Session 25: Ethical AI & Responsible Insurance

1. Why Ethics in Insurance AI?

Insurance is not just another industry adopting AI. Insurance is a social good — it provides financial security, enables risk-taking (entrepreneurship, home ownership, investment), and acts as a safety net for individuals and businesses when things go wrong. When an insurance AI makes a biased decision, the harm is not a wrong product recommendation or an inaccurate search result — it is a family denied a life insurance payout, a small business that cannot get property insurance in its neighbourhood, or a person charged a higher premium because of their zip code, which correlates with their race.

Ethical AI in insurance is not about "being fair" in an abstract sense. It is about: (a) regulatory compliance — IRDAI and the DPDP Act require explainability and fairness in automated decisions, (b) customer trust — insurance depends on trust more than any other financial service, and a single ethical failure can destroy years of brand investment, and (c) business sustainability — a model that systematically disadvantages a customer segment will eventually be discovered, and the resulting regulatory and reputational damage can exceed any short-term profit the model generated.

1.1 Real Cases of Discriminatory AI in Insurance

2. Ethical AI Principles

Seven principles form the foundation of ethical AI in insurance. These principles are derived from: the OECD AI Principles (adopted by 40+ countries), the EU AI Act (which classifies insurance underwriting as "high-risk AI"), IRDAI's consultation papers on AI governance, and the industry's own learnings from the cases described in Section 1. Every principle translates into a specific requirement for insurance AI systems.

3. Algorithmic Bias — Types & Detection

Algorithmic bias is not a single problem — it is five distinct problems that can arise at different points in the ML lifecycle. An ethical AI team must be able to identify which type(s) of bias affect their system and apply the appropriate mitigation strategy.

3.1 The Five Types of Bias in Insurance ML

3.2 Detecting Bias in Practice

import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.metrics import roc_auc_score, classification_report

# Load dataset — we'll create a scenario where bias may exist
df = pd.read_csv('data/insurance_cleaned.csv')

# Simulate a protected characteristic (e.g., "urban vs. rural")
np.random.seed(42)
df['area_type'] = np.random.choice(['Urban', 'Rural', 'Semi-Urban'], size=len(df), p=[0.55, 0.2, 0.25])

# Create target: claim filed or not
claims_per_policy = df.groupby('policy_id').agg(
    has_claim=('claim_id', lambda x: 1 if len(x) > 0 else 0),
    premium=('premium', 'first'),
    age=('age', 'first'),
    income=('income', 'first'),
    credit_score=('credit_score', 'first'),
    policy_type=('policy_type', 'first'),
    area_type=('area_type', 'first')
).reset_index()

# Split by area type for bias analysis
bias_check = claims_per_policy.copy()
features = ['premium', 'age', 'income', 'credit_score', 'policy_type']
area_groups = claims_per_policy['area_type'].unique()

print("=" * 65)
print("BIAS DETECTION — MODEL PERFORMANCE BY GROUP")
print("=" * 65)
print(f"Checking for differential performance across area types...\n")

for area in area_groups:
    group_data = claims_per_policy[claims_per_policy['area_type'] == area].copy()
    if len(group_data) < 100:
        continue

    X_g = group_data[[c for c in features if c in group_data.columns]]
    y_g = group_data['has_claim']

    cat_g = X_g.select_dtypes(include=['object']).columns.tolist()
    num_g = X_g.select_dtypes(include=[np.number]).columns.tolist()

    if len(group_data) > 200:
        # Train simple model on this group
        X_tr, X_te, y_tr, y_te = train_test_split(X_g, y_g, test_size=0.3, random_state=42, stratify=y_g)
        preproc = ColumnTransformer(transformers=[
            ('num', StandardScaler(), num_g),
            ('cat', OneHotEncoder(drop='first', handle_unknown='ignore', sparse_output=False), cat_g)
        ])
        model = Pipeline(steps=[('preprocessor', preproc), ('classifier', RandomForestClassifier(
            n_estimators=100, class_weight='balanced', random_state=42))])
        model.fit(X_tr, y_tr)
        y_prob = model.predict_proba(X_te)[:, 1]
        auc = roc_auc_score(y_te, y_prob)
        claim_rate = y_g.mean() * 100
        print(f"{area:15s}: n={len(group_data):,}, claim_rate={claim_rate:.1f}%, "
              f"model AUC={auc:.3f} {'⚠' if auc < 0.6 else '✓'}")

print(f"\n{'─' * 65}")
print(f"Bias detected if: AUC differs by >0.1 between groups OR")
print(f"claim rates differ by >5% without explanation.")
print(f"{'─' * 65}")
print(f"\nIf the model performs significantly worse for one group:")
print(f"1. Check if that group is underrepresented (representation bias)")
print(f"2. Check if the relationship between features and risk differs (aggregation bias)")
print(f"3. Consider: separate models per group, or features better suited to that group")
print(f"4. Document the disparity — even if unmitigated, transparency is required")

4. Bias Detection Exercise — Proxy Discrimination

Proxy discrimination occurs when a facially neutral variable (e.g., zip code, occupation, credit score) is correlated with a protected characteristic (e.g., race, caste, religion) — and using that variable in a model has the same practical effect as using the protected characteristic directly. Proxy discrimination is the most common and most dangerous form of algorithmic bias in insurance AI because it is invisible: the model does not contain any prohibited variables, yet it produces discriminatory outcomes.

4.1 The Proxy Discrimination Exercise

In this exercise, we examine whether using "locality" as a pricing factor produces outcomes that disadvantage a particular social group — even though "locality" itself is not a protected characteristic.

# Simulate a dataset for proxy discrimination analysis
np.random.seed(42)
n = 10000

# Create synthetic data
data = pd.DataFrame({
    'customer_id': range(1, n + 1),
    'locality': np.random.choice(['Area_A', 'Area_B', 'Area_C', 'Area_D'], size=n),
    'income': np.random.lognormal(mean=11.5, sigma=0.5, size=n),
    'age': np.random.randint(18, 80, size=n),
    'vehicle_age': np.random.randint(0, 15, size=n),
    'claim_history': np.random.poisson(0.2, size=n),
    'credit_score': np.random.normal(720, 80, size=n)
})

# Create systematic relationship: locality is correlated with socioeconomic status
# Area_A and Area_B are higher-income areas (proxy for advantaged group)
# Area_C and Area_D are lower-income areas (proxy for disadvantaged group)
locality_map = {'Area_A': 1, 'Area_B': 0.9, 'Area_C': 0.5, 'Area_D': 0.3}
data['socioeconomic_score'] = data['locality'].map(locality_map) + np.random.normal(0, 0.1, n)

# Adjust credit_score and claim_history to correlate with locality
data.loc[data['locality'].isin(['Area_A', 'Area_B']), 'credit_score'] += 30
data.loc[data['locality'].isin(['Area_C', 'Area_D']), 'credit_score'] -= 20
data['credit_score'] = np.clip(data['credit_score'], 400, 900)

# Simulate claim risk: lower socioeconomic areas have higher claim risk
# (could be due to worse infrastructure, older vehicles, more accident-prone roads)
risk_base = 0.08 + 0.04 * (1 - data['socioeconomic_score'])
data['claim_risk'] = np.clip(risk_base + np.random.normal(0, 0.02, n), 0.02, 0.30)

# Calculate premium using a model that DOES NOT directly use locality
# But uses credit_score and vehicle_age — which are correlated with locality
data['premium'] = (5000 +
                   2000 * (data['claim_risk'] / 0.10) +
                   1500 * (1 - data['credit_score'] / 800) +
                   500 * (data['vehicle_age'] / 10))

# Check average premium by locality — this is the proxy discrimination test
print("=" * 70)
print("PROXY DISCRIMINATION ANALYSIS")
print("=" * 70)
print(f"{'Locality':12s} {'Count':>8s} {'Avg Premium':>15s} {'Avg Credit':>15s} {'Avg Risk':>15s}")
print("-" * 65)
for loc in ['Area_A', 'Area_B', 'Area_C', 'Area_D']:
    sub = data[data['locality'] == loc]
    print(f"{loc:12s} {len(sub):>6,d}   ₹{sub['premium'].mean():>8,.0f}   {sub['credit_score'].mean():>6.0f}   {sub['claim_risk'].mean()*100:>4.1f}%")

# The differential premium between Area_A and Area_D may be "explained" by credit_score
# But if credit_score itself reflects historical socioeconomic disadvantage,
# then locality is a proxy, and the premium differential is discriminatory

prem_diff = data[data['locality'] == 'Area_A']['premium'].mean() - data[data['locality'] == 'Area_D']['premium'].mean()
print(f"\n{'─' * 65}")
print(f"Premium difference: Area_A vs. Area_D = ₹{prem_diff:,.0f}")
print(f"\nInterpretation:")
print(f"The model does NOT use 'locality' as a feature, but the premium")
print(f"difference across localities is ₹{prem_diff:,.0f}. This difference is driven by")
print(f"correlation between locality and credit_score/vehicle_age — which are")
print(f"themselves correlated with socioeconomic status.")
print(f"\nThis is PROXY DISCRIMINATION if:")
print(f"1. Locality correlates with a protected characteristic (caste, race, income)")
print(f"2. The insurer cannot justify the premium difference with a legitimate")
print(f"   business rationale that is independent of the protected characteristic")
print(f"3. A different pricing approach could achieve similar risk accuracy")
print(f"   with less discriminatory impact")

# Disparate Impact Ratio (DIR): compare selection rates across groups
# Using a hypothetical premium threshold
threshold = data['premium'].median()
for loc in ['Area_A', 'Area_B', 'Area_C', 'Area_D']:
    sub = data[data['locality'] == loc]
    above_threshold = (sub['premium'] > threshold).mean()
    print(f"  {loc}: % above median premium = {above_threshold*100:.0f}%")

# The Disparate Impact Ratio compares the % of favourably treated individuals
# in the disadvantaged group vs. the most advantaged group
favourable_A = (data[data['locality'] == 'Area_A']['premium'] <= threshold).mean()
favourable_D = (data[data['locality'] == 'Area_D']['premium'] <= threshold).mean()
dir_ratio = favourable_D / favourable_A

print(f"\nDisparate Impact Ratio (Area_D vs. Area_A): {dir_ratio:.2f}")
if dir_ratio < 0.80:
    print("⚠ DIR < 0.80 — this is the US 'four-fifths rule' threshold for disparate impact.")
    print("  The premium model may be discriminatory in effect even if not in intent.")
else:
    print("✓ DIR ≥ 0.80 — within the acceptable range under the four-fifths rule.")

5. Bias Mitigation Strategies

When bias is detected, three categories of mitigation are available — corresponding to three stages of the ML lifecycle. The choice of strategy depends on: when in the development process the bias is detected, the availability of data for the affected groups, and the business context of the model (a fraud detection model may tolerate different tradeoffs than an underwriting pricing model).

5.1 The Three Mitigation Approaches