The Black Box

Why Explainability Matters

🎱

Complex models (deep nets, random forests) make great predictions but can't explain themselves. Accuracy ≠ trustworthiness.

🔦

LIME

Local Interpretable Model-agnostic Explanations. Perturb the input slightly and see which changes flip the output.

📊

SHAP Values

SHapley Additive exPlanations. Game theory tells us each feature's exact contribution to the final prediction.

⚖️

Legal Right

EU AI Act 2024 + GDPR Article 22 give citizens a legal right to explanation for automated AI decisions.

🎱

Wizzy the AI Tutor

Imagine applying for a loan and the bank's AI says "DENIED" — with no explanation! 😠 This happens millions of times a day worldwide. The AI has 200 million parameters and no one knows why it said no. This is the black box problem. Let's crack it open!

Step 1 — The Black Box Problem

🎱 AI Loan Officer

200 million parameters · No explanation

→

Adjust the applicant's profile:

The AI makes its decision but tells you nothing about why. Change the inputs — can you figure out what it cares about?

📊 Decision History

Current decision—

Approved tests0

Denied tests0

Your guessed reasonUnknown

⚖️ Real Case: In 2019, Apple Card was found to give women lower credit limits than men with identical financial profiles. The bank said "the algorithm decided" — no explanation possible.

🎱

Wizzy the AI Tutor

LIME works by poking the AI! 🔦 It creates hundreds of slightly different versions of the input (perturbing values up and down) and asks the AI each time. Then it finds which changes flipped the decision — those are the important features! Click "Perturb" to see LIME in action.

Step 2 — LIME: Perturbing the Input

// Press Run LIME Analysis to start perturbation...

Perturbation results (which changes flip the decision):

🔦 LIME Results

Run LIME to see results

LIME creates a simple, interpretable model that approximates the black box locally — around the specific input being explained.

🎱

Wizzy the AI Tutor

SHAP values use game theory! 📊 Imagine the features as players in a team. SHAP asks: "How much did each player contribute to the win (or loss)?" Positive SHAP = pushed toward approval. Negative SHAP = pushed toward denial. Every prediction can be decomposed into feature contributions!

Step 3 — SHAP Values: Feature Contributions

Base prediction + feature contributions = final score

Base: 0.50 = —

📊 SHAP Summary

SHAP values sum to the difference between the actual prediction and the base rate prediction.

Game Theory: SHAP is based on Shapley values from cooperative game theory. It's the only method that satisfies efficiency, symmetry, dummy, and additivity axioms.

🎱

Wizzy the AI Tutor

Counterfactuals answer: "What would have to change for the AI to say YES?" 🔄 This is the most human-friendly explanation — it gives actionable advice. "If your credit score was 50 points higher, you'd be approved." Click Generate Counterfactual to find the minimum change needed!

Step 4 — Counterfactual Explanations

❌ Original Application — DENIED

Counterfactuals find the minimum-edit version of the input that flips the decision. They give actionable, human-understandable advice.

🔄 Counterfactual Paths

EU AI Act Article 86: High-risk AI systems must provide "meaningful information about the logic involved, as well as the significance and the envisaged consequences" of automated decisions.

🎱

Wizzy the AI Tutor

🎊 You can now explain any AI decision! You understand LIME (local perturbation), SHAP (feature attribution via game theory), and counterfactuals (minimum-edit explanations). These are the tools AI engineers use every day to make AI trustworthy!

🎱

XAI Expert Badge!

You mastered Explainable AI — LIME, SHAP, and counterfactuals!

🎱 WhizzStep AI Lab

This certifies that

Student Name

has mastered Explainable AI — LIME, SHAP & Counterfactuals

XAI Expert

SHAP Master

AI Auditor

whizzstep.in

Key Concepts Mastered

Black Box

🎱 The Problem

Complex models like deep nets can't explain their own decisions. High accuracy comes at the cost of interpretability.

LIME

🔦 Local Explanation

Perturb the input, observe output changes, fit a simple interpretable model locally. Works on any ML model — model-agnostic.

SHAP

📊 Feature Attribution

Shapley values from game theory. Distributes the "credit" for a prediction fairly across all input features.

Counterfactual

🔄 Actionable Advice

"If X was different, the decision would change." The minimum edit needed to flip the output. Human-friendly and actionable.

EU AI Act

⚖️ Legal Requirement

High-risk AI (loans, hiring, justice) must explain decisions. XAI is now a legal requirement in the EU, not a nice-to-have.

Trust

🤝 Why It Matters

Humans won't use AI they don't trust. Explanations build trust — and catch bias and errors before they cause harm.