Course contentsShow
← AI Engineering
Lesson 1556 of 1,88638. Bias, Fairness, and AlignmentFree lesson

Sources of Bias: Training Data

Understand how historical bias, sampling bias, and label bias in training data propagate into model behavior.

Sources of Bias: Training Data

What You'll Learn

You'll learn how historical bias, sampling bias, and label bias embedded in training data automatically propagate into model predictions and behavior.

The Core Problem

Even a perfectly designed model learns only what its training data teaches. If that data contains biases—and it almost always does—the model will encode and amplify them. Three major sources lurk in training datasets:

Historical Bias

This reflects real-world inequalities captured in historical records. For example, if you train a hiring model on 20 years of company data where 90% of engineers hired were male, the model learns to associate "engineer" with "male." The data accurately reflects history—but that history was biased.

Analogy: Teaching a child solely from 1950s textbooks would pass along outdated gender roles, even if you quoted the textbooks perfectly.

Sampling Bias

This occurs when your training data doesn't represent the full population the model will serve. If you train a medical diagnosis model primarily on data from urban hospitals, it may perform poorly for rural patients with different demographics or health patterns.

Label Bias

Human labelers bring their own biases when annotating data. One annotator might label assertive language from women as "aggressive" while labeling identical language from men as "confident." These subjective judgments become ground truth the model optimizes for.

Why This Matters

Models don't "know" they're learning bias—they optimize to match patterns in the data. Without careful data auditing (which you studied in concepts 1555 and earlier), biased training data creates biased AI systems that harm underrepresented groups and violate fairness principles.

Key Takeaway: Historical bias, sampling bias, and label bias in training data automatically become model behavior—you cannot build fair AI without addressing data quality at the source.