Mod-Guide: An LLM-based Content Moderation Feedback System to Address Insensitive Speech toward Indigenous Ethnic and Religious Minority Communities

LLMs struggle to recognize culturally insensitive speech on their own. The paper reveals that standard large language models have "epistemic limits" in detecting language that marginalizes minority communities through implicit erasure, misrepresentation, or normative framing—rather than just overt hostility—particularly toward Bangladesh's Hindu and Chakma communities.

Community input dramatically improves AI moderation. By co-creating a corpus of insensitive speech directly with community members and integrating their lived experiences into the moderation system using retrieval augmented generation (RAG), the researchers found that AI responses became "more contextually accurate" and better reflected minority perspectives.

The same content is perceived differently across ethnic lines. Mixed-method evaluations showed that RAG-enhanced moderation responses were "perceived differently across ethnic lines," suggesting that what majority groups might miss as problematic, minority communities recognize as harmful—highlighting why their input is essential.

Current AI moderation risks amplifying existing inequities. Without culturally grounded approaches, LLM-based content moderation systems may inadvertently perpetuate marginalization of minority communities by failing to catch insensitive speech that erases or misrepresents their perspectives.

Restorative justice principles should guide AI design. The paper advocates for incorporating restorative justice approaches in content moderation systems, suggesting that AI tools should be designed not just to remove harmful content, but to center the voices and experiences of historically underrepresented communities.