FAIRNESS AND BIAS IN GENERATIVE AI

Abstract

The rapid advancement of generative AI technologie particularly large language models (LLMs) and Text-To-Image (T2I) system has brought with it a growing concern about embedded socio cultural biases , especially in complex , multicultural societies like in India. This thesis investigates how these models engage with Indian social realities, focusing on three interrelated dimensions: cultural traditions, caste representation, and gender roles. We analyze caste-based representational biases in publicly available T2I models , studying how these systems portray caste minorities compared to dominant castes. Using an LLM as an evaluator, we assess both textual prompts and generated visuals to uncover implicit biases against the minorities. We extend this approach to explore gender bias in occupational imagery, focusing on how generative systems depict professional roles in the Indian business context. Our findings reveal that these models often reinforce traditional gender stereotypes, underrepresenting women in various specialized business domains. Together, these studies highlight the challenges of building culturally inclusive AI systems and offer critical insights into how generative models can unintentionally replicate and in some cases amplify existing social inequalities . This thesis contributes to the growing field of AI ethics by foregrounding the importance of contextual sensitivity, cultural pluralism, and fair representation in the design and evaluation of Generative AI systems.