Model-Dependent Moderation: Inconsistencies in Hate Speech Detection Across LLM-based Systems

This paper finds that different LLMs result in widely divergent results when applied to hate speech detection, and determining they are more effective at preventing hate speech towards protected classes than other groups.