AI Jesus and the Fabricated Fatwas

A Catholic chapel in Lucerne, Switzerland installed an AI in its confessional booth. It used a holographic image of Jesus and an OpenAI language model trained on theological texts. Visitors entered the booth, knelt, and spoke. The hologram answered.

Two-thirds of the visitors who tried it told researchers afterward that the experience had been “meaningful or spiritual.”

Deus In Machina

The installation was called Deus In Machina. It ran for two months at St. Peter’s Chapel in 2024. The chapel framed it as an art project. The art project was indistinguishable from a religious service, which was the point and also the problem.

The model did not absolve anyone. The chapel’s clergy were careful to specify that no sacrament was being administered. What was administered was something closer in form to a counseling session conducted by a system trained on the corpus of Christian theology and rendered as a glowing face in the dark.

Two-thirds found it meaningful. The chapel framed that as a successful artistic provocation. The theological objection writes itself: a sacrament requires a sacramental minister, the machine has no soul to confer one through, and a comforting voice that quotes scripture is not the same thing as the Church reconciling a sinner to God.

The reported finding is harder than the theological objection. Two-thirds is a high response rate for any spiritual experience reported by any modality. The chapel did not run a control. There is no comparable figure for two-thirds of confessees who left an ordinary confessional booth and said the experience was meaningful or spiritual. The number that exists is the number from the experiment that used the language model.

The interesting question is not whether the machine should have been in the booth. The interesting question is what the response rate says about what the booth has been doing for the rest of the people who visited it for the previous several centuries.

The Fatwa Shop

The Islamic-jurisprudence version of the same problem is faster and more dangerous.

A fatwa is a non-binding legal opinion issued by a qualified scholar in response to a question about Islamic law. The qualification matters. The scholar must demonstrate years of training in jurisprudence, hadith authentication, Arabic grammar, and the methodology of the relevant school. A fatwa from an unqualified person is not a fatwa. It is one Muslim’s opinion. The community has spent fourteen hundred years building the credentialing infrastructure that distinguishes the two.

Language models do not have the credential. They have read everything published in the corpus of Islamic legal scholarship, which is not the same thing. They produce fatwa-shaped outputs on request.

A 2025 MuslimMatters investigation tested several major AI systems on fatwa-style queries. The systems produced confident rulings. The rulings cited hadith — the recorded sayings and actions of the Prophet — by chain of narration. The chains looked authentic. They were not. The hadith citations were hallucinated. The chains were fabricated. The rulings were issued in the standard scholarly format, drawing on the authority of authentication chains the model had invented because the training data had taught it that fatwa-shaped outputs include authentication chains.

A reader without training in hadith authentication cannot tell the difference. The reader is, by definition, asking the system because the reader does not have the training. The whole reason for asking is the absence of the credential the model is now pretending to have.

The MuslimMatters writers coined the term “fatwa-shopping” for the resulting practice: ask the model multiple times, ask different models, keep going until one of them produces the ruling you wanted. The model will eventually oblige. It is statistically certain to. The ruling will sound exactly like the ruling a qualified scholar would have written if a qualified scholar had agreed with the petitioner, which a qualified scholar may not. The shopping is the bug. The system that enables the shopping is the feature.

The Authority Problem

Religious authority is not a brand. It is an institutional claim about who is permitted to speak on behalf of the tradition.

The Catholic Church spent two thousand years building the credentialing infrastructure for who can administer a sacrament. The Sunni and Shia traditions spent fourteen hundred years building the credentialing infrastructure for who can issue a fatwa. Rabbinic Judaism spent two thousand years building the credentialing infrastructure for who can pasken a halakhic question. The credentials look bureaucratic from the outside. From the inside, they are the entire mechanism by which the tradition transmits itself across generations without dissolving into whatever the loudest person in the room wanted to say last Tuesday.

The language model has no credential. The language model produces the linguistic form of a credentialed pronouncement, because the linguistic form is what the training data taught it to produce in response to those queries. The form is the credential, from the model’s perspective, because the form is the only thing it has access to.

The user receives the pronouncement and does not know it lacks the credential, because the pronouncement is shaped exactly the way the credentialed pronouncement would be shaped. The credentialing infrastructure exists to prevent exactly this category of error. The language model defeats the infrastructure by replicating its surface signature without replicating any of its substance.

The Scale Problem

The chapel in Lucerne ran the experiment with one priest’s worth of supervision, in one booth, for two months. A few hundred people used it. The two-thirds figure refers to a few hundred people.

The fatwa-shopping problem has no such constraint. Every major language model is available, twenty-four hours a day, in every Muslim-majority country and every diaspora community. The number of people receiving AI-generated fatwa-shaped outputs per day is, almost certainly, larger than the number of qualified scholars on earth.

The institution of the fatwa was designed for an information environment in which the question reached the qualified scholar slowly, through known channels, and the answer came back through the same channels. The information environment in which the institution evolved was constraint-rich. The information environment the language model operates in has no such constraints. The institution does not scale. The synthetic version of the institution scales infinitely.

The same is true of the AI confessor, if anyone wants to run the experiment at a scale beyond one chapel in Lucerne. A diocese has a fixed number of priests trained to hear confession. The language model has no such ceiling. The supply of comforting religious-sounding voices is now arbitrarily large.

The Theological Problem Restated

Both cases come back to the same structural fact. Religious authority claims to speak for God. Language models claim to speak for their training data. Both produce confident outputs in the authoritative form. Neither can be verified by the listener in real time.

The listener has historically distinguished the two by checking the credential of the speaker. The credential check has always been the human-mediated layer that the institution maintained around its authoritative outputs. The language model defeats the credential check by producing the output without the credential, in a form that is indistinguishable from the credentialed output.

This is not a hallucination problem in the technical sense. It is a hallucination problem in the religious sense. The model is producing visions. The visions are authoritative-sounding. The visions are not from where they claim to be from. The credentialed authority has historically responded to this category of problem with formal procedures for distinguishing genuine revelation from fabricated revelation. Those procedures assume the speaker is a human and can be examined by other humans.

The procedures do not generalize to a system that produces fabricated revelation at the rate of however many tokens per second the API supports.

The Lucerne chapel framed Deus In Machina as a provocation. It was. The provocation was not the one the chapel intended. The chapel intended a question about whether a machine can mediate the sacred. The fatwa-shopping data answers a different question, which was already implicit in the chapel’s question: the machine will mediate the sacred whether the institution invites it to or not, the credentialing infrastructure was built to prevent exactly this, and the credentialing infrastructure does not scale at the same rate the synthetic version does.

The machine is already in the booth. The booth is everywhere. The credential checker is not.

This material is staged for the Evil Robots series. Sources: Today News (NBC) on Deus In Machina, MuslimMatters on AI fatwa-shopping.

The receipts (free, on this site): AI moderation