I. Legal Architecture - Why Explainability is Mandatory

1. INDIA - RBI - 2024 to 2026 Framework

a) Digital Lending Guidelines 2025: Technology can change channel through which credit is delivered but does not dilute responsibility of regulated lender. Automation does not transfer responsibility from lender to technology.

b) Draft Guidance on Regulatory Principles for Model Risk Management, 2026 - Released 24 June 2026, Press Release 2026-2027/528 - Comments closed 24 July 2026:

This is India's first AI rulebook for banks/NBFCs. Key obligations:

• Explainability, not black boxes: Models must be explainable, with control boundaries to contain risks of generative AI, hallucinations. Banks must be able to explain model's decision, especially for important things like lending. Where they cannot fully explain it, they must add extra checks.
• No blame the algorithm: Banks and NBFCs may use AI, ML and third-party systems, but cannot transfer accountability to algorithm or vendor. Outsourcing model to vendor doesn't remove regulatory accountability.
• Lenders should assume duty of explanation and ensure outputs are explainable, transparent and fair by knitting ethical AI design to fabric of FinTech. Outcomes shall be consistent, unbiased, explainable and verifiable.
• 5 Immediate Priorities from RBI Governor Sanjay Malhotra: Complete inventories of AI systems, board-approved governance frameworks, explainable AI decisions, regular stress testing and meaningful human oversight + kill-switch - override, suspension, or deactivation mechanisms.

RBI Financial Stability Report June 2025: AI systems now underpin more than 70% of loan approvals.

2. EU - AI Act + GDPR

EU AI Act Regulation 2024/1689 entered into force Aug 2024, high-risk obligations apply from 2 Dec 2027 per Digital Omnibus deferral:

• High-Risk Classification: AI systems used in credit scoring, creditworthiness assessment of natural persons are classified as high-risk under Annex III.
• Obligations (Art 13, 14, 50): Technical documentation, data governance, human oversight, transparency - system must be designed to allow human overseer to correctly interpret system's output before acting on it. Provide clear explanations for credit denials.
• Two levels required:
  • Global Explainability: Which features carry most weight overall
  • Local Explainability: Why specific individual refused - e.g., debt-to-income ratio exceeds threshold
• GDPR Art 22: Right not to be subject to solely automated decision with legal effect + right to meaningful information about logic involved. SHAP/LIME explainer layers required.

Penalty: €35 million or 7% global turnover.

3. USA - ECOA + CFPB - The Oldest Explainability Law

• Equal Credit Opportunity Act 1974 + Reg B: Creditor that denies or limits credit must provide specific principal reasons, and model's complexity does not suspend obligation.
• CFPB Circular 2022-05 and 2023-03: Complex algorithmic models do not excuse ECOA adverse-action requirements. Black-box excuse rejected.
• SR 11-7 replaced by SR 26-2 April 2026 interagency guidance: Generative and agentic AI carved out of model risk scope but NOT carved out of consumer protection law.

II. What Must Be Explained - The Standard

III. Case Laws - Where Banks Failed

Case 1: USA - Massachusetts AG v. Earnest Operations LLC - $2.5 Million Settlement - 10 July 2025 - The Landmark AI Lending Case

Facts: Private student loan company used AI model that included:

• Cohort Default Rate (CDR) - dataset from US Dept of Education showing default rate of applicant's college
• Knockout rule denying non-green-card holders

Investigation: AG found Earnest's use of CDR as weighted subscore penalized HBCU graduates (Historically Black Colleges) regardless of personal creditworthiness. Used CDR which disproportionately hurt approval rates for Black and Hispanic applicants.

Held:

• ECOA Disparate Impact Violation: AI model created disparate impact against Black, Hispanic, noncitizen applicants - zero exceptions for complexity.
• Proxy Discrimination: CDR = proxy for race.
• Transparency Failure: Inaccurate adverse-action notices - gave generic reasons like "education" not specific principal reasons.
• Settlement: $2.5M, model remediated, CDR removed, adverse action notices rewritten with SHAP values.

This is India's warning - RBI cited this case in FREE-AI Committee Report Aug 2025 - Framework for Responsible and Ethical Enablement of AI.

Case 2: CFPB v. Upstart Network - Ongoing Supervision 2022-2024

CFPB found Upstart's AI model using education and employment as variables. Even though model approved more Black applicants overall, CFPB demanded proof of less discriminatory alternative (LDA). Upstart had to produce:

• Global explainability - feature importance
• Local explainability - SHAP values per denied applicant
• Disparate impact testing per variable

Established principle: More approvals overall does not excuse disparate impact on subgroup if proxy variable exists.

Case 3: Apple Card - Goldman Sachs - NYS DFS Investigation 2019-2021

Algorithm gave man 20x credit limit vs wife with shared finances. Goldman claimed "algorithm gender blind".

NY DFS found no intentional discrimination but found explainability failure - could not explain why. Required Goldman to:

• Implement explainability layer
• Retrain with causation, not correlation
• Provide adverse action with specific reasons

Now codified in EU AI Act Art 13.

Case 4: Spain/EU - Tribunal Supremo STS 735/2021 - Bankia KYC Failure + Cl@ve

Supreme Court found bank liable for allowing account opening with low-assurance Cl@ve PIN without second factor, because AI fraud model could not explain why it approved high-risk onboarding. Pushed adoption of eIDAS High LoA + human review.

Case 5: India - RBI Enforcement Actions (not public names, but pattern from LawBeat 2026)

RBI draft message: Banks cannot blame algorithm. In 2025 inspections, RBI found digital lenders rejecting loans in seconds with reason "AI model decision". RBI direction:

• Loan rejected in seconds? RBI wants human able to step in. Banks must have override.
• Algorithms should be able to explain reasoning behind decisions to customers: RBI Murmu statement.
• If black-box (deep learning), must add extra monitoring, limit to low-stakes decisions, or use interpretable model (logistic regression, decision tree) for final decision.

IV. How to Comply - Explainability Stack for Lending

For Bank/NBFC in 2026:

1. Inventory: List all AI models affecting credit - sourcing, scoring, pricing, collections. Board-approved.
2. Classify Risk: Credit scoring = High-Risk under EU AI Act + Material under RBI Model Risk.
3. Choose Interpretable Model: For adverse action, use model where you can generate SHAP/LIME. If deep learning, build explainer layer.
4. Adverse Action Notice - Must contain:
   • 4 principal reasons: e.g., 1. Debt-to-income ratio 52% exceeds 40% threshold, 2. 3 recent 30+ DPD in last 6 months, 3. Insufficient credit history (6 months <12 required), 4. Loan amount requested 150% of annual income
   • Not: "Credit score low", "Model rejected"
5. Disparate Impact Test: Test model on protected groups - race (US), caste/gender (India proxy), age. Remove proxies - zip code, college name, language.
6. Human in Loop + Kill Switch: Human credit officer must be able to view explanation and override. Draft guidance requires override, suspension, deactivation mechanisms.
7. Documentation: Technical doc, data lineage, bias test results, explainer validation.

Failure Consequence:

• India: RBI penalty under Sec 47A BR Act + Digital Lending Guidelines violation + consumer court deficiency + loan declared void for lack of transparency (applying STS 17 April 2018 doctrine - convenience test does not exempt reporting duties)
• EU: Up to €35M + loan contract void for unfair term under Directive 93/13 (Aziz doctrine)
• US: ECOA violation = unlimited punitive damages + CFPB enforcement

Bottom Line: Explainability is not about opening source code. It is about proving why a human would have made same decision with same data, without proxy discrimination.