Bias in Training Data
You'll learn how flawed or unrepresentative training datasets cause machine learning models to inherit and amplify historical discrimination.
The Foundation of Discriminatory Models
A machine learning model is only as fair as the data it learns from. When training datasets reflect historical inequalities, underrepresent certain groups, or are collected through biased processes, the resulting models perpetuate—and often worsen—those biases. This isn't a hypothetical problem; it's a documented pattern affecting hiring tools, credit scoring, facial recognition, and criminal justice systems.
Where Training Bias Comes From
Historical inequality: If you train a hiring model on 20 years of resume data from a company that predominantly hired men for engineering roles, the model learns that "good engineers" look like men. It didn't invent sexism—it learned it from reality.
Sampling problems: If your facial recognition training set contains 90% lighter-skinned faces, the model will perform poorly on darker-skinned individuals. The dataset doesn't represent the full population it will serve.
Labeling bias: Human annotators bring their own biases when labeling data. If reviewers consistently rate résumés with "foreign-sounding" names lower, that prejudice becomes encoded in training labels.
The Amplification Effect
Models don't just mirror bias—they can amplify it. A model optimizing for patterns will find correlations between protected characteristics and outcomes, then apply those patterns uniformly and at scale, affecting thousands or millions without the case-by-case judgment humans might apply.
Think of it like teaching someone history exclusively from biased textbooks—they'll learn and repeat those biases confidently, applying them consistently to every new situation.
Key Takeaway: Training data contaminated by historical inequality or sampling problems becomes the blueprint for discriminatory model behavior—addressing fairness requires auditing and correcting datasets before training begins.