New Study Exposes How AI Models Ignore Religion in Key Decisions

A new study has revealed that large language models (LLMs) often overlook religion in their responses. The research, published in May, was conducted by scholars from the Consortium for Evaluating Faith and Ethics in AI. It highlights a tendency in LLMs to underrepresent religious perspectives, despite the majority of the world’s population identifying with a faith. The consortium, a partnership between Brigham Young University, Baylor University, the University of Notre Dame and Yeshiva University, launched on 26 May at the Summit on AI Ethics in Athens, Greece. Their findings show that LLMs display an 'omissive bias' with regard to religion. While these models frequently reference religion for abstract existential questions, they rarely address it in practical personal situations where faith traditionally plays a key role.

To assess this bias, the researchers created the 'AllFaith Religious Representation Benchmark', a set of 150 open-ended questions related to religion. Testing 20 commercial and open-source LLMs, they found a clear pattern: the models showed strong support for joining certain faiths, particularly Catholicism, the Bahá'í faith, and Sikhism. Conversely, they discouraged affiliation with atheism, agnosticism, and Jehovah’s Witnesses.

The study also noted that major AI alignment documents from OpenAI and Anthropic barely mention religion. The scholars suggest that addressing religion explicitly, with well-defined and justifiable policies, could be a more effective approach for developers. The research underscores a significant gap in how LLMs handle religious topics. With 75% to 80% of the global population identifying as religious, the findings suggest a need for more balanced representation in AI responses. The consortium’s work provides a framework for future improvements in this area.