After studying this material, you should be able to:

1. Explain why AI governance is necessary at scale
2. Describe the five pillars of Uber's Responsible AI program
3. Understand how visibility, explainability, and governance are implemented technically
4. Identify key tools and methods used (Model Catalog, SHAP, PFI, Integrated Gradients)
5. Apply the concept of "shift-left governance" to AI development workflows

---

Step-by-Step Teaching Guide

Step 1: Why Does Responsible AI Matter?

The Core Problem

Imagine a company where hundreds of AI models make decisions every second:

• Calculating your ride price
• Predicting your driver's arrival time
• Detecting fraud on your account

Key Question: How do you ensure all these models are fair, reliable, and accountable?

The Stakes

Core Insight: As AI scales, governance cannot be an afterthought. It must be built into the system.

---

Step 2: Understanding the Five Pillars Framework

Uber's solution rests on five interconnected pillars. Think of them as building blocks stacked on each other:

        [ADOPTION]          ← Scaling it company-wide
           ↑
       [EDUCATION]          ← Teaching people to use it
           ↑
       [GOVERNANCE]         ← Enforcing the rules
           ↑
     [EXPLAINABILITY]       ← Understanding how models work
           ↑
        [VISIBILITY]        ← Knowing what models exist

Each pillar depends on the one below it. You cannot govern what you cannot see. You cannot explain what you do not understand.

---

Step 3: Pillar 1 — Visibility (Knowing What Exists)

The Tool: Model Catalog

Think of the Model Catalog like a library catalog for AI models.

Just as a library catalog tells you:

• What books exist
• What they are about
• Where to find them

The Model Catalog tells you:

• What AI models are deployed
• What they do
• How they perform

The Model Card

Each model gets a Model Card — a standardized document containing:

Why This Matters

Without a centralized inventory, teams duplicate work, governance teams cannot audit models, and no one has a complete picture of AI risk.

Key Design Principle: Many fields are automatically populated from system signals, reducing manual burden on engineers.

---

Step 4: Pillar 2 — Explainability (Understanding How Models Work)

The Black Box Problem

Most AI models are like a black box:

INPUT → [????BLACK BOX????] → OUTPUT

You put data in, you get a prediction out, but you don't know why.

Explainability turns this into a glass box:

INPUT → [TRANSPARENT MODEL] → OUTPUT
         ↓
    "Feature X contributed 40%
     Feature Y contributed 35%
     Feature Z contributed 25%"

Three Levels of Explanation

Uber implements three complementary methods:

---

Level 1: Global Attributions — The Big Picture

Method: Permutation Feature Importance (PFI)

How it works:

1. Train your model normally
2. Take one feature (e.g., "distance")
3. Randomly shuffle its values across all data points
4. Measure how much model performance degrades
5. Large degradation = that feature is very important

Example:
Normal accuracy: 92%
After shuffling "distance": 71%  → Distance is VERY important (−21%)
After shuffling "weather": 91%   → Weather is less important (−1%)

Use case: Understanding overall model behavior and risk

---

Level 2: Local Attributions — The Individual Prediction

Method: SHAP (SHapley Additive exPlanations)

How it works:

• Borrowed from game theory
• Asks: "How much did each feature contribute to THIS specific prediction?"

Example — Why was this ETA 12 minutes?

Base prediction:     8 min
+ Traffic feature:  +3 min
+ Distance feature: +2 min
+ Time of day:      −1 min
= Final ETA:        12 min

For tree-based models (like decision trees, random forests), Uber uses TreeSHAP, which is computationally efficient.

Use case: Debugging individual predictions, explaining decisions to users

---

Level 3: Integrated Gradients — Deep Learning Models

Method: Integrated Gradients

How it works:

• Used for neural networks
• Starts from a baseline input (e.g., all zeros)
• Gradually moves toward the actual input
• Accumulates gradients along this path to measure each feature's contribution

Baseline Input → → → → → Actual Input
(all zeros)    gradients   (real data)
               accumulated
               = attribution

Use case: Explaining neural network predictions where other methods don't apply

---

Summary Comparison

---

Step 5: Pillar 3 — Governance (Enforcing Responsible Practices)

The Shift-Left Approach

Traditional governance catches problems late:

Design → Build → Test → Deploy → [GOVERNANCE CHECK] ← Too late!

Shift-left moves governance checks earlier:

[GOVERNANCE CHECK] → Design → Build → Test → Deploy
         ↑
    Catch problems here, when they're cheap to fix

How Uber Implements This

1. Document Review Entry Point

   • Engineers write PRDs (Product Review Documents) and ERDs (Engineering Review Documents)
   • Governance checks are embedded at this stage
   • Problems are caught before any code is written

2. In-Product Enforcement

   • Engineers are prompted to complete Model Cards before deployment
   • Acts as a gating mechanism — you cannot deploy without completing governance steps

Analogy: Like a seatbelt warning in a car. The car won't let you ignore safety. Governance is built into the workflow, not bolted on afterward.

---

Step 6: Pillar 4 — Education (Building AI Literacy)

Why Education Is a Pillar

Tools and processes only work if people understand them.

Uber's education strategy includes:

The Goal: Responsible-First Mindset

Every engineer, product manager, and business owner should ask: "Is this AI being built responsibly?" — not just "Does this AI work?"

---

Step 7: Pillar 5 — Adoption (Scaling Across the Company)

The Challenge

Building a governance system is one thing. Getting every team to use it is another.

Uber faced two populations:

• New models → Catch them early with shift-left governance
• Existing models → Bring them into compliance retroactively

Key Concepts for Scaling

1. Life-Cycle-Aware Governance

Governance adapts to where a model is in its life cycle:

Experimentation → Development → Staging → Production
     ↓                ↓             ↓           ↓
Light touch      More checks    Full review  Ongoing monitoring

2. Adaptive, Technology-Agnostic Controls

• Rules are dynamic, not static
• New AI techniques (e.g., new GenAI frameworks) are automatically accommodated
• No need to rewrite governance rules every time technology changes

3. Precision with Context

• Not every model carries the same risk
• Governance uses contextual signals to distinguish:
  • High-impact production models (need full review)
  • Experimental notebooks (lighter touch)

4. System-Level Visibility

• AI at Uber spans multiple platforms and codebases
• Governance operates at the system level, not just individual files

---

Step 8: Putting It All Together — The Complete Picture

Here is how all five pillars work as an integrated system:

ENGINEER BUILDS A MODEL
         ↓
[VISIBILITY] — Model Card created in Model Catalog
         ↓
[EXPLAINABILITY] — Feature importance automatically computed
         ↓
[GOVERNANCE] — Shift-left checks at PRD/ERD stage
               Gating mechanism before deployment
         ↓
[EDUCATION] — Engineer trained on responsible AI principles
         ↓
[ADOPTION] — Model brought into company-wide governance framework
         ↓
MODEL DEPLOYED RESPONSIBLY

---

Step 9: Key Takeaways

---

Quick Knowledge Check

Test yourself with these questions:

1. What problem does the Model Catalog solve? (Answer: Provides a centralized, searchable inventory of all AI models — visibility)

2. What is the difference between PFI and SHAP? (Answer: PFI gives global/overall feature importance; SHAP explains individual predictions)

3. What does "shift-left" mean in governance? (Answer: Moving compliance checks to the earliest stages of development, not at deployment)

4. Why is education considered a pillar of responsible AI? (Answer: Tools only work if people understand and adopt a responsible-first mindset)

5. What is the "glass box" approach? (Answer: Making AI systems transparent and interpretable, as opposed to opaque black boxes)

---

Final Thought: Responsible AI is not a one-time project. It is an ongoing commitment — embedding accountability into platforms, workflows, and culture so that innovation and trust can grow together.