Bias, Fairness & Ethics

Where Bias Comes From

1. Training Data Bias

Internet data over-represents certain groups and perspectives. If 80% of 'doctor' mentions use 'he', the model learns that association.

2. Labeling Bias

Human annotators bring their own biases to what they consider appropriate or high-quality.

3. Measurement Bias

Optimizing for overall accuracy may hide disparities across demographic groups.

Detecting Bias

Create stratified test sets with equal representation

Measure performance disparities across groups

Test counterfactual fairness (change names, pronouns)

Mitigation Strategies

Your Turn!

Design a bias testing protocol for a resume screening AI. What would you test and how?