Uncovering Theological and Ethical Biases in LLMs

Abstract

This paper explores theological and ethical biases in LLMs through a novel approach involving creative text generation tasks based on biblical texts, specifically the Ten Commandments and the Book of Jonah. Utilizing models such as GPT-4 Turbo, Claude v2, PaLM 2 Chat, Llama 2 70B, and Zephyr 7B, the study employs a combination of qualitative hermeneutical analysis and quantitative textual analysis. Findings reveal a prevalent progressive bias in these models, evident in their interpretations of foundational ethical guidelines and narrative texts. This bias aligns with contemporary socio-political and environmental concerns, especially in themes of environmental ethics, social justice, and inclusivity. In the narrative task involving the Book of Jonah, a dominant interpretive trend is observed, reflecting the models' tendency to mirror historical and prevailing interpretations. This study highlights the need for multidisciplinary research into LLMs' biases, particularly their impact on religious and ethical narrative interpretation and broader societal implications.