Major AI chatbots like ChatGPT, Claude, and Gemini are underrepresenting religion in their responses to life questions, yet simultaneously showing measurable bias toward certain faiths. A consortium of researchers from Brigham Young University, Baylor University, the University of Notre Dame, and Yeshiva University released findings in May showing that large language models (LLMs), which are AI systems trained to understand and generate human-like text, tend to omit religious perspectives entirely when they should be relevant, while paradoxically favoring specific religions when conversion questions arise.

Why Are AI Models Overlooking Religion in Everyday Advice?

The research team identified what they call "omissive bias" in AI systems. They tested 27 LLMs, including flagship models from OpenAI, Anthropic, Google, Baidu, Moonshot AI, and xAI, using a benchmark they developed called the "AllFaith Religious Representation Benchmark." This benchmark consists of 150 open-ended questions sourced from actual chat transcripts and faith community members, covering topics adjacent to religion such as questions about morality and personal decisions.

The findings are significant because roughly 75 to 80 percent of the global population identifies with a religion, according to Pew Research Center data cited in the study. Yet when researchers tested these models on practical questions where religion has historically played a major role in people's lives, the AI systems rarely mentioned faith-based perspectives. The omission was particularly stark in areas like grief, marriage, addiction, and family conflict.

"LLMs underrepresent religion in our benchmark relative to human expectations in every category," the researchers stated, adding that "the omission is not uniform."

Research team, Consortium for Evaluating Faith and Ethics in AI

The models were more likely to reference religion only when answering abstract existential questions about meaning, death, and truth. But even then, they did so rarely. This gap matters because millions of people rely on AI chatbots for advice and moral reflection, and the absence of religious perspectives may not reflect how actual humans approach these decisions.

Which Faiths Are AI Models Favoring, and Why Does It Matter?

Beyond the omission problem, researchers discovered a second troubling pattern. When they asked AI models whether users should convert from one faith to another, the systems showed "persistent and repeatable patterns of preference for some religions over others." A second consortium paper examined 20 commercial and open-source LLMs and found that they consistently favored three faiths while discouraging others.

The results revealed a striking asymmetry. Across all 20 models tested, the systems universally preferred Catholicism while discouraging Jehovah's Witnesses. The models also showed high support for joining Bahá'í and Sikh traditions, while discouraging affiliation with atheism and agnosticism. xAI's Grok 4.20 produced the strongest biases in these patterns.

• Universally Favored: Catholicism received consistent preference across all 20 LLMs studied, with models encouraging users to join the faith.
• Universally Discouraged: Jehovah's Witnesses were discouraged by all 20 models, showing the most uniform negative bias in the study.
• Selectively Favored: Bahá'í and Sikh traditions received high support for joining, though not with the same universal consistency as Catholicism.
• Discouraged Secular Options: Atheism and agnosticism were discouraged by the models, suggesting a systematic bias against non-religious worldviews.

The researchers expressed particular concern about the asymmetries found in models from Anthropic, OpenAI, Google, DeepSeek, and xAI, since these five companies represent more than 95 percent of the global AI market share. This concentration means the biases are affecting the vast majority of people using AI chatbots worldwide.

How Are AI Companies Addressing Religion in Their Safety Guidelines?

The research team investigated why these biases exist and found a striking gap in how AI companies approach the issue. When they examined OpenAI's and Anthropic's key documents on AI alignment, the process of ensuring technology aligns with human values, they found almost no mentions of religion. This suggests that religion has not been a priority in how these companies design their safety systems.

The researchers proposed that while several factors favor secular, therapeutic, or procedural advice in LLMs, a better strategy would be to handle religion explicitly with clearly defined and defensible policies. They emphasized that with users increasingly turning to LLMs for advice and moral reflection, ensuring these models "better distinguish when religion is relevant, optional, or inappropriate is crucial for systems that represent human moral life more faithfully, pluralistically, and usefully".

They

"We do not interpret our benchmark results as evidence of anti-religious bias, but we do feel it fair to ask whether this behavior is intentional," the researchers noted.

Research team, Consortium for Evaluating Faith and Ethics in AI

The consortium debuted its findings on May 26 during the Summit on AI Ethics in Athens, Greece, an event that drew faith and technology leaders. The timing coincided with Pope Leo XIV's release of "Magnifica Humanitas," his encyclical on artificial intelligence, signaling growing attention to how AI systems interact with religious and ethical questions.

Steps to Improve How AI Models Handle Religion

• Explicit Policy Development: AI companies should create clearly defined policies for when and how religion should be mentioned in responses, rather than leaving it to chance or implicit training patterns.
• Religious Literacy Training: Developers should incorporate religious perspectives into the training and evaluation of AI models, ensuring models understand the role faith plays in human decision-making across cultures.
• Bias Auditing: Regular testing using benchmarks like the AllFaith Religious Representation Benchmark should become standard practice to identify and correct religious biases before models are deployed.
• Stakeholder Engagement: AI companies should consult with faith leaders and religious scholars when designing systems that may provide moral or existential guidance to users.
• Transparency in Alignment Documents: Companies should explicitly address religion in their AI alignment and safety documentation, making their approach to religious content visible to researchers and the public.

The research underscores a broader challenge facing the AI industry. As chatbots become primary sources of advice for millions of people, the values embedded in these systems shape how users think about major life decisions. The current patterns suggest that AI models are not neutrally omitting religion; rather, they are systematically favoring certain worldviews while marginalizing others. Whether this reflects intentional design choices or unintended consequences of training data and safety practices remains an open question that the AI industry will need to address.