Auditing and Redesigning LLM Moderation Guardrails for Cultural and Genre Sensitivity

Mahomed et al. (2024) uncover how LLM moderation endpoints (like those in ChatGPT) often flag scripts based on genre and maturity ratings, resulting in over-censorship of certain cultural products. However, there’s little research on how to systematically audit and then redesign these guardrails to be more culturally and contextually sensitive. This project would conduct large-scale, cross-cultural audits of LLM moderation outcomes, identify patterns of bias, and then involve users and creators in co-design workshops to prototype new moderation criteria, perhaps leveraging federated or community-curated “cultural exemption” lists. The work is novel in bridging algorithm audits with participatory policy design, directly tackling the challenge of algorithmic cultural imperialism and the self-censorship of creative expression.

References:

1. Auditing GPT's Content Moderation Guardrails: Can ChatGPT Write Your Favorite TV Show?. Yaaseen Mahomed, Charlie M. Crawford, Sanjana Gautam, Sorelle A. Friedler, Danaë Metaxa (2024). Conference on Fairness, Accountability and Transparency.
2. Auditing GPT's Content Moderation Guardrails: Can ChatGPT Write Your Favorite TV Show?. Yaaseen Mahomed, Charlie M. Crawford, Sanjana Gautam, Sorelle A. Friedler, Danaë Metaxa (2024). Conference on Fairness, Accountability and Transparency.
3. Auditing GPT's Content Moderation Guardrails: Can ChatGPT Write Your Favorite TV Show?. Yaaseen Mahomed, Charlie M. Crawford, Sanjana Gautam, Sorelle A. Friedler, Danaë Metaxa (2024). Conference on Fairness, Accountability and Transparency.