Are LLMs Trustworthy Psychometric Hate Speech Judges?

Content moderation demands trustworthy, interpretable annotation at scale, yet a fundamental question remains unanswered: what kind of psychometric raters are LLMs? We apply the Many-Facet Rasch Model (MFRM) to evaluate five frontier LLMs (GPT-5, GPT-5-mini, Claude, Gemini, and DeepSeek) as hate speech annotators across ten dimensions, benchmarked against 582 human raters from the Measuring Hate Speech (MHS) corpus. Our findings reveal that most LLMs are systematically more lenient than human annotators but more consistent as a group. LLM annotation bias is also severity-dependent, thus concentrating annotation error where early-stage moderation matters most. Modular prompting induces greater leniency than monolithic while preserving rank-order agreement, and the non-Western model (DeepSeek) shows disproportionate sensitivity to prompting strategy. Moreover, the three-class hate speech dimension on which humans show highest inter-rater reliability (IRR) produces the greatest cross-model LLM disagreement, thus motivating a recalibration framework to align LLM-derived scores with human measurement standards.