Knowledge MaxxingLabel bias
Label bias occurs when the 'correct answers' used to train a model themselves reflect human prejudice or inconsistent judgment, rather than an objective ground truth.
Prerequisite: Sampling bias
What matters
Because a model learns to reproduce its labels, if the labels themselves were assigned unfairly, the model can learn to systematically replicate that same unfairness.
Example
If past hiring decisions used as training labels reflected biased human judgments, a model trained on them can learn to reproduce those same biased outcomes.
Common misconception
Label bias is not fixed simply by having more labelers; if the labeling process or criteria themselves are systematically biased, more labelers can still reproduce the same bias.
Recall question
Why can a model trained on biased human-assigned labels reproduce that same bias?