1. Why Ethical AI is Critical in Finance?

AI in Banking = Credit Scoring, Fraud Detection, Robo-Advisory, AML, Chatbots, Algorithmic Trading, Insurance Underwriting.

Risks: Bias, Discrimination, Black Box, Data Privacy, Financial Exclusion, Systemic Risk.

Regulators say: If you cannot explain AI decision, you cannot use it for lending.

2. Global Regulatory Framework

A. EU AI Act - Regulation 2024/1689 - Effective 1 Aug 2024 - First AI Law in World

• Financial AI = High-Risk AI - Article 6 + Annex III(5)(b) and (c): AI systems used to evaluate creditworthiness and credit scoring, risk assessment for insurance are High-Risk.
• Obligations for High-Risk AI [Article 8-25]:
  1. Risk Management System
  2. Data Governance - Training data must be representative, free from bias
  3. Technical Documentation
  4. Record Keeping / Logging
  5. Transparency - Must inform customer they are interacting with AI and logic involved
  6. Human Oversight - Human must be able to override
  7. Accuracy, Robustness, Cybersecurity
• Prohibited AI: Social scoring for loan, exploitative AI targeting vulnerable.
• Penalty: Up to €35M or 7% global turnover.

B. US - No Federal AI Act, but 5 regulators:

1. CFPB Circular 2022-03 + 2023: Lenders using AI must provide Adverse Action Notice with specific reasons - Black box not excuse. "Complexity of AI is not defense for ECOA violation."
2. ECOA + Fair Lending: AI must not have disparate impact on protected class.
3. SEC - Robo-Advisor: Must disclose conflict, data used.
4. Federal Reserve SR 11-7: Model Risk Management - All AI models must be validated, back-tested.

C. India - RBI Framework

1. RBI Free-AI Committee - Framework for Responsible and Ethical Enablement of AI in Financial Sector - Report 8 Dec 2024: 7 Sutras: Trust is foundation, People first, Fairness, Accountability, Transparency, Privacy, Robustness.
2. RBI Digital Lending Guidelines 2022 + 2023: AI lending apps must disclose all data collected, cannot access contact list, must explain credit decision.
3. SEBI Circular on use of AI/ML in Stock Market - 4 Jan 2019 + 2024: Robo-advisors must be registered Investment Adviser, disclose logic, back-testing, maintain audit trail.

3. 7 Pillars of Ethical AI in Financial Services

1. Fairness & Non-Discrimination - No Bias
Training data must be balanced. Test for disparate impact - e.g., if AI rejects 60% women vs 30% men for same credit score, it is biased.

2. Explainability - XAI
Customer has right to explanation under GDPR Art 22 + EU AI Act + ECOA. Must provide reason codes - e.g., "Rejected due to high DTI ratio 55%, low bureau score 620, not due to gender."

3. Transparency & Disclosure
Must inform customer: "You are interacting with AI Chatbot", "Loan decision made by AI model XYZ, data sources: Bureau, Bank Statement, etc."

4. Accountability & Human-in-Loop
Every AI decision must have human owner. For high-value loan > Rs.1 Cr, RBI Free-AI recommends human override mandatory.

5. Privacy & Data Minimization
RBI Digital Lending - Cannot scrape SMS, contacts, media. Only need-based data. Consent must be explicit, purpose-limited.

6. Robustness & Safety
Model must be tested for adversarial attacks, data drift, model decay. Must have fallback.

7. Auditability
All AI decisions must be logged for 7 years, model version, data snapshot stored.

4. Leading Case Laws - Ethical AI Failure in Finance

1. Apple Card - Goldman Sachs - New York DFS Investigation - 2019-2021 - Gender Bias Case - Most Famous

Facts: David Heinemeier Hansson tweeted - He got 20x credit limit than his wife despite joint tax returns, same assets. Steve Wozniak same complaint. NYDFS investigated Goldman Sachs Apple Card AI model.

Finding: Algorithm did not intentionally discriminate, but used proxies like credit history length that had disparate impact on women. Goldman could not explain model.

Outcome: NYDFS 2021 Report - No intentional discrimination found, but ordered Goldman to improve explainability and re-train model. Set precedent - Even unintentional disparate impact = ECOA violation if AI not explainable.

Principle: "Black Box is not defense"

2. CFPB vs Upstart Network - 2021-2023 - Consent Order

Upstart uses AI with 1000+ variables - Education, field of study, GPA for credit scoring. CFPB found - Model approved more borrowers but charged higher APR to Black and Hispanic borrowers.

CFPB required Upstart to monitor for disparate impact, report, and improve adverse action notices. First CFPB action on AI lending bias.

3. Netherlands - SyRI Case - Dutch Court - Hague District Court - 5 Feb 2020 - C-09-550982

SyRI - System Risk Indication - Dutch welfare fraud detection AI used by Government + Banks to detect fraud. Court held - System violated Article 8 ECHR Privacy, lacked transparency, discriminated against low-income neighborhoods.

Struck down entire AI system. Applied to banking - If bank's fraud AI targets only low-income pin codes, it is discriminatory and illegal.

4. UK - FCA vs WealthTech - Robo-Advisor Failures - 2023

FCA fined WealthTek £84k and banned robo-advisor for unsuitable advice - AI recommended high-risk products to low-risk clients without explainability. FCA held - Firm cannot hide behind AI.

5. US - State Farm - Home Insurance AI - Class Action - 2022 - Pending

Plaintiffs allege State Farm's Claims Fraud AI flagged Black homeowners disproportionately for fraud investigation. Case under Fair Housing Act. Shows insurance AI also high-risk.

6. LendingClub - FTC $18M Settlement - 2023

LendingClub AI charged hidden fees, misled customers about AI approval - FTC held - AI transparency must include fees and final APR, not just teaser rate.

7. India - No direct AI bias case yet, but:

RBI Action on Digital Lending Apps - 2021-2022: RBI banned 200+ apps for unethical AI - Scraping contacts, shaming borrowers, AI deciding credit limit based on contacts and photos. Banned under Digital Lending Guidelines.

SEBI vs Investment Adviser using AI - 2023: SEBI penalized registered IA for using ChatGPT to generate stock recommendations without disclosure - Held - AI-generated advice must disclose that it is AI and must be supervised by human IA.

5. How to Build Ethical AI Compliant Lending System - RBI + EU Checklist

For Bank / NBFC / Fintech:

1. Model Risk Documentation - SR 11-7: Data sheet, feature importance, bias test - Disparate Impact Ratio >0.8 [4/5th rule]
2. Adverse Action Notice: Provide 4 specific reasons - Not "AI model rejected"
3. Human Override: Credit Officer must be able to override AI with reason logging
4. Consent Architecture: Explicit consent - "We will use your bank statement data from Account Aggregator to train AI model - Purpose: Credit Assessment"
5. Audit Trail: Log - Model version v2.3, Data snapshot ID, Score 720, Decision, Human override Y/N
6. Fairness Test Quarterly: Test model on protected groups - Gender, Caste proxies like surname, Pin code
7. Explainability Tool: Use SHAP / LIME to generate reason codes

Penalty in India:

• RBI - Under BR Act 1949 - Penalty up to Rs.1 Cr + Cancellation of digital lending license
• SEBI - Under SEBI Act - Rs.25 Cr or 3x profit
• Consumer Court - Under Consumer Protection Act 2019 - AI bias = Deficiency in service - Compensation

6. Future - 2026 Onwards

• EU AI Act High-Risk obligations become mandatory from 2 Aug 2026 for financial AI.
• RBI will issue final guidelines on Free-AI in 2025 - Expected to make AI Audit mandatory for banks with asset > Rs.50,000 Cr
• SEBI will require BRSR Core disclosure of AI use - "What % of lending decisions made by AI, what bias testing done"