1. Legal Definition - Two Types of Discrimination

A. Disparate Treatment - Direct use of protected attribute.
E.g., Model includes gender, race, religion, caste. Even if you don't include it, if you include title = Mr/Mrs, it's direct.

B. Disparate Impact - Neutral variable has disproportionate adverse effect on protected group, and no business justification.
This is 95% of algorithm cases. Model never sees gender, but it uses:

• PIN code = proxy for caste/community/race
• Phone model = Android low-end = proxy for income/race
• University = HBCU / Tier 3 college = proxy for race
• Working hours = night shift = proxy for gender
• Shopping at specific stores = proxy for religion

Under Texas Dept. of Housing vs Inclusive Communities - US Supreme Court 2015, disparate impact is illegal under Fair Lending even without intent, if plaintiff shows statistical disparity and lender cannot show business necessity.

2. Where Discrimination Enters Scoring Algorithm

Stage 1 - Training Data Bias: Historical loans were biased - e.g., women got less credit in past - model learns that pattern.

Stage 2 - Feature Selection Bias: Alternative data like social media, device data correlates with protected class.

Stage 3 - Label Bias: You train on default but default is influenced by bank's own harsh collection on certain groups.

Stage 4 - Feedback Loop: Model rejects group -> group gets less credit history -> future model sees them as thin-file risky -> loop.

3. Legal Frameworks

USA:

• ECOA Reg B - prohibits discrimination on race, color, religion, national origin, sex, marital status, age
• FHA - housing loans
• CFPB Circular 2022-05 - Black-box is no defense - must provide specific adverse action reasons
• Fair Lending Act - Disparate impact ratio must be >0.8 (4/5ths rule)

India:

• RBI Fair Practices Code - No discrimination on gender, caste, religion
• Article 14, 15 of Constitution - State action includes public sector banks
• Digital Lending Guidelines 2022 - Reason for rejection mandatory
• DPDP Act 2023 Sec 7 - Automated decision must be transparent

EU:

• AI Act 2024 - Credit scoring = High Risk - banned if leads to discrimination
• GDPR Art 22 - Right not to be subject to solely automated decision with legal effect

4. Landmark Case Laws

CASE 1: Apple Card / Goldman Sachs - NYDFS 2019-2021 - Gender Discrimination by Proxy

Facts: David Heinemeier Hansson got 20x credit limit vs wife with same income, same joint account. Steve Wozniak same complaint. Alleged algorithm discriminated women.

NYDFS Investigation: 400,000 NY applications reviewed, thousands of pages, interviews.

Result: No intentional bias found. But NYDFS forced Goldman to prove model: Used credit score, utilization, missed payments, not gender. However Goldman had to change model to add manual review.

Risk Lesson: Even if cleared, cost of investigation is huge. Case established that "similarly situated applicants must get similar limits" - If model gives different limits to husband/wife with same finances, proxy discrimination presumed.

CASE 2: CFPB vs Upstart Network - 2021-2024 - Education Discrimination

Upstart used AI with 1600 variables - education: college name, GPA, field of study, employment.

Finding: Model charged higher APR to graduates of Historically Black Colleges (HBCUs) and to women with certain degrees (e.g., teaching vs engineering). APR difference 27%.

Consent Order: Upstart must:

• Stop using college name as predictor unless it proves predictive beyond bias
• Monitor adverse impact by race using BISG proxy method
• Provide specific adverse action reasons, not "education risk"

This is now template for all fintechs.

CASE 3: SafeRent Solutions (Formerly CoreLogic) - 2022-2023 - Massachusetts AG - Credit Score + Criminal History

SafeRent provided tenant credit scoring algorithm to landlords. Algorithm used credit score + criminal record to generate risk score.

Black and Hispanic applicants were rejected 3x more. Algorithm used ZIP code as proxy.

Settlement: $2.2M, agreement to remove criminal history from model, stop using credit score below threshold for housing.

Principle: Credit scoring algorithm used for housing = FHA violation if disparate impact.

CASE 4: Navy Federal Credit Union - 2023 - DOJ + CFPB - Largest Credit Union

Navy Federal mortgage model: Black borrowers denied 2x vs white, Latino 1.5x, even after controlling for income, credit score.

DOJ analysis: Model used "branch location" and "membership duration" which correlated with race - branches in military bases had less diverse staff.

Settlement: $95M compensation, model overhaul.

CASE 5: Texas Department of Housing vs Inclusive Communities Project - US Supreme Court 2015 - Foundation

Held: Disparate impact claims are cognizable under Fair Housing Act. Plaintiff must show statistical disparity, burden shifts to defendant to show legitimate business interest.

This case is cited in every algorithm discrimination case - allows statistical proof without showing intent.

CASE 6: State Farm / Allstate Credit-Based Insurance Scoring - Multiple US States 2005-2018

Insurance credit scoring used credit score to price auto insurance. Minorities with same driving record paid 30-40% higher because credit score correlated with race.

NAACP vs State Farm settlement: State Farm agreed to remove credit scoring in 5 states.

Extension to Lending: Same logic applied to credit scoring - if your score uses data that correlates with protected class, it's illegal unless you prove necessity.

CASE 7: HDFC Bank / Indian Fintechs - PIN Code Discrimination - RBI Ombudsman Cases 2022-24 (Unreported but pattern)

Several Indian fintechs rejected loans for PIN codes 1100xx (Delhi's certain minority concentration areas) or North East states, using geolocation risk.

RBI Ombudsman held: Using PIN code as risk factor without individual assessment violates Fair Practices Code - indirect discrimination based on religion/region. Banks ordered to remove PIN code filter.

This is India's Apple Card equivalent - not published as judgment but as ombudsman order.

5. How to Test Discrimination Risk - The 4/5ths Rule + SHAP

Step 1 - Statistical Test:
Calculate Adverse Impact Ratio = Approval rate of protected group / Approval rate of control group

If <0.8, disparate impact presumed.

E.g., Women approval 40%, Men 60% = 40/60 = 0.66 <0.8 = violation.

Step 2 - Proxy Hunt:
Run SHAP values for each variable. If device_type or zip code has high SHAP and correlates with protected class (check correlation matrix), remove it.

Step 3 - Less Discriminatory Alternative:
If variable is predictive (e.g., college), test if alternative variable (e.g., income) achieves same predictive power with less bias. If yes, must use alternative.

6. How to Make Your Model Compliant

For RBI FREE-AI compliance:

1. Ban List: Never include - gender, religion, caste, marital status, PIN code first 3 digits, phone model, language.
2. Documentation: Keep Model Card - what data, why each feature, bias test results.
3. Adverse Action: Provide top 4 specific reasons - not "low score" but "high utilization 78%, 3 missed payments in last 6 months, debt-to-income 52%".
4. Human Review: Auto-rejection above Rs 50k must have human override option.
5. Annual Bias Audit: By external auditor using BISG method.