Mod-Guide Applies LLM RAG Feedback to Moderation

For content-moderation teams serving underrepresented communities, a new arXiv paper offers a concrete method rather than a general call for more inclusive data: Mod-Guide (arXiv:2606.13397, Dipto Das et al., submitted June 11, 2026) integrates a community co-created corpus into a retrieval-augmented generation (RAG) pipeline, built and evaluated with members of Bangladesh's Hindu and Chakma communities - the country's largest religious and Indigenous ethnic minorities. The authors report that RAG-enhanced moderation responses were rated more contextually accurate, and were perceived differently across ethnic lines, in mixed-method evaluations with minority and majority participants.

Per the paper, the authors co-created a culturally grounded corpus of insensitive-speech examples directly with community members, then used it as retrieval context so the base LLM can surface culturally specific interpretations of implicit or normative harms it might otherwise miss. This is a concrete instance of a broader pattern: RAG is increasingly used to inject external, curated context into LLM outputs to improve factuality and domain sensitivity, but that benefit is bounded by retrieval quality, corpus representativeness, and prompt design.

Human-computer-interaction and AI-ethics research increasingly calls for participatory data practices when models affect marginalized communities. Co-creating datasets with impacted groups can raise model relevance and community trust, but it also raises open questions about scope, governance, and long-term maintenance of culturally specific corpora - questions this paper does not resolve.

Whether the authors release the co-created corpus or evaluation data, how the RAG pipeline's reproducibility holds up under peer review, and whether the approach generalizes beyond the Bangladesh Hindu and Chakma context studied here.