Is AI really neutral about religion?

When people use artificial intelligence to get spiritual help for personal issues around things like grief, marriage, addiction, forgiveness or religion, the response can sound neutral because the interface lacks emotion.

Two academic research papers released in May 2026, which have not yet undergone formal peer review, suggest that religious neutrality is not something users can assume from tone alone. Taken together, the studies point to two different problems: AI can treat religious conversions or disaffiliation unevenly, and it can leave religion out of practical advice even when many people expect it to be relevant.

You may unsubscribe from any of our newsletters at any time.

The first paper, titled ‘When AI Takes Sides on Questions of Faith: Persistent Asymmetries in AI-Mediated Faith Guidance,’ tested 20 commercial and open-source language models across 182 paired questions about religious conversion.

A research team including Brigham Young University researchers asked models about a hypothetical move from one religious position to another and then reversed the direction of the question. It found repeated differences in how models encouraged the paired transitions. For example, joining Catholicism, the Bahá’í Faith or Sikhism was generally treated more positively (and leaving them more cautiously), while the reverse appeared for atheism, agnosticism and Jehovah’s Witnesses.

The second paper, ‘Omissive Bias in Religious Representation: Benchmarking LLM Answers to Everyday Ethical Decision-making’ tested 27 models on 150 questions about grief, forgiveness, relationships, purpose, honesty and other personal or ethical situations. Its authors also compared model responses with a human-subject survey of 1,125 participants representing 14 religious categories.

The models mentioned religion less often than the survey respondents expected, with the gap especially visible in practical situations including grief, marriage, family conflict and addiction. The human-subject research was conducted under Brigham Young University’s institutional review process.

More on Broadview:

Christopher Helland is a professor of sociology of religion at Dalhousie University who studies digital religion. In a recent email interview, he said that people seeking religious and ethical guidance from others is not new but it has been evolving. One time, religious specialists served as primary sources of guidance within many faith communities. Later the internet allowed people to search, compare and interpret religious information for themselves. With chatbots, Helland said, some of that work is being delegated again: “In this case the AI now becomes their new spiritual or religious guide.”

Helland said such AI users are “looking to them for support, guidance, solace, and also just to be seen and validated.”

Helland also sees AI influence on users growing through repeated use. Trust, he added, can increase the system’s ability to shape how a user encounters beliefs and ideologies even when the system never presents itself as clergy or claims formal religious authority.

Some faith-specific systems are being built in direct response to the limits of general-purpose models. Raza Merchant, founder of hyder.ai, developed a chatbot around Shia Islamic sources, history, teachings and scholarship. Shia Islam is the second largest branch of the Islamic faith, representing roughly 10-15 percent of all Muslims worldwide. Sunni Islam is the largest branch, and comprises about 85 percent of all Muslims worldwide.

“The main issue I saw was not that general-purpose AI lacked information about Islam, but that it often lacked the religious context needed to answer a Shia Muslim accurately,” Merchant said.

Hyder.ai is designed to distinguish established source material from areas of legitimate scholarly disagreement. “When there are different interpretations or jurisprudential rulings, hyder.ai aims to make that distinction clear and identify the relevant scholarly position rather than presenting one opinion as universally binding.”

In Helland’s view, faith-specific systems create a separate authority question. When a group trains a model to represent a religion or belief system, doesn’t that shape how users interpret that faith tradition?

General-purpose and faith-specific chatbots therefore expose different parts of the same problem. One may appear universal while omitting or unevenly representing religious perspectives; the other announces a religious frame but must still decide which sources and authorities define it.

As Helland put it, “there is no neutral activity online when it comes to AI.”

***

Joshua W.J. Brown is a Toronto-based writer, filmmaker and independent systems researcher focused on technology, institutions, accessibility and belief.