Executive Summary
Respondents agree that ECOA/Regulation B requires specific, principal reasons for adverse actions, but most see gaps when these obligations meet AI-driven underwriting. A clear set of technical approaches exists to extract applicant-level reasons from complex models, yet many commenters call for additional guidance to translate model explanations into compliant, consumer-meaningful notices. With 61.84% responding “No” on sufficiency of clarity for Question 15, the dominant view is that Regulation B needs AI-specific direction. A minority maintains that Regulation B is already clear and flexible, provided reasons are specific and accurate.
Key takeaways:
- 61.84% No vs. 38.16% Yes on whether Regulation B provides sufficient clarity for AI adverse action reasons (coverage: 100%).
- Explainability methods (e.g., Shapley-based techniques) can identify principal reasons from AI models.
- Institutions must justify decisions at a local/individual level to generate reason codes; the outcome definition can alter the reason code.
- It is technically possible to attribute factors that contributed to an AI decision.
- Surrogate models can be used to assess and communicate reasons from black-box algorithms.
- Documentation describing how adverse action notices are created is emphasized for compliance.
- Some assert Regulation B is already clear; specificity and accuracy are required, and no further clarity is needed.
- Others stress that legacy techniques won’t work for AI/ML and that updated guidance is needed to produce borrower-specific reasons.
Bottom line:
Approaches to identify AI-driven adverse action reasons include Shapley-based explainability, surrogate models, individual-level reason coding, and robust documentation. Most respondents find Regulation B’s current guidance insufficient for AI contexts and seek detailed, AI-specific expectations, though a minority views existing standards as clear and flexible.
The Question (Ref #15)
The Equal Credit Opportunity Act (ECOA), which is implemented by Regulation B, requires creditors to notify an applicant of the principal reasons for taking adverse action for credit or to provide an applicant a disclosure of the right to request those reasons. What approaches can be used to identify the reasons for taking adverse action on a credit application, when AI is employed? Does Regulation B provide sufficient clarity for the statement of reasons for adverse action when AI is used? If not, please describe in detail any opportunities for clarity.
Direct Response to the Catalog Question
Use model explainability techniques—such as Shapley-based methods—to extract principal, applicant-level factors from AI decisions for adverse action reasons.
Leverage surrogate/explainable models and local explanations to translate black-box outputs into specific, actionable reason codes; note that the exact model outcome being explained can change the resulting code.
Establish governance and documentation describing the mechanism for generating adverse action notices, ensuring transparency, specificity, and consumer-understandable language.
Most respondents say Regulation B lacks sufficient AI-specific clarity (61.84% No), citing uncertainty around mapping complex model outputs to compliant reason statements and requesting updated guidance.
Some respondents argue Regulation B already provides clear, flexible standards requiring specificity and accuracy, and that creditors need not describe how or why a factor adversely affected the application.
Opportunities for clarity include: acceptable AI/XAI methods, examples of compliant reason-code mappings from attributions, expectations for reason granularity and model-outcome definitions, and treatment of non-explainable models.
By-the-numbers — Question 15
| Metric | Value |
|---|---|
| Total Yes | 29.0 |
| Total No | 47.0 |
| Total (Yes+No) | 76.0 |
| % Yes | 38.16% |
| % No | 61.84% |
| % of answers (coverage) | 100.0% |
Introduction
Question 15 asks how creditors can identify principal reasons for adverse action when AI is used, and whether Regulation B provides sufficient clarity for those reasons. The materials consistently reference ECOA/Regulation B’s requirement to notify applicants of specific reasons, while probing how that obligation functions with modern AI/ML underwriting.
Historic Lessons in the Evidence
Respondents’ reasoning reflects that unexplainable or black-box systems frustrate ECOA/Reg B obligations, driving a shift toward per-decision explainability and documentation. They also note that traditional, legacy techniques are inadequate for AI/ML, prompting the adoption of game-theoretic attributions, surrogates, and governance processes to produce specific, accurate, and consumer-meaningful reasons.
Recent Developments
Not observed in the provided materials.
The Challenge
Practically, lenders must bridge technical attributions to consumer-friendly reason codes while managing complex, less transparent models. Commenters flag that neural networks may not yield clear reasons, that the outcome definition can change reason codes, and that uncertainty persists around how to align AI explanations with Regulation B’s requirement for specific, principal reasons without needing to explain how/why a factor affected the decision.
Evolving Metrics
Respondents justify feasibility by pointing to local/individual attributions (e.g., Shapley-based contributions) and the technical ability to attribute factors to decisions. They emphasize specificity and accuracy for reason codes, documentation of the mechanism for creating notices, and, in some cases, multiple tests to ensure reliable, compliant explanations. Several highlight that reason-code validity can hinge on the precise outcome being explained.
A Framework Inspired by the Inputs
An implicit approach emerges: pair AI models with explainability (local attributions or surrogates), map the top contributing factors to standardized reason codes, and maintain end-to-end documentation of how notices are generated. This is anchored by Regulation B’s flexible standard for specific reasons, tempered by calls for AI-specific guidance to ensure borrower-level clarity and compliance.
Case Study
A representative pattern shows a lender using ML underwriting, applying Shapley-based explanations to identify the top drivers for each declined applicant, then mapping those drivers to standardized reason codes. Where models are opaque, a surrogate model produces comparable explanations. The lender documents the mechanism for generating notices and validates that explanations remain specific and accurate, while acknowledging uncertainty about how best to define outcomes and translate technical attributions into compliant, consumer-meaningful reasons.
Recommendations
- Adopt Shapley-based, local explainability to surface applicant-level principal reasons from AI models.
- Use surrogate or inherently interpretable models when primary models are too opaque to yield compliant reason codes.
- Define and standardize the exact model outcome to be explained, then validate that resulting reason codes are stable and appropriate.
- Map attributions to clear, standardized reason codes and ensure consumer-understandable phrasing that satisfies specificity and accuracy.
- Document the full mechanism for generating adverse action notices, including data, model, explanation method, and reason-code mapping.
- Perform periodic testing to confirm explanations remain consistent, accurate, and aligned with ECOA/Reg B obligations.
- Seek or support AI-specific regulatory guidance clarifying acceptable methods, examples of compliant reason statements, and expectations for black-box models.
- Leverage existing flexibility in Regulation B while avoiding unexplainable models that cannot yield borrower-specific reasons.
Conclusion
Effective approaches exist to identify principal reasons for adverse action in AI underwriting—particularly Shapley-based explanations, surrogates, local reason codes, and rigorous documentation. However, most respondents find Regulation B’s current guidance insufficient for AI contexts and request specific clarifications to translate technical attributions into compliant, consumer-meaningful reasons. A minority maintains that Regulation B is already clear and flexible, provided reasons are specific and accurate. Aligning explainability practices with explicit regulatory expectations will close the gap between AI model outputs and ECOA-compliant adverse action notices.
This analysis will continue in our next publication. Don’t miss the next installment.
Follow us, stay informed, stay secure, and let’s navigate the risk landscape together.
