Mod-Guide uses RAG to improve LLM sensitivity to indigenous minorities

A team of researchers led by Dipto Das has introduced Mod-Guide, a content moderation feedback system that addresses insensitive speech toward indigenous ethnic and religious minorities. The work, published on arXiv (2606.13397) and presented at a major HCI venue, focuses on two marginalized groups in Bangladesh: the Hindu community (the country's largest religious minority) and the Chakma community (the largest Indigenous ethnic minority). The researchers argue that large language models (LLMs) often fail to recognize culturally insensitive speech—language that marginalizes communities through implicit erasure, misrepresentation, or normative framing rather than overt hostility. To solve this, they co-created a culturally grounded corpus of insensitive speech with community members and integrated these narratives into LLM moderation pipelines using Retrieval Augmented Generation (RAG). RAG allows the system to pull contextual cues from lived experiences, making the LLM more sensitive to minority viewpoints.

Through mixed-method evaluations involving both minority and majority participants, the team found that RAG-enhanced moderation responses were more contextually accurate and were perceived differently across ethnic lines. The study advances human-computer interaction, AI ethics, and social computing by foregrounding restorative justice and hermeneutical inclusion in content moderation design. Mod-Guide demonstrates how minority communities can actively shape moderation tools rather than being passive subjects of top-down AI systems. The researchers emphasize that ignoring cultural context in LLM-based moderation risks perpetuating the very harms these tools aim to reduce. This work provides a practical framework for building more inclusive AI moderation systems that respect the epistemic contributions of historically underrepresented groups.