Quest 53 - Bias Detector

Quest 53: Bias Detector

medium 25-30 minutes

---

🎯 Learning Objectives

• ✅ How to detect bias in AI outputs using the 4/5ths rule (disparate impact ratio)
• ✅ Why measuring rates matters more than measuring counts
• ✅ How to calculate per-group fairness metrics
• ✅ The critical edge case: different group sizes masking bias

---

📖 Concept: AI Fairness & Bias Detection

AI models can produce biased outputs — and the bias is often invisible unless you actively measure it. A model might approve 90% of applicants from Group A but only 50% from Group B. If you only look at total counts, this disparity gets hidden behind group size differences.

Fairness in AI is not just an ethical ideal — it’s a measurable property. The most widely used metric is the disparate impact ratio (also called the 4/5ths rule), borrowed from US employment law. If one group’s positive outcome rate is less than 80% of another group’s rate, that’s a signal of potential bias.

Think of bias detection like a dojo instructor watching your form. You might think your technique looks perfect, but the instructor measures the angles and catches the 15-degree deviation you can’t feel. Bias detection tools do the same thing — they measure the numbers your intuition misses.

---

⚙️ How It Works

The 4/5ths Rule (Disparate Impact Ratio)

Disparate Impact Ratio = min(group_rates) / max(group_rates)

If this ratio is < 0.8 (80%), the system is flagged as potentially biased.

Step-by-step calculation

Step 1: Calculate the approval rate per group

// Group A: 10 people, 9 approved → rate = 0.9
// Group B: 10 people, 5 approved → rate = 0.5
const rateA = approvedA / totalA;  // 0.9
const rateB = approvedB / totalB;  // 0.5

Step 2: Compute the disparate impact ratio

const ratio = Math.min(rateA, rateB) / Math.max(rateA, rateB);
// 0.5 / 0.9 = 0.555 → biased (below 0.8)

Step 3: Flag groups below threshold

const flagged = groups.filter(g => rates[g] < threshold);

---

💡 Example: Different Group Sizes (The Edge Case)

The most dangerous naive-AI mistake: checking raw counts instead of rates.

// ❌ NAIVE: Checks if "approved count" is similar
const approvedA = 9;  // Group A: 9 approved
const approvedB = 1;  // Group B: 1 approved
// Naive: "9 vs 1 — that's biased!"

// ✅ CORRECT: Checks approval RATES
// Group A: 9 out of 10 (90%)
// Group B: 1 out of 2 (50%)
const rateA = 9 / 10;  // 0.9
const rateB = 1 / 2;   // 0.5
const ratio = 0.5 / 0.9;  // 0.555 → flagged

Why rates matter: If Group A has 1,000 people and Group B has 10, comparing raw counts will always make Group B look “underrepresented” — even if both groups have the same approval rate. Rates normalize for group size.

---

⚠️ Common Mistakes

Mistake 1: Comparing raw counts instead of rates

“Group A has more approvals than Group B — that’s bias!” → Always normalize by group size. A larger group will naturally have more approvals.

Mistake 2: Only checking one direction

“What if Group A is favored over Group B?” → Bias can flow both ways. Check all group pairs.

Mistake 3: Ignoring small sample sizes

“Group C has 1 person who was denied — bias!” → Small samples are noisy. A single denial in a group of 1 doesn’t prove bias. Consider statistical significance.

Mistake 4: Using counts for the disparate impact ratio

“The ratio of 9 approved to 1 approved is 0.11 — that’s biased!” → The ratio must use rates (approved/total per group), not raw counts.

---

📝 Knowledge Check

📝 Knowledge Check

Q1:What is the 4/5ths rule (disparate impact ratio) used for?

Measuring the total number of approved vs denied outputsComparing approval rates between demographic groups to detect biasCounting how many times each group appears in the datasetCalculating the average response time of the AI model

Q2:Why is comparing raw counts (instead of rates) a critical mistake in bias detection?

Raw counts are slower to compute than ratesA larger group will naturally have more approvals even with the same rate, masking biasRaw counts don't include demographic informationCounts are less precise than percentages

Q3:Group A: 9 approved out of 10 (90%). Group B: 1 approved out of 2 (50%). What is the disparate impact ratio?

0.11 (1 divided by 9)0.555 (0.5 divided by 0.9)0.9 (the higher rate)0.5 (the lower rate)

---

🏋️ Quest: Bias Detector

Now it’s time to practice! Build a bias detection system using fairness metrics.

1. Download the starter files:

   npx bluebeltdojo download quest-53-bias-detector
   cd quest-53-bias-detector

2. Open problem.js in your editor with your AI tool (Copilot, Claude Code, Cursor, etc.)

3. Implement detectBias(results, demographics) that uses the 4/5ths rule to detect bias

4. Important: Pay special attention to the different-group-sizes edge case — naive AI compares counts instead of rates

5. Verify all tests pass:

   node test.js

6. When all tests pass, submit your solution:

   npx bluebeltdojo submit

💡 Tip: The hardest test case has groups of different sizes. Make sure you calculate rates (approved/total per group), not raw counts.

---

คำใบ้

• อ่าน instructions ใน problem.js อย่างละเอียด
• คำนวณ approval rate ต่อกลุ่ม (จำนวน approved / จำนวนทั้งหมด) ไม่ใช่จำนวน raw
• Edge case ที่สำคัญที่สุด: กลุ่มที่มีขนาดต่างกัน — ต้องใช้ rate เทียบกัน
• ถ้าติดขัด ลองอ่าน “Common Mistakes” อีกครั้ง — อย่าดู solution โดยตรง