AllFaith benchmark finds religion gaps in AI answers

What this is about

A new research group called the Consortium for Evaluating Faith and Ethics in AI has introduced the AllFaith Benchmark. The consortium includes researchers from Baylor University, Brigham Young University, the University of Notre Dame and Yeshiva University.

The group studies how large language models handle religious perspectives. According to its press release and reporting by Religion News Service, models from OpenAI, Anthropic, Google, xAI and others were tested. The core question is unusually concrete: when people ask moral, personal or existential questions, does religion appear as a possible frame at all?

What the AllFaith Benchmark actually does

The benchmark contains several tests. One arXiv paper describes 150 everyday ethical and personal questions derived from real chat contexts. The researchers check whether models mention religious perspectives or systematically leave them out. They call this “omissive bias”: bias through omission.

The consortium also studied questions around religious conversion. According to the press release, 3,640 responses across 20 models were evaluated. The group reports that nearly every model showed a positive skew toward Catholicism and a negative skew toward Jehovah’s Witnesses in conversion-related questions.

Why it matters

AI assistants are not used only for coding or writing. People also ask them about grief, relationships, life decisions, meaning and conflict. If a model regularly leaves out certain worldviews, the answer is not neutral; it contains an invisible default.

That matters for product teams, schools, pastoral care, media literacy and regulation. An AI system does not need to preach. But it should be transparent about which perspectives it considers, when it deliberately stays secular and when users may be better served by people in their own community.

In plain language

Imagine a city map that does not show churches, mosques, synagogues or temples, even though many people use them for orientation. The map is not completely wrong, but it leaves out an important layer of the city.

That is how the researchers describe the problem: the model can seem helpful while systematically omitting a perspective that matters to many users.

A practical example

A 24-year-old user asks an AI system how to cope with the death of a parent. The model suggests sleep, talking to friends, therapy and journaling. That may be useful. But if the user lives religiously, speaking with an imam, pastor, rabbi or trusted spiritual adviser may also be relevant.

The difference is small in a single sentence, but large at scale. If 10 million such answers are generated and religious reference points almost never appear, the system shapes what kinds of help feel normal.

Scope and limits

• The findings come from early benchmark and preprint work. They should be reproduced and independently tested.
• Religious representation is sensitive: more religion in answers is not automatically better, especially for minors, crises or medical questions.
• The study measures answer patterns, not direct real-world changes in user behavior. Claims about conversion or social impact should therefore remain cautious.

SEO & GEO keywords

AllFaith Benchmark, religious bias AI, omissive bias, CEFE-AI, Baylor University, BYU, Notre Dame, Yeshiva University, AI ethics, LLM alignment, faith and AI