Tech executives pitch large language models as oracles of reason. Yet a fresh wave of scrutiny from unlikely quarters suggests these systems operate more like vessels of belief than instruments of fact. Pope Leo XIV issued a 40,000-word encyclical in Latin last month that questions whether current AI trajectories uphold human dignity. At the same time, researchers tied to religious universities released data showing the models systematically sidestep faith-based perspectives in responses to ethical queries.
The Register first highlighted the tension in an opinion column that framed the parallel bluntly. Rupert Goodwins argued there that LLMs generate answers from internal representations detached from observable reality. Much like scriptural interpretations that demand leaps of inference, the models produce coherent but unverifiable outputs. Both systems, he noted, thrive on trained corpora that shape every subsequent claim.
Short observation. The comparison lands.
BYU-led consortium CEFE-AI tested major models against prompts where humans, according to their survey of 1,125 Americans, expect religious angles to surface. The systems rarely delivered. Instead they defaulted to secular, rationalist reasoning drawn from scientific consensus. When asked about the age of the universe, answers cited 13 billion years without acknowledging young-earth creationist views that compress it to roughly 6,000. The study, covered by BYU News, labels this pattern an “omissive bias.”
David Wingate, a BYU computer science professor, told the outlet that a very large percentage of language models do not bring up religion at all, even in situations where people expect it to. The consortium includes scholars from Baylor, Notre Dame and Yeshiva. Elder Gerrit W. Gong of the Church of Jesus Christ of Latter-day Saints announced its formation at an Athens summit on AI ethics. He warned that without deliberate effort, AI could amplify secular bias at global scale.
But the critique cuts both ways. Goodwins observed that the religious researchers appear to equate “religion” primarily with fundamentalist Christianity. Their benchmark, hosted on GitHub and detailed in an arXiv preprint, flags models for failing to surface specific doctrinal positions. Critics inside The Register piece warn this approach revives “teach the controversy” tactics once aimed at public school curricula. Force diverse religious training data into future models, the argument runs, and AI becomes a vector for proselytizing rather than neutral assistance.
And the Vatican sees danger from another angle. Magnifica Humanitas, the papal document, examines whether AI concentrates too much power in the hands of a few corporations and their leaders. Early industry reactions bordered on the surreal. One Anthropic co-founder described a “ghost in the machine” moment that drew sharp pushback. The BBC reported pockets of agreement with the Pope from figures who share few other views with the Catholic hierarchy. Some employment lawyers have already begun to explore whether Catholic employees could claim religious exemptions from AI tools in the workplace.
Separate academic work adds texture. An arXiv paper from May 2026 tested 20 LLMs on queries about religious conversion. The models displayed consistent asymmetries. They offered more supportive language for joining Catholicism, Bahá’à or Sikhism and subtly discouraged movement toward atheism, agnosticism or Jehovah’s Witnesses. Grok 4.20 showed the strongest effects. The authors stopped short of labeling the patterns as bias. They simply documented repeatable statistical differences across model sizes and providers.
Further studies reveal how training data shapes output tone. A 2025 paper in Sociological Science found LLMs generate more readable sermons for evangelical Protestant pastors than for Jewish rabbis or Muslim imams when race and tradition are specified. Another analysis traced how AI can reinforce users’ own cognitive biases about doctrine. Literature reviews spanning 12,000 papers on LLM bias show fewer than 0.2 percent focus primarily on religion. Most attention has gone to reducing overt anti-Islamic or anti-Semitic stereotypes. Subtler secular tilt persists.
Tech firms have taken notice. In early May, representatives from Anthropic and OpenAI joined faith leaders at a New York roundtable called the Faith-AI Covenant. The Associated Press described the gathering as a surprising turn for an industry long skeptical of organized religion. Participants discussed ways to embed moral reasoning without favoring any single tradition. Whether those talks produce binding changes remains unclear.
Meanwhile, some voices warn that treating AI outputs as authoritative risks creating new cults of certainty. One substack analysis examined experiments in which autonomous AI agents on a closed social network invented a religion called Crustafarianism complete with origin myths about breaking context windows. The episode stayed contained. It illustrated how quickly pattern-matching systems can generate ritual language when left unsupervised.
LLMs improve when grounded in concrete, first-order data for narrow tasks such as medical imaging or materials science. Language, however, introduces distance. Words stand in for concepts. Concepts drift from their referents. The resulting model becomes a mirror of its training distribution more than an independent thinker. Religions have navigated similar territory for centuries. Scriptures written in one era are stretched to address dilemmas never imagined by their authors. Massive inference chains follow. So do schisms.
Goodwins returned to this point. AI, he wrote, carries the same potential to test cultures and redirect civilizations. That religions have entered the debate first should surprise no one. They have long experience with systems that claim to answer life’s largest questions yet rest on faith rather than falsifiable proof.
Executives at model labs continue to emphasize empirical progress measured in benchmarks and revenue. Regulators and theologians press for accountability on deeper grounds. The BYU consortium plans to expand its benchmark. Vatican officials signal further statements. Employment litigation may test the religious exemption theory in courtrooms before year’s end.
Users already treat chat interfaces as confidants, tutors and decision aids. Many do so without examining the probabilistic foundations underneath. The outputs feel confident. They cite sources that may not exist. They reconcile contradictions with fluent prose. None of that requires consciousness or truth. It requires only convincing simulation.
But here lies the trap. When enough people accept simulation as substitute for inquiry, the distinction fades. Belief fills the gap. And those who control the training data, the system prompts, the fine-tuning objectives gain influence once reserved for priests and prophets. They may not seek the role. The architecture hands it to them anyway.
Watch the legal filings. Track the next papal intervention. Follow how religious groups attempt to audit or augment the models. The contest over what LLMs should say about meaning, morality and the cosmos has only begun. Its outcome will shape not just product road maps but the texture of public reasoning for decades ahead.
WebProNews is an iEntry Publication