“The model said no.” That’s not an explanation. When AI influences loan decisions, hiring recommendations, or content moderation, stakeholders deserve to understand why. Explainability isn’t a nice-to-have – it’s a requirement for trust.

The Three Pillars of Responsible AI

Explainability: Opening the Black Box

Two techniques dominate the explainability space: LIME and SHAP. Both answer the same question – “which inputs mattered?” – but approach it differently.

LIME: Local Interpretable Model-agnostic Explanations

LIME explains individual predictions by perturbing inputs and watching how outputs change:

1. Take a prediction you want to explain
2. Create variations of the input (change words, values, features)
3. See how the model’s output changes for each variation
4. Build a simple model that approximates this behavior locally
5. The simple model’s weights become the explanation

SHAP: Shapley Additive Explanations

SHAP uses game theory. It treats each input feature as a “player” and calculates each player’s contribution to the final “payout” (prediction):

# SHAP values answer: "How much did each feature contribute?"
shap_result = {
    "feature_contributions": {
        "income": +0.38,      # Pushed toward approval
        "credit_score": +0.31, # Pushed toward approval
        "debt_ratio": -0.22,   # Pushed toward denial
        "age": +0.05          # Minimal impact
    },
    "base_value": 0.5,       # Average prediction
    "final_prediction": 0.82 # Sum of base + contributions
}

The key insight: SHAP values are additive. You can literally add them up to get the final prediction. This makes explanations mathematically rigorous.

When to Use Which

• LIME: Fast, good for text/image explanations, easier to understand
• SHAP: More theoretically grounded, better for tabular data, consistent explanations
• Both: When you need cross-validation of explanations

Fairness: Beyond Individual Explanations

Explainability shows what influenced a decision. Fairness analysis asks whether the system treats different groups equitably.

The 4/5 Rule (Disparate Impact)

A common fairness benchmark: the selection rate for a protected group should be at least 80% (4/5) of the selection rate for the majority group.

IBM AIF360 Integration

The AI Fairness 360 toolkit provides standardized fairness metrics. Running it locally means you can:

• Test fairness before deployment
• Monitor fairness metrics in production
• Compare different model versions for fairness improvements
• Generate compliance reports with specific metric values

# Fairness analysis structure
fairness_report = {
    "protected_attribute": "gender",
    "privileged_group": "male",
    "unprivileged_group": "female",
    "metrics": {
        "disparate_impact": 0.87,
        "statistical_parity_difference": -0.08,
        "equal_opportunity_difference": -0.12
    },
    "passed_4_5_rule": True
}

Impact Assessment: Before and After

Before deploying AI, assess potential impacts. After deployment, track actual outcomes.

Continuous Monitoring

Impact assessment isn’t one-and-done. Track metrics over time to catch drift:

# Impact monitoring tracks trends
impact_trend = {
    "period": "2024-Q4",
    "decisions_made": 28500,
    "approval_rate": 0.67,
    "fairness_metrics": {
        "disparate_impact": {
            "current": 0.87,
            "previous_period": 0.89,
            "trend": "declining"  # Flag for investigation
        }
    },
    "appeals_filed": 142,
    "appeals_successful": 23
}

Why Running This Locally Matters

Cloud AI platforms offer “responsible AI” features. But can you:

• See the actual SHAP values, not just a summary?
• Configure which fairness metrics matter for your use case?
• Run impact assessments on your specific data?
• Export detailed audit trails for compliance?

Running explainability locally gives you:

• Full transparency: Every calculation is visible and verifiable
• Customization: Define fairness metrics that match your requirements
• Data control: Sensitive data never leaves your environment
• Audit capability: Generate compliance reports with actual numbers

The Responsible AI Stack

These components work together:

1. Before prediction: Impact assessment identifies risks
2. During prediction: Guardrails catch obvious problems (Part 3)
3. After prediction: LIME/SHAP explain the decision
4. Across predictions: Fairness metrics measure aggregate behavior
5. Over time: Trend monitoring catches drift

Coming Up Next

Explainability shows why decisions were made. But who has access to make those decisions? Who can see the data? In the next post, we’ll explore Governance – access control, audit trails, and cost management that complete the trust picture.