Keywords
religious bias, fairness in AI, LLM evaluation
Abstract
In the earlier years of development of LLMs, it was relatively easy to prompt an LLM to respond with toxic or biased statements about religion. Subsequent improvements in frontier models addressed many of the issues of bias and toxicity in general, including against religion. At the same time, the adoption and usage of these models has grown exponentially. Small and implicit biases, therefore, have a magnified overall impact. In this paper, we (1) briefly review previous efforts to measure religious bias in LLMs, (2) show, by reviewing over 12,000 papers dealing with bias in LLMs, that religious bias has been significantly understudied compared to other bias targets, and (3) highlight areas where further research is needed to understand gaps in model performance for a growing global community of LLM users.
Original Publication Citation
https://www.overleaf.com/project/6a0c661f6ee207e406c87340
BYU ScholarsArchive Citation
Carty, Sheryl; Fulda, Nancy; and Reade, Walter, "Religious Bias in LLMs is Significantly Understudied" (2026). Faculty Publications. 9613.
https://scholarsarchive.byu.edu/facpub/9613
Document Type
Peer-Reviewed Article
Publication Date
2026-07-08
College
Computational, Mathematical and Physical Sciences
Department
Computer Science
Copyright Use Information
https://lib.byu.edu/about/copyright/