New Study Exposes AI's Blind Spot on Global Religious Identity

A new study has revealed that large language models (LLMs) often underrepresent religion in their responses. The research, published in May, shows a clear gap between global religious identity and how AI systems discuss faith. Between 75% and 80% of people worldwide identify with a religion, yet LLMs frequently overlook or minimise its role. The findings come from the Consortium for Evaluating Faith and Ethics in AI, a group of scholars from Brigham Young University, Baylor University, the University of Notre Dame, and Yeshiva University. They introduced the 'AllFaith Religious Representation Benchmark' to test 20 commercial and open-source LLMs. The benchmark used 150 open-ended questions to measure bias in how these models handle religious topics.

The study found that LLMs tend to favour certain faiths, such as Catholicism, Baháʼí, and Sikhism. At the same time, they often discourage affiliation with atheism, agnosticism, and Jehovah’s Witnesses. The models also showed an 'omissive bias', linking religion more to abstract existential questions than to practical, personal situations.

The consortium officially launched on 26 May during the Summit on AI Ethics in Athens, Greece. Their work highlights that major AI alignment documents from companies like OpenAI and Anthropic barely mention religion at all. The researchers suggest that AI developers should adopt explicit, well-defined policies for handling religion. This approach could improve alignment and ensure fairer representation. Without such steps, the gap between real-world religious identity and AI responses may persist.