Computer Science - Computation and Language

Vicarious Offense and Noise Audit of Offensive Speech Classifiers

Testing 9 different AI content moderation systems on 92 million YouTube comments reveals wildly inconsistent results, while human annotators show strong political bias—proving that …