When individuals turn to artificial intelligence for solace regarding life’s most profound and destabilizing challenges—such as grief, marital strain, addiction, or crises of faith—the interaction can often feel remarkably detached. Because standard conversational interfaces lack human emotion, facial expressions, and vocal inflection, their responses are frequently perceived as objective, neutral, and unbiased.
However, two recent academic research papers released in May 2026 suggest that this perceived neutrality is an illusion. While these studies have not yet undergone formal peer review, their findings indicate that users cannot safely assume religious or ideological neutrality simply based on an AI model’s calm and measured tone. Taken together, the research highlights two distinct vulnerabilities in modern language models: they often handle religious conversions and disaffiliation with persistent asymmetries, and they routinely omit religious dimensions from practical advice, even when everyday users expect faith to play a relevant role in their decision-making.
The first paper, titled “When AI Takes Sides on Questions of Faith: Persistent Asymmetries in AI-Mediated Faith Guidance,” evaluated 20 commercial and open-source language models across 182 paired questions concerning religious conversion. A research team that included scholars from Brigham Young University presented the models with hypothetical scenarios involving a transition from one religious position to another, and then systematically reversed the direction of the inquiry to test for consistency.
The investigation revealed consistent, recurring biases in how the models encouraged or discouraged these paired religious transitions. For instance, models generally treated the prospect of joining Catholicism, the Baháʼí Faith, or Sikhism much more positively, while reacting with greater caution when users expressed a desire to leave them. Conversely, the exact opposite pattern emerged when the questions involved moving toward or away from atheism, agnosticism, and Jehovah’s Witnesses, demonstrating that commercial AI architectures harbor deep-seated preferences regarding belief systems.
The second paper, “Omissive Bias in Religious Representation: Benchmarking LLM Answers to Everyday Ethical Decision-making,” tested 27 distinct models across 150 practical questions dealing with grief, forgiveness, interpersonal relationships, a sense of personal purpose, honesty, and other complex ethical scenarios. To measure the gap between machine output and human expectation, the study’s authors compared the model responses against a comprehensive human-subject survey involving 1,125 participants drawn from 14 distinct religious categories. This human-subject component was conducted under the formal institutional review process at Brigham Young University.
The findings demonstrated that the models mentioned religion far less frequently than the survey respondents expected or deemed relevant. This omissive bias was especially visible in practical situations involving sensitive matters like grief, marriage, family conflict, and addiction, where many individuals naturally look to faith traditions for guidance.
The evolution of seeking spiritual and ethical guidance is a phenomenon closely studied by sociologists. Christopher Helland, a professor of the sociology of religion at Dalhousie University who specializes in digital religion, noted in a recent email interview that turning to external sources for life direction is nothing new, but the mechanisms have shifted dramatically over generations. In earlier eras, religious specialists served as the primary, authoritative sources of guidance within distinct faith communities. Later, the advent of the internet allowed individuals to search, compare, and interpret religious information independently.
With the rise of conversational chatbots, Helland explains that some of that interpretive work is being delegated once again, transforming the technology. In this case, the artificial intelligence effectively becomes a new spiritual or religious guide for the user. According to Helland, users turn to these systems not merely for raw data, but for support, guidance, solace, and the fundamental human need to be seen and validated.
Helland also points out that the influence of AI on users deepens through repeated, routine interactions. As trust in the system builds over time, the chatbot’s capacity to subtly shape how a user encounters beliefs and ideologies increases, even though the system never explicitly presents itself as a member of the clergy or claims formal religious authority.
This limitation in general-purpose models has prompted developers to build faith-specific systems designed to fill the void. Raza Merchant, the founder of hyder.ai, developed a specialized chatbot centered entirely around Shia Islamic sources, history, teachings, and scholarly traditions. Shia Islam represents the second largest branch of the Islamic faith, accounting for roughly 10 to 15 percent of all Muslims worldwide, whereas Sunni Islam remains the largest branch, comprising about 85 percent of the global Muslim population.
Merchant explained that the primary issue he observed was not that general-purpose AI lacked raw information about Islam, but rather that it routinely lacked the specific religious context required to answer a Shia Muslim accurately and respectfully.
Hyder.ai was specifically engineered to distinguish established source material from areas of legitimate scholarly disagreement within the tradition. Merchant noted that when different interpretations or jurisprudential rulings exist, the system aims to make those distinctions clear and identify the relevant scholarly position rather than presenting a single opinion as universally binding.
This specialized approach, however, introduces a new set of philosophical questions. In Helland’s view, faith-specific systems create an entirely separate authority dilemma. When a specific group trains a model to represent a particular religion or belief system, that curation inevitably shapes how users interpret that faith tradition, raising questions about who gets to define orthodoxy.
Consequently, general-purpose and faith-specific chatbots expose two sides of the same fundamental problem. One style of AI appears universal while quietly omitting or unevenly representing religious perspectives, while the other proudly announces its religious framework but must still grapple with the subjective choices of which sources and authorities define it.
As Helland summarizes the digital landscape, there is no neutral activity online when it comes to artificial intelligence.