Whose harm gets detected? A structured review and conceptual framework for misogyny detection and Dari–Pashto marginalization in AI content moderation

Women in Dari and Pashto digital spaces face gendered abuse expressed through proverbs, religious framing, and moral instruction forms that existing AI content moderation systems often fail to detect. This paper presents a theory-driven structured review of misogyny detection research, complemented by a structured empirical mapping of representative studies. Drawing on a corpus of 84 publications published between 2015 and 2025, we examine how misogyny is operationalized across four dimensions of the NLP moderation pipeline: linguistic surface modeling, cultural-norm encoding, annotation epistemology, and computational adaptation. Our analysis reveals a pronounced structural imbalance. Across a coded subset of representative studies ( n = 20), all focused on linguistic surface modeling and computational adaptation, while only a small minority addressed annotation epistemology, and none explicitly operationalized culturally grounded norm systems as a dedicated modeling layer. These dimensions are not mutually exclusive. Across datasets and benchmarks, recurring limitations include overreliance on explicit lexical toxicity, Western-centric resource concentration, limited robustness under domain shift, and the absence of misogyny-specific resources for Dari and Pashto. We argue that these limitations are systemic rather than purely architectural. Framing misogyny detection as a sociotechnical moderation pipeline, we introduce the Culturally Grounded Misogyny Detection Framework (CG-MDF) and outline design principles for culturally grounded dataset development, evaluation, and human-in-the-loop moderation, with particular emphasis on establishing a research foundation for Afghan languages.