Majority-vote labeling in hate speech datasets doesn't just lose nuance — it actively misleads models trained on the result.
Researchers studying HateXplain, a widely used hate speech benchmark, found that 42.6% of all annotator disagreements cluster at the boundary between hate and merely offensive content. That concentration is not random: a chi-squared test (135.199, df=2, p<0.0001) confirms annotators are applying different internal thresholds for where offense ends and hatred begins. Both a standard hard-label BERT model and a soft-label alternative dropped about 22 percentage points in accuracy when moving from posts annotators agreed on (~80% accuracy) to posts they disagreed on (~58%). A per-annotator multi-head model widened that gap to 28 points. Three downstream fixes — each more sophisticated than the last — all failed to close it.
The deeper problem is what happens to model confidence. The hard-label BERT model expressed significantly higher confidence on its boundary-case errors (0.710) than the multi-head model did (0.495, p<0.0001). That means standard evaluation metrics won't flag the failure — the model looks sure of itself precisely when it is most likely wrong. Content moderation systems built on these models inherit a kind of institutional false certainty.
Majority vote is the norm in annotation pipelines because it is cheap and produces clean labels. But the researchers argue the fix has to come before training, in how annotations are designed — not after, through model architecture. That's a harder sell to teams already sitting on labeled datasets, but the alternative is moderating contested speech with a classifier that doesn't know what it doesn't know.