ML models learn from data. If your data has bias, your model will too. Bias in, bias out.

How Bias Enters Models

Historical bias: Past data reflects past discrimination

Example: Hiring model trained on 90% male resumes learns to prefer male candidates.

Sampling bias: Training data doesn't represent everyone

Example: Facial recognition trained mostly on light-skinned faces fails on darker skin tones.

Label bias: Human labelers have biases

Example: "Professional photo" labels reflect labeler's cultural assumptions.

Real Consequences

• Loan applications denied based on zip code (proxy for race)
• Resume screening filters out qualified candidates
• Recidivism prediction unfairly targets minorities
• Healthcare algorithms under-allocate resources

What You Can Do

1. Audit your data:

• Check demographic representation
• Look for proxy variables (zip code, names)
• Question your historical data

2. Test for disparate impact:

• Measure outcomes by group
• Compare false positive/negative rates
• Use fairness metrics (demographic parity, equalized odds)

3. Design with fairness in mind:

• Remove sensitive attributes (race, gender) AND their proxies
• Balance training data
• Consider fairness constraints during training

The Hard Truth

Perfect fairness is impossible. Different fairness definitions contradict each other. But you must still try.