Annotator disagreement has often been treated as a nuisance in content moderation datasets, something to be smoothed away by majority vote. A new paper on arXiv argues the opposite: contested items carry useful information about the subjectivity of harmful content. By decomposing disagreement into its sources, the authors propose a way to predict which items are likely to divide annotators.

The practical payoff is in annotation effort. If disagreement can be anticipated, teams can route contested items to more diverse or additional annotators, or use the predicted contestation as a signal in model training. The abstract emphasizes that this is not just a single-language phenomenon; the work explicitly addresses harmful content across languages.

Because the source is only an abstract, details on the decomposition method and evaluation are not available here. Still, the framing aligns with a broader shift in NLP toward treating disagreement as data rather than error. The authors stop short of claiming a full solution, but the direction is clear: disagreement is not a bug, it is a feature worth predicting.