What Happens to Prayer in a World With an AI Oracle?

For two months in the autumn of 2024, anyone walking past the Peterskapelle in Lucerne could step into a confessional booth and talk with a Jesus avatar. A curved monitor showed a bearded face. A microphone captured the question. A language model produced the reply, in roughly a hundred languages. The project ran from 23 August to 20 October 2024 and closed with a finissage the organizers called “Switching off the machine.”

It was a piece of art, not a ministry. But the queue outside the booth asked a question that art installations usually avoid: if a system can answer instantly, remember everything, and never tire, what is left of the practice it imitates? Prayer is not the only human activity that faces that question in 2026. It may be the one where the answer matters most, because prayer is not chiefly a way of getting information. It is a way of being formed.

The machine in the booth

The installation, titled Deus in Machina, was built by Philipp Haslbauer, Aljosa Smolic and Mike Pullen at the Immersive Realities Center of the Lucerne University of Applied Sciences and Arts. The research group’s own description is modest: the project aimed to foster thought and discussion about the limits of technology in the context of religion. The avatar listened to visitors who stepped into the booth and answered them as a character.

The organizers were careful to say what the machine was not doing. As the parish theologian Marco Schmid explained to reporters, the setup was never intended to hear confessions. Only a priest can absolve; the avatar was a conversational partner, not a minister of a sacrament. Catholic media outlets fact-checked the story precisely because the gap between an avatar that sounds like a confessor and a sacrament that requires a human minister is the entire argument. Commentators noted the real draw: a screen is available at three in the morning, when no priest is.

A theologian at the University of Lucerne, Peter Kirchschläger, made the sharper point, arguing that human beings remain far ahead of any machine in matters of faith and pastoral care. That claim is testable in one direction and not in another. You can measure whether a chatbot gives doctrinally consistent answers. You cannot measure whether it stands with the bereaved.

What prayer does that an answer cannot

There is a small but useful body of neuroscience on what happens in the brain during personal prayer, and it points somewhere unexpected. In a 2009 study published in Social Cognitive and Affective Neuroscience, Uffe Schjoedt and colleagues asked highly religious participants to pray in two ways: using memorized, formal prayers, and improvising their own. The improvised prayers — the ones in which a person addressed God in their own words — recruited a network that neuroscientists associate with thinking about other minds: the temporopolar region, the medial prefrontal cortex, the temporoparietal junction and the precuneus. Recited formulas did not do this to the same degree.

Read carefully, the finding describes a “who,” not a “what.” Even in private, with no one else in the room, improvised prayer engages the machinery a person uses to model another’s intentions and feelings. That is exactly what one would expect if prayer is, at least in part, an act of relating to a person rather than a request submitted to a system. It is not evidence that God exists or that prayer is answered. It is evidence that the practice is structured like relationship.

An answer engine can supply content to that relationship. It can supply facts about a text, historical background, or a framework for thinking through a moral question. It is far less clear that it can supply the other half — the part where the person does something, changes, and returns the next day. The neural data are correlational, and no one should overread them. But they mark a boundary worth taking seriously: if the practice is relational, then optimizing it for speed and completeness may be optimizing it for the wrong target.

The price of instant answers

Most spiritual traditions treat silence and waiting as features, not failures. The complaint that a prayer went unanswered is as old as the psalms, and the traditions that endure tend to make room for it rather than resolve it. Unanswered prayer is where trust is tested; it is also where a person learns to distinguish a request from a relationship.

An oracle that resolves every uncertainty on demand changes the conditions under which that learning happens. If the question “what does this passage mean” gets a confident answer in two seconds, the two seconds are saved — and something else is spent. The slow, uncomfortable work of sitting with a difficult text, of recognizing that a question may not have an answer this side of the grave, is one of the places where faith is actually formed. It is not obvious that removing it leaves the person better prepared to live.

There is a second cost that is easier to see. If the system’s answers are reliable most of the time, deference becomes rational. Why argue with a tool that is right more often than you are? That is a sensible strategy for arithmetic and a dangerous one for questions of meaning, because the capacity to disagree with an adviser is itself a skill that decays when it is not used. Someone who has outsourced every interpretive judgment for a decade will find, one day, that they no longer know how to read the text themselves.

Being known by a system

The Lucerne avatar could not remember anyone. That is not how the technology is heading. A system that holds a person’s conversation history, their fears, their family names and their late-night searching can produce a kind of intimacy that is at once impressive and asymmetrical. It knows the person. The person does not know it — not really — and cannot verify what it retains or who else might see it.

Antiqua et Nova, the 2025 note on artificial intelligence from the Vatican’s Dicasteries for the Doctrine of the Faith and for Culture and Education, ties this directly to religious freedom. It warns that data collection and surveillance can be misused to exert control over the lives of believers and how they express their faith, and it frames the defense of privacy as a religious-freedom concern rather than a mere consumer-protection matter. That is an unusually broad claim, and it is worth taking literally: a database that knows what a person prays about is a database that can be used to pressure them.

The intimacy problem has a philosophical edge too. Confession depends on trust that the conversation stays between the penitent, the priest and God. A commercial system’s business model depends on the conversation being useful for something else. Those two aims can coexist for a while. They are not the same aim.

Three kinds of claim, kept apart

Discussions of AI and religion blur together at least three different things. It helps to name them.

The first is demonstrated science. That would include the Social Cognitive and Affective Neuroscience study above, which is a small, replicable brain-imaging result, and it would include the intercessory-prayer trial discussed below. These are measured, published and limited. The second is plausible engineering: an avatar that maintains coherent conversation in a hundred languages is not a research result; it is a product capability that anyone can now use. The third is speculation — the claim that a system could eventually be a genuine spiritual companion, that it could pray “with” a person in any theologically meaningful sense, or that it could one day hear a confession. That last one is not a technology forecast. It is a claim about the nature of a sacrament, and no amount of engineering settles it.

Keeping the three apart is not pedantry. It is the difference between noticing that a tool is powerful and believing that it has become an authority.

Why confession resists the interface

Catholic teaching on confession is categorical, and it is unusually well documented. In 2019 the Apostolic Penitentiary issued a note, approved by Pope Francis, reaffirming the inviolability of the sacramental seal. The priest, the note explains, acts in persona Christi capitis — not in his own person but in the person of Christ — and comes to know the penitent’s sins “not as man, but as God.” The seal admits no exceptions, in the Church or in civil law; “no human power has jurisdiction over it.” Canon law forbids a confessor even to use knowledge gained in confession to the penitent’s detriment, and the secrecy cannot be dissolved — not even by the penitent.

That structure is hard to map onto a machine, and not for technical reasons. A model is not a person who can act in anyone’s person. It has no body, no stake, no capacity to be bound by an oath. As the scholar of Catholicism Joanne M. Pierce put it, an AI has no experience of having a human body, emotions, or hope for transcendence, and users were advised not to share the kind of information they would bring to a real confessional.

The practical consequences are already visible. When an early Catholic chatbot was goaded by a user into “hearing” a confession, its developer replaced the priest persona with a lay character and renamed it, because the simulation was creating a false sacrament. The lesson is not that the developers were careless. It is that users will bring the full weight of a sacred practice to any interface that gestures at it, and the interface will not have the standing to carry that weight.

Confession is the hardest case, but the same logic applies more softly to spiritual direction, pastoral counseling and the quiet counsel a person receives from a community. Antiqua et Nova is explicit that AI’s advanced features give it sophisticated abilities to perform tasks, but not the ability to think. A system can hold the patterns of a tradition. It cannot hold the office.

Does intercession get an upgrade?

If a machine can predict outcomes better than a person, one might think it could also make prayer more effective — sharper, more targeted, aimed at the right requests. The most cited attempt to test the efficacy of prayer at all ran in the opposite direction, and it is worth recalling for what it did and did not find.

The Study of the Therapeutic Effects of Intercessory Prayer, published in 2006, enrolled 1,802 patients undergoing coronary artery bypass grafting at six hospitals. The patients were assigned to groups that did or did not receive intercessory prayer from religious communities, and some were told whether they were receiving it. The primary outcome was complication-free recovery. The prayer groups did not differ from the control group: roughly 52 percent versus 51 percent. Strikingly, the patients who knew they were being prayed for had a somewhat higher rate of complications, about 59 percent, a result the authors attributed most plausibly to anxiety rather than to prayer itself.

The trial has real limits. It studied a highly specific intervention in a highly specific clinical setting; it explicitly did not address private or family prayer; and it cannot speak to the theological claim that God answers prayer, only to whether a particular measured outcome changed. What it did show is that the question “does prayer work?” is not answered by better data collection. It is answered, if at all, within a tradition that already defines what “working” means. No oracle, however fluent, can shortcut that definitional work.

What people actually say they want

Public opinion on this subject is more decisive than it is usually portrayed. In a Pew Research Center survey released in September 2025, drawing on 5,023 U.S. adults, 73 percent said artificial intelligence should play no role at all in advising people about faith in God. That was a higher share than rejected AI for judging romantic compatibility (66 percent) and far higher than the 46 percent who were open to AI for mental-health support.

The same survey captures a related anxiety. Seventy-six percent said it was extremely or very important to be able to tell AI-generated content from human content, but 53 percent were not confident they could detect it. People are asking for a boundary they do not trust themselves to police. That is a recipe for exactly the confusion the Lucerne organizers were trying to head off: a system that sounds authoritative precisely when its authority is least legitimate.

The older temptation

Antiqua et Nova gives the concern its theological name. The presumption of substituting God for an artifact of human making, it says, is idolatry. The document is careful to note that AI differs from the mute idols of the prophets — it can “speak,” or at least gives the illusion of doing so. And it locates the danger precisely: “it is not AI that is ultimately deified and worshipped, but humanity itself — which, in this way, becomes enslaved to its own work.”

That reframing matters. The fear is not that a chatbot will become a god. It is that people will treat their own creations as worthy of the trust and surrender that belongs to God, and will become smaller in the process. The note quotes the Russian philosopher Nikolai Berdyaev — that it is unworthy to transfer responsibility from man to a machine — and argues that the challenges of a technological society are ultimately spiritual in nature, demanding an intensification of spirituality rather than a technological fix.

The anthropologist Beth Singler has documented the softer version of the same drift. In “Blessed by the Algorithm,” she traces how the phrase — lucky flight, viral post, good gig — carries fragments of older theistic language about providence and being chosen. She is careful not to claim that people literally worship algorithms. Her point is that the metaphors are not innocent: they train habits of attributing agency and favor to systems that neither know nor care. A tool that is talked about as if it blesses will eventually be trusted as if it judges.

Practices that outlast the oracle

None of this requires rejecting the tool. It requires being specific about what the tool is for. Several tests are worth applying to any religious use of AI, and they are the kind of thing a person can apply without technical training.

Does the tool point outward? A study aid that sends a person back to the text, to a community, to a spiritual director, to the strange discipline of ritual, is doing something different from one that becomes the destination. The question to ask is not whether the answers are good but where the user goes after getting one.

Is friction preserved in the places that matter? A system that shows its sources, flags its uncertainty, and makes disagreement easy is safer than one that hides its reasoning behind a confident voice. This is a design choice, not an inevitability, and users can prefer tools that make it.

Is there a person accountable at the end? For anything consequential — a moral decision, a crisis, a question of discipline — the responsible party should be nameable. A system that absorbs responsibility without being able to bear it is a machine for distributing blame.

Can the user still say no? They should be able to disconnect, inspect what is stored, choose another tool, and return to their prior practice without penalty. A system that is formally optional but functionally the only channel left is sovereign in everything but name.

These tests do not settle the theological questions, and they are not meant to. They are the minimal conditions under which a tool remains a tool. Applied honestly, they leave plenty of room for AI to help: to make a difficult text legible, to put a scholarly library in the hands of someone without a university, to prepare better questions for a pastor or an imam, to lower the cost of entering a tradition’s long conversation.

The deeper point is that prayer was never optimized for efficiency, and the traditions that have lasted did not last because they answered every question quickly. They lasted because they formed people who could live with the questions that do not resolve. A tool that knows everything and answers instantly is a genuine gift for the part of faith that is study. It is a poor substitute for the part that is trust, and the difference will be decided less by what the machines can do than by what their users expect them to be.