Understanding Bias in Machine Learning

Bias starts in the data

Models learn the patterns — and prejudices — present in their training data. If history was biased, an unchecked model will be too.

Where it enters

• Sampling — groups under-represented in the data.
• Labels — subjective or historically skewed annotations.
• Proxies — features that stand in for protected attributes.

What to do

• Evaluate per group, not just overall accuracy.
• Audit features for hidden proxies.
• Involve affected people in defining fairness for the use case.

Fairness isn't one metric — it's a deliberate, ongoing practice. Measure across subgroups before you ship.