Label Bias and Annotator Disagreement

Supervised models learn from labels, but labels may be judgments rather than observations of an objective fact. Label bias occurs when annotation practices, instructions, institutional categories, or proxy outcomes encode systematic viewpoints. In subjective tasks, disagreement may reflect real ambiguity or different social perspectives, not merely careless annotators. If a team collapses all judgments into one “gold” label, it can erase that variation and make one group’s interpretation appear universal. Research by Sap and colleagues on hate-speech detection found associations between annotator identity or beliefs and toxicity ratings, including for African American English. Their earlier study found that models trained on several widely used datasets could label AAE tweets and posts by self-identified Black users as offensive at elevated rates. They also found that making annotators aware of dialect context could reduce offensive labels in that study. These findings are specific to datasets and tasks; they do not mean every annotator shares the same bias or every mention of a dialect is mislabeled. Labels may also be institutional proxies rather than annotator judgments. Obermeyer and colleagues showed that using health-care cost as a proxy for health need disadvantaged Black patients because comparable illness had historically generated lower spending. That issue is often described as measurement or target bias, not simply annotation error, but it illustrates why teams must ask what a label actually represents. Improve labeling by defining the construct, providing context-sensitive instructions, recruiting appropriately diverse annotators, and preserving multiple judgments where disagreement is meaningful. Audit disagreement by subgroup and example type. A majority vote can be useful for some operational tasks, but it should not silently erase ambiguity. Decide whether the task needs consensus, a distribution of views, escalation, or abstention, and document how labels were produced.