Prompted Publics: Generative AI, Agentic Auditing, and Representational Bias in Indian Social Media Literature

Aditi Kuila, Suman Roy

Keywords: Generative AI, Computational Journalism, Representational Bias, Multilingual NLP, Agentic Auditing, Digital Humanities

TL;DR: Prompts are ideological acts: commercial LLMs erase 75% of caste markers from Indian social media; sovereign models with persona-based prompting double retention, demanding auditable AI in journalism.

Abstract:

Generative AI and Large Language Models (LLMs) are increasingly deployed in computational journalism and digital humanities, yet their application to Global South literatures raises critical concerns about representational bias and algorithmic erasure. This paper introduces the Prompted Publics framework, which reconceptualizes prompts, system instructions, and adversarial jailbreaks not as neutral technical inputs but as ideological interpretive acts that shape how marginalized digital publics are rendered legible by AI systems.

We operationalize this framework through a mixed-method pipeline that combines Multilingual Retrieval-Augmented Generation (mRAG) with Agentic Auditing protocols, evaluated on a curated multilingual corpus of Indian social media literature spanning Hindi, Bengali, Hinglish, and English. Our experiments demonstrate that commercial LLMs exhibit severe caste-marker erasure under default prompting (retention rates as low as 25.0%), while sovereign Indian models combined with positional prompting significantly improve retention (up to 50.9%).

Topic modeling identifies four coherent thematic clusters capturing caste resistance, intersectional feminism, AI fairness discourse, and vernacular literary movements. These findings underscore the urgent need for culturally grounded auditing frameworks in AI-assisted journalism.

Submission Number: 7