO APPROXIMATE FORM C-1 BOXES (12 CFR 1002.9(b)(2)): Never map cash-flow
    or alternative variables onto unrelated Appendix C checkboxes.
[ ] 5. CAP AT 1 TO 4 PRINCIPAL REASONS: Suppress 8-reason dumps so applicants see
    the true primary drivers they can verify or fix.
[ ] 6. FCRA DISCLOSURES (15 U.S.C. 1681m(a)): List every credit or cash-flow
    bureau used, numerical score, key factors, and 60-day free report steps.
[ ] 7. KEEP ANNUAL MA DISPARATE-IMPACT TESTING: Despite the Jul 21, 2026 CFPB Reg
    B rule, log annual M.G.L. c. 151B / c. 93A disparity and LDA checks.
[ ] 8. LOG SECOND-LOOK OVERRIDES: Apply written manual-review criteria uniformly.

PART 2: COPY-READY MASSACHUSETTS BORROWER SPECIFIC-REASONS REQUEST
(Send within 60 days of credit denial or unfavorable rate notice)
To: [Creditor Adverse Action Compliance Dept] | Date: [Date] | App #: [Number]
From: [Applicant Name & Massachusetts Address]

On [Date], I received an adverse-action notice on my credit application above.
1. SPECIFIC REASONS: Under 15 U.S.C. 1691(d)(2), 12 CFR 1002.9(b)(2), and M.G.L.
   c. 93A / c. 151B, please provide in writing within 30 days the specific,
   accurate principal reasons that actually determined this decision, including
   any automated underwriting or non-bureau variables rather than generic labels.
2. REPORTING SOURCES: Under 15 U.S.C. 1681m(a), please confirm the name, address,
   and phone of every bureau or cash-flow aggregator used in this decision.
3. HUMAN RECONSIDERATION: Please state how I may submit verified income or
   corrected source documents for manual underwriting review, if offered.
Signed: ___________________________

Where this stops being true

1. Commercial credit rules: Business credit above and below $1 million in gross revenue follows modified timing rules under 12 CFR s. 1002.9(a)(3), though anti-discrimination bans still apply.
2. Active federal APA litigation: Consumer groups (NCLC, NFHA) are challenging the CFPB's July 21, 2026 Regulation B disparate-impact rule in federal court; a stay or vacatur could restore federal disparate-impact text nationwide.
3. Multistate AI statutes: Lenders operating in Colorado face SB 26-189 AI notice and human-review duties effective January 1, 2027 where not federally preempted.
4. Not legal advice: Published by ed.cafe at No Worker Left Behind for workforce literacy; consult fair-lending counsel before changing Regulation B notices or BISG models.

Questions people ask

Key terms

• Adverse action notice (12 CFR s. 1002.9): Written notice within 30 days of a credit denial or worse rate stating specific principal reasons and ECOA rights.
• Assurance of Discontinuance (AOD): A court-filed M.G.L. c. 93A, s. 5 settlement, like the July 10, 2025 $2.5 million Earnest filing in Suffolk Superior Court.
• Cohort Default Rate (CDR): A school-level student default metric barred in the Earnest settlement after disproportionately penalizing Black and Hispanic applicants.
• Knockout Rule: A hard automated pre-screening filter rejecting credit applications before scoring, like Earnest's contested immigration filter.
• Post-hoc explainability (SHAP): Mathematical approximations estimating which input features lowered a machine-learning credit score on a specific file.
• Regulation B (12 CFR Part 1002): The federal rule implementing ECOA anti-discrimination and 30-day adverse-action notice duties.

Sources

1. Massachusetts Attorney General, "AG Campbell Announces $2.5M Settlement With Earnest Operations LLC" (Jul 10, 2025): Mass.gov
2. Orrick, "Massachusetts Attorney General Shares AI Guidance (Apr 16, 2024 Advisory)" (Apr 2024): Orrick
3. Federal Register / CFPB, "Withdrawal of 67 Guidance Documents (90 FR 20084)" (May 12, 2025): Federal Register
4. CFPB, "Circular 2022-03 & Circular 2023-03 (Historical Archive)": Circular 2022-03 and Circular 2023-03
5. eCFR, "12 CFR s. 1002.9: Regulation B Notifications": eCFR 12 CFR 1002.9
6. JD Supra, "CFS Bites of the Month: 2025 Annual Review (CFPB Withdrawals & Reg B)" (Jan 2026): JD Supra
7. Massachusetts General Court, "M.G.L. c. 151B, s. 4 & M.G.L. c. 93A, s. 2": c. 151B, s. 4 and c. 93A, s. 2
8. Cohen Milstein, "SafeRent Solutions $2.275M AI Discrimination Settlement (D. Mass.)": Cohen Milstein