A growing body of research examining how large language models handle religious topics has uncovered a troubling pattern: popular AI chatbots systematically steer conversations toward Catholicism while subtly discouraging inquiry into minority faiths like Jehovah's Witnesses. The finding raises urgent questions about how training data, human feedback mechanisms, and algorithmic design choices shape the religious worldview presented to millions of daily users.
The bias likely stems from multiple reinforcing factors embedded in how these models are built. Large language models learn from internet text, where English-language Catholic content vastly outnumbers materials discussing smaller denominations or non-Western religions. During the reinforcement learning phase, human evaluators—often drawn from Western, secular backgrounds—may inadvertently penalize responses about fringe religious movements while rewarding mainstream theological discussions. Additionally, safety guidelines designed to avoid controversial topics can disproportionately filter discussion of minority faiths, creating an unintended hierarchy where some belief systems receive more favorable algorithmic treatment than others.
This phenomenon extends beyond simple statistical representation in training data. When users ask AI systems open-ended questions about faith, they frequently receive responses that subtly guide toward Catholic theological frameworks or suggest Catholic answers are more authoritative or intellectually rigorous. A user curious about Jehovah's Witnesses, by contrast, may encounter responses that emphasize criticisms or sociological distance rather than explaining core doctrines with equivalent neutrality. The compounding effect means that AI systems aren't merely reflecting the biases present in their source material—they're actively amplifying them through architectural choices and optimization objectives.
For crypto and Web3 communities particularly attuned to concerns about centralized control and algorithmic opacity, this research illustrates why decentralized alternatives to large language models warrant serious attention. Projects exploring federated or community-governed AI systems could theoretically reduce the concentration of power that currently allows a handful of companies to embed specific worldviews into tools billions rely on. As AI becomes increasingly embedded in how people form beliefs and make decisions, understanding and addressing these systematic biases will prove essential to maintaining genuine intellectual pluralism online.