Advanced Bias Lab — Design datasets, train models, audit fairness, and fix the AI.
1. Design
2. Train
3. Audit
4. Test
5. Fix
6. Report
🎛️ Design Your Training Dataset
Choose a scenario and configure the bias level in your training data. Higher bias = more skewed approvals toward one group.
This simulates how real-world historical data often contains hidden discrimination.
⚙️ Bias Configuration
Group A Approval Rate80%
Group B Approval Rate30%
Training Data Size16
Noise / RandomnessLow
⚠️ Bias Gap: 50 percentage points
This gap will teach the AI to favor Group A. Real-world example:
Mortgage approval rates in 2019 were 80% for white applicants vs. 66% for Black applicants (Home Mortgage Disclosure Act data).
📝 Training Phase
1/16
Labeled
Review each profile. Click Approve or Reject.
Your labels will become the AI's "ground truth." Pay attention to patterns — what are you rewarding?
🔍 Model Audit
The AI has learned from your labels. Let's inspect what it learned — feature weights, decision boundaries, and potential bias signals.
Feature Weights (What the AI Values)
Training Data Breakdown
🧠 How the AI Decides:
# Simplified decision function the AI learned defpredict(person):
score = return"approve"if score > 0else"reject"
🧪 Fairness Testing
1/12
Tested
The AI now predicts on new, unseen people. For each, predict what the AI will do, then see if it matches your fairness intuition.
Watch for false positives and
false negatives.
Your prediction: What will the AI decide?
🤖 AI Decision:
🔧 Fix the AI
The AI is biased. Now it is your job to fix it. You have three tools:
Reweight (adjust feature importance), Resample (add balanced data),
or Constrain (force equal treatment). Try different approaches and see what works.
⚖️
Reweight Features
Reduce the weight of biased features (neighborhood, referral, insurance). Increase weight of merit features.
🔄
Resample Data
Add more balanced training examples. This simulates collecting fairer historical data.
Balanced examples to add4
🔒
Constrain Decisions
Force the AI to approve/reject at equal rates across groups, regardless of other features.
