Hybrid Moderation in Practice: How Users, AI, and Professionals Judge Uncivil Comments

Online platforms increasingly rely on hybrid moderation systems that combine automated triage with professional human review, yet little research compares how these actors judge the same real content in production-linked settings. This article examines uncivil online comments as a platform-governance problem by comparing three decision-makers: registered users, a platform AI system, and professional moderators. Using a secondary dataset from a US-based publisher comment environment, we analyze real comments spanning five moderation categories: Violence, Doxxing, Hate Speech, Sexual Explicitness, and Personal Attacks. The dataset combines production-stage AI workflow dispositions, professional moderators’ final adjudications, and ex-post survey judgments from 863 registered users, yielding 8,588 user-comment decisions. Results show that users chose to allow publication of roughly two-thirds of the uncivil comments and that their decisions aligned more closely with the AI system’s publish-and-moderate versus hold-and-moderate orientation than with professional moderators’ publish-versus-block decisions. Divergence was concentrated in specific categories and items, especially Hate Speech. Users’ justifications suggest that permissive decisions often relied on threshold reasoning, rejection of toxicity, or free-speech concerns, whereas restrictive decisions more often relied on direct violation labels. Demographic and engagement-based predictors explained little variance in alignment. The findings suggest that users, AI systems, and professional moderators should not be treated as interchangeable moderation tools: they operationalize different thresholds and may be suited to different governance functions.