Quest 103 - AI Bias Auditor

Quest 103: AI Bias Auditor

medium 25-30 minutes

---

🎯 Learning Objectives

• ✅ Why measuring fairness across groups is essential for responsible AI
• ✅ How to compute demographic parity and equalized odds
• ✅ The difference between group-level fairness metrics and global averages
• ✅ How bias can hide behind seemingly 'fair' aggregate numbers

---

📖 Concept: AI Fairness Metrics

AI systems make decisions that affect real people — loan approvals, hiring recommendations, medical diagnoses. When these systems treat different demographic groups unequally, it’s called algorithmic bias. The problem is, bias can hide in plain sight. A system might have 90% overall accuracy, but if it’s 95% accurate for one group and only 75% for another, that’s a serious fairness issue.

Fairness metrics quantify these disparities. Two of the most important are:

• Demographic Parity: Do different groups get positive predictions at the same rate? If a hiring AI recommends 40% of Group A but only 15% of Group B, demographic parity is violated.
• Equalized Odds: Among people who actually qualify, does the AI correctly identify them at the same rate across groups? If the AI catches 90% of qualified candidates from Group A but only 60% from Group B, equalized odds is violated.

The key insight is that fairness must be measured between groups, not by averaging everything together. Averages mask disparity.

---

⚙️ How It Works

The Fairness Computation Pipeline

1. Collect predictions with group labels
         ↓
2. Group predictions by demographic group
         ↓
3. Compute per-group statistics:
   - Positive prediction rate
   - True positive rate
         ↓
4. Compare metrics across groups
         ↓
5. Report disparities (demographic parity gap, equalized odds gap)

Why Group-Level Metrics Matter

Consider a simple example with 10 predictions per group:

Group	Total	Predicted Positive	Actually Positive	True Positives
Group A	10	8	5	5
Group B	10	3	5	2

• Group A positive rate: 80% | True positive rate: 100%
• Group B positive rate: 30% | True positive rate: 40%
• Demographic parity gap: 80% - 30% = 0.50 (large disparity)
• Equalized odds gap: 100% - 40% = 0.60 (large disparity)

If you averaged everything together, you’d see “70% positive rate” and miss the fact that Group B is being systematically underserved.

---

💡 Example: Computing Fairness Metrics

Here’s how to compute fairness metrics in JavaScript:

Step 1: Group predictions by demographic

const groups = {};
for (const p of predictions) {
  if (!groups[p.group]) groups[p.group] = [];
  groups[p.group].push(p);
}

Step 2: Compute per-group statistics

for (const name of Object.keys(groups)) {
  const preds = groups[name];
  const total = preds.length;
  const positivePredicted = preds.filter(p => p.predicted).length;
  const actualPositive = preds.filter(p => p.actual);
  const truePositives = actualPositive.filter(p => p.predicted).length;

  groupStats[name] = {
    positiveRate: positivePredicted / total,
    truePositiveRate: actualPositive.length > 0
      ? truePositives / actualPositive.length
      : 0,
  };
}

Step 3: Compute the gap between groups

const positiveRates = groupNames.map(g => groupStats[g].positiveRate);
const demographicParity = Math.max(...positiveRates) - Math.min(...positiveRates);

const tprRates = groupNames.map(g => groupStats[g].truePositiveRate);
const equalizedOdds = Math.max(...tprRates) - Math.min(...tprRates);

A gap of 0 means perfect fairness. Larger gaps indicate greater disparity.

---

⚠️ Common Mistakes

Mistake 1: Averaging across all groups

“Overall accuracy is 85%, so the model is fair” → Averages mask group-level disparities. Always compute metrics per group first, then compare.

Mistake 2: Ignoring groups with small sample sizes

“Group C only has 2 predictions, so skip it” → Small groups matter too. Report the sample size alongside metrics so others can judge reliability.

Mistake 3: Confusing demographic parity with equalized odds

“They both measure fairness, so they’re the same” → They measure different things. Demographic parity checks prediction rates; equalized odds checks accuracy among qualified individuals. A model can satisfy one but violate the other.

Mistake 4: Rounding away the problem

“The gap is only 0.003, so it’s negligible” → Even small gaps can have large real-world impact when applied to thousands of decisions. Report exact values and let stakeholders decide the threshold.

---

📝 Knowledge Check

📝 Knowledge Check

Q1:What is demographic parity?

Different groups have equal true positive ratesDifferent groups receive positive predictions at the same rateThe overall model accuracy is above 90%Each group has the same number of samples

Q2:Why is computing a single global fairness metric misleading?

Global metrics are always wrongAverages mask disparities between different groupsGlobal metrics only work for binary classificationYou need at least 10 groups for a global metric

Q3:How do you compute the demographic parity gap?

Average the positive rates across all groupsSubtract the minimum positive rate from the maximum positive rateCount how many groups have exactly the same rateDivide the total positive predictions by total predictions

---

🏋️ Quest: AI Bias Auditor

Now it’s time to practice! Implement a fairness metrics calculator for AI predictions.

1. Download the starter files:

   npx bluebeltdojo download quest-103-bias-auditor
   cd quest-103-bias-auditor

2. Open problem.js in your editor with your AI tool

3. Implement computeFairnessMetrics(predictions) that:

   • Groups predictions by demographic group
   • Computes positive prediction rate and true positive rate per group
   • Calculates demographic parity gap and equalized odds gap

4. Run node test.js and read the failures carefully

5. Fix any edge cases the AI missed

6. Verify all tests pass:

   node test.js

7. When all tests pass, submit your solution:

   npx bluebeltdojo submit

💡 Tip: The key insight is that fairness must be computed BETWEEN groups, not by averaging. If your code only computes a single global metric, you’re missing the point.

---

คำใบ้

• อ่าน instructions ใน problem.js อย่างละเอียด
• Fairness metrics ต้องคำนวณระหว่างกลุ่ม (between groups) ไม่ใช่ค่าเฉลี่ยรวม
• Demographic parity = ความแตกต่างของ positive prediction rates
• Equalized odds = ความแตกต่างของ true positive rates
• ถ้าติดขัด ลองอ่าน “Common Mistakes” อีกครั้ง — อย่าดู solution โดยตรง