AI Ethics

# Example: Detecting and Mitigating Bias in a Hiring AI
import pandas as pd
from sklearn.linear_model import LogisticRegression
from fairlearn.metrics import demographic_parity_difference, equalized_odds_difference
# Sample hiring data
data = pd.DataFrame({
    'experience': [5, 3, 7, 2, 4, 6, 8, 3],
    'education': [1, 0, 1, 0, 1, 1, 1, 0],  # 1=degree, 0=no degree
    'gender': [1, 0, 1, 0, 1, 0, 1, 0],  # 1=male, 0=female
    'hired': [1, 0, 1, 0, 1, 0, 1, 1]
})
# Train model WITHOUT considering fairness
model_unfair = LogisticRegression()
X = data[['experience', 'education', 'gender']]
y = data['hired']
model_unfair.fit(X, y)
predictions_unfair = model_unfair.predict(X)
# Check for demographic parity - should be close to 0 for fairness
# Measures if selection rates are equal across groups
dp_diff = demographic_parity_difference(
    y_true=y,
    y_pred=predictions_unfair,
    sensitive_features=data['gender']
)
print(f"Demographic Parity Difference (unfair model): {dp_diff:.3f}")
print("Value > 0.1 indicates potential bias")
# MITIGATION STRATEGY 1: Remove sensitive features
print("\n--- Mitigation 1: Remove Gender Feature ---")
X_fair = data[['experience', 'education']]  # Remove gender
model_fair = LogisticRegression()
model_fair.fit(X_fair, y)
predictions_fair = model_fair.predict(X_fair)
dp_diff_fair = demographic_parity_difference(
    y_true=y,
    y_pred=predictions_fair,
    sensitive_features=data['gender']
)
print(f"Demographic Parity Difference (fair model): {dp_diff_fair:.3f}")
# MITIGATION STRATEGY 2: Use fairness constraints
from fairlearn.reductions import ExponentiatedGradient, DemographicParity
print("\n--- Mitigation 2: Fairness Constraints ---")
constraint = DemographicParity()
mitigator = ExponentiatedGradient(
    LogisticRegression(),
    constraints=constraint
)
mitigator.fit(X, y, sensitive_features=data['gender'])
predictions_constrained = mitigator.predict(X)
dp_diff_constrained = demographic_parity_difference(
    y_true=y,
    y_pred=predictions_constrained,
    sensitive_features=data['gender']
)
print(f"Demographic Parity Difference (constrained): {dp_diff_constrained:.3f}")
# BEST PRACTICE: Regular auditing and monitoring
print("\n--- Best Practices ---")
print("1. Audit datasets for representation across protected groups")
print("2. Monitor fairness metrics in production")
print("3. Establish thresholds for acceptable bias levels")
print("4. Regular retraining with updated, balanced data")
print("5. Human review for borderline cases")