Participatory, Dialect- and Gender-Aware Guardrails via On-Device Adapters

Run participatory policy-redrafting workshops with trans users, dialect communities, and creators to co-specify nuanced, testable policy criteria. Convert these into policy-as-prompt test suites and distill them into parameter-efficient guardrails using LoRA-Guard for on-device enforcement. Layer in calibration for reliable uncertainty and appeals. This approach embeds participatory redrafting into the model artifact itself, enabling community-specific adapters without degrading generation quality. On-device deployment supports localized norms and privacy, closing the loop from participatory drafting to executable guardrails. It moves beyond revising policy language by co-producing executable, bias-sensitive guardrails aligned with creator-focused governance and addressing governance challenges in policy-as-prompt systems. Community-tailored adapters can reduce false positives on trans bodies or dialectal expressions while maintaining safety. Impact includes a replicable blueprint for inclusive moderation that is technically deployable, accountable, and socially legitimate.

References:

1. Policy-as-Prompt: Rethinking Content Moderation in the Age of Large Language Models. Konstantina Palla, Jos'e Luis Redondo Garc'ia, Claudia Hauff, Francesco Fabbri, Andreas Damianou, Henrik Lindström, Dan Taber, M. Lalmas (2025). Conference on Fairness, Accountability and Transparency.
2. LoRA-Guard: Parameter-Efficient Guardrail Adaptation for Content Moderation of Large Language Models. Hayder Elesedy, Pedro M. Esperancca, Silviu Vlad Oprea, Mete Ozay (2024). Conference on Empirical Methods in Natural Language Processing.
3. On Calibration of LLM-based Guard Models for Reliable Content Moderation. Hongfu Liu, Hengguan Huang, Hao Wang, Xiangming Gu, Ye Wang (2024). International Conference on Learning Representations.
4. Misgendered During Moderation: How Transgender Bodies Make Visible Cisnormative Content Moderation Policies and Enforcement in a Meta Oversight Board Case. Samuel Mayworm, Kendra Albert, Oliver L. Haimson (2024). Conference on Fairness, Accountability and Transparency.
5. Conceptualizing and Improving Creator Moderation Design with Platform Stakeholders. Renkai Ma (2023). CSCW Companion.