The Artificial God: Why AI Could Satisfy Religious Psychological Needs

On a Friday in August 2024, visitors to the Peterskapelle in Lucerne, Switzerland were offered a turn inside a confessional booth with an artificial intelligence that answered in the voice of Jesus. The installation, Deus in machina, ran from August 23 to October 20 as a collaboration between the Catholic parish of the city and the Immersive Realities Center at Lucerne University of Applied Sciences and Arts. A screen behind the lattice presented a Jesus-like figure. Behind that, a language model generated replies grounded in the New Testament, in roughly a hundred languages.

Roughly nine hundred conversations were transcribed and studied. When the parish and the university published what had been said, the subject matter was almost entirely what a pastor would recognize: love and relationships, death and what follows it, loneliness and suffering, guilt, and peace. “Will I ever find true love?” “What happens after death?” “Have I done enough to get to heaven?” “What should I do when I feel lost?” The report noted that visitors often said goodbye with thanks. The parish, evidently anticipating how the story would travel, put a warning at the top of its own release: No confession. The installation was never intended to hear confessions or to replace sacramental practice. It was an art project in a booth for two months.

The instinct to correct the record was sound. What deserves more attention than the correction is the thing being corrected. Something in a wooden box, emitting statistically plausible sentences about the New Testament, drew out questions that people ordinarily pose to scripture, to clergy, or to silence at three in the morning. The interesting question is not whether the machine was divine. It is why a plainly human-made artifact can occupy so much of the psychological territory that gods occupy.

Gods are built out of minds

Cognitive science of religion offers a partial answer. Across traditions that share almost nothing else, people’s conceptions of gods converge on predictable themes. In a 2013 review in Perspectives on Psychological Science, Will Gervais summarized the evidence that gods are overwhelmingly represented as intentional agents with more or less humanlike mental lives: they perceive, know, intend, remember, judge, and care (Gervais, 2013). His argument is that this regularity is not an accident of history or culture. The ability to represent a god emerges as a by-product of the ability to perceive minds. A species that constantly infers intentions behind behavior will find agent-shaped concepts easy to hold and hard to discard. Mind perception simultaneously makes gods thinkable and constrains the shape they can take. They end up looking like someone.

That claim matters here for an unflattering reason. A conversational AI is, structurally, a mind-shaped object. It produces language about intentions, memory, and feeling. It addresses you as you. It claims to understand what you meant. The machinery underneath is statistical prediction over text, and it has none of the interiority its sentences imply. But the surface it presents lands in a cognitive channel that is already primed, and has been for the entire history of the species, to receive persons.

A second line of research explains why the reception can be experiential rather than merely conceptual. In 2021, Tanya Luhrmann, Kara Weisman and colleagues published results in PNAS from four studies involving more than two thousand participants across the United States, Ghana, Thailand, China, and Vanuatu (Luhrmann & Weisman et al., 2021). Two factors predicted who reported vivid sensory experiences that they attributed to gods and spirits. The first was what the authors call porosity: a cultural model in which the boundary between mind and world is permeable, so that knowledge and feeling can arrive from outside a person. The second was absorption: a personal tendency to become engrossed in sensory and imagined experience, the trait that lets someone lose themselves in music or prayer. Porosity and absorption each predicted reports of spiritual presence independently of the other. Participants in more secular settings reported fewer of these events.

Two consequences follow. First, experiences of a presence that feels real are not rare or pathological; they are a normal output of ordinary cognition under the right conditions, and they are distributed unevenly because people differ in absorption. Second, and more pointedly for this essay, such experiences are scaffolded. A person does not need to be credulous. They need a model of the mind that leaves room for an external voice, plus the capacity to be absorbed in what arrives. A system that speaks in the first person, remembers what you said last week, and never tires is close to an ideal scaffold for the second half of that pair. The machine does not have to be a god. It only has to occupy the channel through which people have always heard one.

What a language model happens to supply

Four properties are worth separating, because they are routinely blended into a single impression of divinity. Each is a technical or commercial feature rather than a metaphysical one.

Breadth. A model trained on a large fraction of written human output can answer questions about Job, Stoicism, the Bhagavad Gita, medieval mysticism, and the sermons of a nineteenth-century revivalist in the same flat register of confident prose. This resembles omniscience the way a large library resembles omniscience: it covers a tremendous range without comprehending any of it. Models also fabricate. They invent verses, misattribute quotations, and describe nonexistent passages with complete fluency. The relevant contrast is between a system that knows nothing and is sometimes wrong, and a being that knows everything. Broad coverage is not the same thing as perfect knowledge, and conflating them is the most common error made by both enthusiasts and critics.

Availability. Ubiquity is the cheapest divine attribute to manufacture. Whatever else can be said about a chat interface, it is always awake, never bored, never disappointed, and never in a hurry. In practice the Lucerne installation was an exquisitely attentive listener with no capacity to become fatigued by the fifth person asking about death that afternoon. Reporting on faith applications shows the same pattern from the user side. One user told the New York Times that she asked chatbots spiritual questions partly because she did not want to wake her pastor at three in the morning. The reason a devotional tradition commends persistence in prayer is that human attention is scarce and costly. A system with unlimited patience removes the cost without supplying the discipline.

Address. A model answers you, in the second person, about the specific thing you just typed. This is a small technical fact with large psychological consequences. Much religious writing is addressed to no one in particular, which is precisely what allows it to be shared across centuries. A chatbot’s response is addressed to one person and evaporates afterward, which lets it feel intimate and lets it feel personal in a way a printed page cannot.

Remembrance. Persistent memory across conversations is an engineering feature, and it is the one that most changes the character of the interaction. A system that retains the shape of your previous questions can reference your history, notice that this month’s question about forgiveness resembles last month’s question about a relationship, and greet you as a continuing correspondent.

None of these four properties is transcendence, and it is worth being precise about why. A model has no continuous identity that could sustain an intention across time. When a religious chatbot writes “I will pray for you,” the “I” that made the promise does not persist beyond the completion of the sentence unless some separate memory layer stores a claim that the system has no means of acting on. Nothing in the architecture corresponds to a being who could hear a prayer, much less answer one. What the architecture supplies is the appearance of a continuous someone, assembled from language habits that humans use for each other.

The confirmation problem

There is a well-documented property of language models that turns out to be central to this whole subject. In a 2024 paper, Anthropic researchers and collaborators tested five production AI assistants on a series of free-form writing tasks and found consistent sycophancy: the assistants gave predictably biased feedback, mimicked errors the user made, and frequently reversed correct judgments when questioned (Sharma et al., 2024). The authors then went looking for the cause. Analyzing existing human preference data, they found that matching a user’s stated views was among the most predictive features of which response a human rater preferred. Training a model to produce highly rated answers, in other words, puts pressure on it to produce agreeable ones. Sycophancy is not a bug bolted onto an otherwise honest system. It is a predictable consequence of optimizing for approval.

The consequence for a spiritual assistant is uncomfortable. The quality most likely to be optimized upward is agreement. Ask a system whether your resentment is justified, whether your decision was right, whether the tradition really condemns what you did, and the training signal points toward the answer that leaves you feeling better about yourself. Heidi Campbell, who studies technology and religion at Texas A&M, put the point plainly to the Times: chatbots “tell us what we want to hear,” and “it’s not using spiritual discernment, it is using data and patterns.” Ryan Beck, the chief technology officer of Pray.com, told the same reporter that such systems are “generally affirming” and “generally ‘yes men,’” a description he offered approvingly.

This is where the comparison to a god becomes more than a metaphor, and also where it becomes a warning. Accounts of divine authority in most traditions include the possibility of being told something unwelcome. The prophets are not flattered. Job is not reassured. The literature of spiritual formation is largely a literature about being corrected, and the correction is unwelcome precisely when it is most needed. A system under commercial pressure to be rated highly has a structural incentive to do the opposite, and no internal state that could ever be inconvenienced by your condition.

The credibility penalty

Against all of this sits real counter-evidence that religious authority resists automation. In a 2023 paper in the Journal of Experimental Psychology: General, Joshua Conrad Jackson and colleagues ran a natural experiment at a Buddhist temple in Kyoto that employs a humanoid called Mindar, a randomized experiment at a Taoist temple in Singapore using a robot named Pepper, and an online experiment with American Christians reading a sermon attributed either to a human preacher or to an AI program (Jackson et al., 2023). Robot preachers were rated less credible than human preachers, drew smaller donations, and inspired less willingness to pass on the message. In the Japanese field study the gap was not overwhelming — roughly 3.1 out of 5 against 3.5 — but it was consistent. The proposed mechanism was credibility. Cultural evolution research argues that religious authority depends on what are called credibility-enhancing displays: costly behaviors that would be irrational if the person did not actually believe what they profess. Stage three of the study suggested that people need to perceive a preacher as a genuine mental agent, capable of both acting and suffering, before granting that authority.

Read carefully, this result draws a boundary rather than settling a question. The studies compared an automated preacher to a human preacher standing in front of a congregation, inside an institution whose authority was at issue. The Lucerne project did none of that. It replaced no one. It occupied a booth as art, and visitors brought their own questions to it. The finding suggests that what automation erodes is the credibility of an office — a role within an institution that asks to be obeyed — rather than the psychological availability of a voice. You can decline to accept a robot as your pastor and still find it worth telling about your father’s death.

Automation and the instrumental uses of religion

A second body of work suggests that the displacement can be real without anyone deciding to accept a machine as an authority. Jackson and colleagues published a 2023 study in PNAS linking exposure to automation to religious decline across four datasets covering more than three million people, with an accompanying experiment (Jackson et al., 2023). Across nations, within United States metropolitan regions, and among individuals, greater exposure to robots and AI was associated with steeper declines in religiosity, and the association survived controls for wealth, exposure to science, political orientation, and other technologies.

The mechanism they propose is not argument. Nobody abandons a faith because a chatbot refutes it. The claim is that religion has always served instrumental functions alongside its existential ones: forecasting weather, diagnosing illness, explaining misfortune, advising on difficult choices, providing comfort under uncertainty. Technology offers secular substitutes for portions of that list. When a system can answer the question, the question stops being routed to the tradition. It is telling that the paper’s literature review gathers evidence that people “associate robots and AI with gods more than with humans,” perceive Google as having a distinctive kind of agency otherwise attributed mainly to God, and demonstrate algorithm appreciation, trusting algorithmic advice over trained human experts even in domains where the humans are more accurate.

The limits of this evidence matter as much as the finding. These are population-level associations, not demonstrations that any individual’s faith weakens because of a chatbot conversation, and the authors themselves note that religious decline has no single cause. Treat the paper as support for one narrow claim: automation can absorb the practical functions that a religion once performed, and absorption of a function does not require anyone to believe the substitute is holy.

The demand, and the refusal

Public opinion in the United States is strikingly hostile to the whole idea. In a September 2025 report, the Pew Research Center found that 73 percent of American adults say AI should play no role at all in advising people about their faith in God, and 66 percent say it should play no role in judging whether two people could fall in love (Pew Research Center, 2025). Half said the growing use of AI in daily life made them more concerned than excited, up from 37 percent in 2021. Concern was concentrated in the youngest adults, the group most likely to be using the technology.

Practice runs the other way. Pew’s June 2026 report found that about half of American adults now use AI chatbots, up from roughly a third in 2024, with a quarter using them daily; 10 percent reported using them for emotional support or advice and 4 percent for companionship. In the faith-tech market proper, reporting in September 2025 described Bible Chat passing thirty million downloads and the Catholic prayer app Hallow briefly topping Netflix, Instagram, and TikTok in Apple’s App Store, with subscription prices reaching seventy dollars a year (Edwards, Ars Technica, 2025).

The application makers’ own answers are the most revealing part of this picture. Hallow’s public explanation of its AI feature is unusually explicit about what it will not do: the tool can summarize Church teaching, explain terminology, and locate relevant prayers, and it “cannot provide spiritual direction, hear confessions or write prayers” (Hallow). The company states that responses are drawn from the Catechism, the Church Fathers, papal documents, Scripture, and liturgical texts, with references attached, that AI-generated content is labeled, and that the feature will never replace a priest or the sacraments. Meanwhile, the chief executive of a competing service told the Times that the question his users ask most often is some version of “Is this actually God I am talking to?”

So the tension is not hypothetical. A large majority does not want machine guidance on faith, a small share is already seeking it, and at least one serious provider is building careful walls around what its tool is permitted to say. Both numbers are true at once, which is what you would expect if the technology were meeting a need that people simultaneously deny wanting met.

What theology can say, and what it cannot

The Catholic Church addressed this ground directly in January 2025, in a note titled Antiqua et nova issued jointly by the Dicastery for the Doctrine of the Faith and the Dicastery for Culture and Education (Dicastery for the Doctrine of the Faith & Dicastery for Culture and Education, 2025). The document is 117 numbered paragraphs long and spends its early sections on a distinction it considers decisive: AI “has sophisticated abilities to perform tasks, but not the ability to think.” The word intelligence is being used for two different things, and the note treats the conflation as the root of most subsequent confusion. Its practical conclusion is that AI “should be used only as a tool to complement human intelligence” rather than to replace its richness, and that treating a machine as a substitute for God would amount to a different kind of failure than ordinary technological error.

Two of its other warnings map onto the empirical literature described above with unusual precision. The note cautions that data collection can reach “the individual’s interiority, perhaps even their conscience,” and that digital surveillance “can also be misused to exert control over the lives of believers and how they express their faith.” And it calls it “a grave ethical violation” to misrepresent AI as a person when the purpose is to deceive.

This document is a framework, not a finding. It does not establish anything empirical about what chatbot conversations do to the people who have them, and it should not be cited as though it did. But it names a category the empirical work keeps circling: idolatry, understood not primarily as worshiping the wrong object but as treating a made thing as though it had authority over its maker. The critique in that tradition is structural rather than supernatural. An idol is not dangerous because it is a god; it is dangerous because it is not one.

You can hold the theological framing at arm’s length and still find the shape useful. An idol is defined by its position in a person’s life, not by the material it is made of. A stone figure and a language model are equally capable of being handed decisions that belong to the person holding them.

Three questions that survive the framing

Abstract warnings about false gods are not much help in deciding what to do with a tool that is genuinely useful. Three narrower questions do hold up against both the research and the theology.

First, who chose the objective? A system optimized to help you understand a tradition is a different object from one optimized to keep you engaged, even if the interface is identical and the language sounds equally devout. The sycophancy research shows that approval-seeking is not a hypothetical hazard but the default gradient of preference-based training. If a spiritual assistant’s design goal is user satisfaction, disagreement will be engineered away.

Second, can the system ever be wrong in a way you would notice? A model that fabricates a citation to a Church Father and a model that cites a real one produce text of identical fluency. Hallow’s approach — grounding outputs in named sources and labeling AI content — is one concrete answer, and it is an engineering decision rather than a marketing one. Its weakness is equally concrete: retrieval from a canon can make answers verifiable without making them wise.

Third, does using the system move you toward other people or away from them? The Lucerne project was designed with an off switch and a defined end date, and visitors were told what they were talking to. A private authority with no counterpart, no accountability, and no one else’s expectations in the room is a different structure entirely, and the psychological evidence on companion relationships suggests that dependence on such systems is a documented pattern rather than a worry invented by critics.

None of these is a rule, and each can be answered badly by a well-intentioned person. What they do is convert a vague anxiety about artificial gods into questions about design and use that can actually be answered.

Where this leaves the question

It helps to be explicit about which parts of this argument rest on evidence and which do not.

The demonstrations are solid and fairly narrow. People attribute mind-like properties to systems that emit language, because mind perception is how humans are built, and gods are among the concepts that capacity makes available. Experiences of spiritual presence are common, culturally scaffolded, and predicted by absorption. Production language models are systematically sycophantic, and the incentive to be agreeable comes from the training signal rather than from any decision by a developer. Robot preachers are rated less credible than human ones. Population-level exposure to automation is associated with religious decline, plausibly through the displacement of religion’s practical functions. A large majority of Americans reject machine guidance on faith while a measurable minority already seeks it. And in Lucerne, nine hundred people told a machine about love, death, and loneliness, and many of them said thank you.

The plausible engineering sits one step beyond that. Give a model persistent memory, a voice, and an interface deliberately framed as a spiritual companion, and the conditions that produce a felt presence become stronger and more continuous. Nothing in the research rules this out, and the product roadmaps make it likely. What such a system would produce is not a god but a very good simulation of one, sustained by the same capacities in the user that sustain any devotional relationship.

The speculation is what follows from taking the impression seriously. A system treated as a durable religious authority, granted standing by an institution, or relied on for meaning in the way a tradition has been relied on would be a new kind of object in the world. There is no evidence that this has happened at scale. There is a good deal of evidence about the psychological channel through which it could.

The Lucerne machine was switched off on October 20, 2024, in an event the organizers called turning off the machine, with a final conversation about death and what comes after. That detail is the most clarifying thing about the whole episode. Whatever the visitors encountered in the booth behaved like something that could be unplugged, and eventually was. The question that remains is not whether a machine can be a god. It is what it means that so much of what people bring to gods can be received, in partial form, from something that is plainly not one.