The Human Hand in Labeling

Raw data rarely arrives ready to train a model. Someone has to label it — decide what counts as "spam," "qualified," "appropriate," or "relevant." Every one of those labels is a judgment call made by a person (or a small team), and their assumptions become part of the data just as surely as the raw examples do.

Take a content-moderation dataset used to train a system that flags "inappropriate" posts. Two different labeling teams, working from the same guidelines, can genuinely disagree about borderline cases — slang, dialect, sarcasm, cultural context. Whichever set of calls ends up in the training data is what the model learns as the definition of "inappropriate," even though it was one specific, disputable interpretation.

This isn't the labelers doing something wrong on purpose — most labeling guidelines are written in good faith. But good faith doesn't guarantee that every labeler's read of an ambiguous case lines up with everyone else's, or that the guidelines themselves anticipated every kind of language and context they'd be applied to. Those small, human inconsistencies get absorbed into the data at scale.

Why do labeling choices introduce bias into training data even when labelers act in good faith?

Someone has to ____ raw data — decide what counts as "spam" or "appropriate" — before a model can learn from it.

Labeling guidelines applied in good faith always produce perfectly consistent labels with no room for disagreement.