Religious bias landscape in language and text-to-image models: analysis, detection, and debiasing strategies

Ajwad Abrar, Nafisa Tabassum Oeshy, Mohsinul Kabir, Sophia Ananiadou

Abstract: The widespread adoption of language models highlights the need for critical examinations of their inherent biases, particularly concerning religion. This study systematically investigates religious bias in language and text-to-image generation models, analyzing open-source and closed-source systems. We curate a dataset of approximately 460 unique, naturally occurring prompts to evaluate religious bias in language models across diverse tasks, such as mask filling, prompt completion, and image generation. To identify biases in image generation, we produce 5,000 images using text-to-image (T2I) models and release them publicly for further classification and analysis. Our experiments reveal concerning instances of underlying stereotypes and biases associated disproportionately with certain religions. In addition, we explore cross-domain biases, examining how religious bias intersects with demographic factors such as gender, age, and nationality. This study further evaluates the effectiveness of targeted debiasing techniques, primarily through corrective prompts, while also exploring complementary model-level approaches. Our findings demonstrate that language models continue to exhibit significant biases in both text and image generation tasks. These findings advocate for the integration of the principles of equity, diversity, and inclusion (EDI) into the development of ethical AI, particularly in addressing biases within generative AI systems for global acceptability.

External IDs:dblp:journals/ais/AbrarOKA26