Regulation — banking-specific
CFPB Circular 2022-03 — AI-driven adverse-action notices must give specific reasons
The Circular that answered "can we just say 'model score' as the reason?" with a clear no. When an AI credit-underwriting engine drives an adverse action, the ECOA / Reg B specific-reason duty still applies — the blackbox is not the bank's shield.
What the Circular actually says
"Federal consumer financial protection law requires creditors to provide statements of specific reasons to applicants against whom adverse action is taken. … Creditors cannot justify noncompliance with ECOA's requirement to provide the specific and accurate reasons for adverse actions by using complex algorithms whose reasoning may be difficult to explain."
CFPB Circular 2022-03 (May 2022) — specific-reason requirement applied to complex models
"A creditor's use of complex algorithms … does not absolve the creditor from its obligation to provide accurate and specific reasons for adverse actions. Whether a creditor is using a decision-making process that is too complex, opaque, or new to allow the creditor to accurately identify the specific reasons is not a defense to noncompliance."
CFPB Circular 2022-03 — blackbox is not a defense
What this means in plain English
The bank cannot use its AI vendor's opacity as a legal shield. Two rules apply:
- Specific reasons. The adverse-action notice must state the specific reasons that drove the decision — not generic language like "credit score below threshold" or "model output."
- Accuracy. The reasons stated must actually be the drivers of the decision. Post-hoc rationalizations do not satisfy the rule.
If the bank cannot explain the AI's decisions specifically and accurately, the CFPB position is that the bank should not be using that AI for adverse-action decisions.
What triggers the exposure in the sample
Silent B3 drift + credit-underwriter-v6.1.4 shipped without explainability discipline for the drifted borrower group. Every B3 decline / price-up in the drift window was accompanied by an adverse-action notice. If those notices cited a generic AI-output reason — and the underlying driver was silent applicant-mix drift the model was not calibrated for — the specific-reason requirement is not met.
What the $499 Snapshot shows against this rule
- Distributional-shape analysis that names the drift-driver — the specific-reason input the compliance team can trace
- Model-version pin per decision — ties the adverse-action notice to the exact model output
- 3 fix-first items including explainability-per-group discipline — the remediation stance
- Independent-verifier declaration — the diligence record CFPB looks for on file
See the baseline-vs-drifted chart the drift-driver falls out of →
$499 Snapshot. 3 business days.
Named-driver evidence for your bank's AI credit-underwriting decisions — the specific-reason record the CFPB is trained to look for.
Buy $499
Snapshot credit applies to Baseline ($2,500) or Enterprise Attestation ($35-55K) upgrade within 30 days.