Ninety-Eight Percent from the Machine

Clergy now ask a question that did not exist four years ago: may I let ChatGPT write the sermon? Congregations ask the other half of it: would I know, and would it matter? Most answers online come from companies selling sermon tools, or from one-line opinions in either direction. This essay does something slower. It reports three occasions when a congregation actually heard a machine speak in a sacred setting, using the accounts of people who were there. It reads the written guidance that exists. Then it sets out the case for and the case against, each as strongly as I can make it.

How the sources were handled. Every quotation below was copied from a page opened with a public request on 30 September 2026. Each one is recorded in this site’s source ledger with its address and the exact sentence. Two of the accounts are in German. Where I quote German, the original wording is given and the English beside it is my own translation, marked as such. Numbers are the sources’ numbers. Where two sources disagree, I give both.

Fürth, June 2023: a service “98% from the machine”

On 9 June 2023 the German Protestant Kirchentag, a church festival held that year in Nuremberg and neighbouring Fürth, put on what its organisers billed as Germany’s first AI service, in the Protestant church of St. Paul in Fürth. The man behind it was Jonas Simmerlein, a theologian and AI artist from Vienna. The Kirchentag’s own report opens with his summary of how much of it was his:

Von dem, was sie heute sehen und erleben, kommen rund 98 % von der Maschine.Jonas Simmerlein, quoted in the Kirchentag’s report “Deutschlands erster KI-Gottesdienst”, June 2023

(My translation: “Of what you see and experience today, around 98% comes from the machine.”) According to the Kirchentag’s report, avatars in human form, projected onto a screen, led the whole programme: greeting, intercessions, psalm, creed, sermon, the Lord’s Prayer and the blessing. Even the music between the parts was composed by an AI. The Protestant news portal evangelisch.de counted more than 400 people in a full church for a service of about 40 minutes. The Kirchentag’s report says about 45. The service ran under the motto “Alexa, starte den Gottesdienst” (“Alexa, start the service”).

The two reports describe the room in detail, and the detail is the useful part.

• People worshipped anyway. The Kirchentag report says that without being asked, the congregation stood for the creed and the blessing and joined in the prayers and psalms, as far as they could keep up with the AI’s speaking pace.
• They also laughed. Both reports record laughter or murmuring at some passages. One avatar opened its sermon with Liebe Gemeinde, es ist mir eine Freude, als erste KI vor Ihnen zu stehen und zu Ihnen zu predigen (my translation: “Dear congregation, it is a joy for me to stand before you as the first AI and preach to you”). When the machine said Möge die Weisheit meines Glaubens dabei helfen, eine gerechtere Welt für alle zu bauen (“May the wisdom of my faith help to build a more just world for all”), evangelisch.de reports that the audience laughed.
• The blessing felt wrong in the body. The avatars were shown to the waist with their arms down. One attendee, quoted by the Kirchentag without a name, said: Ich hätte schon erwartet, dass sie zum Segen die Hände hebt. (“I would have expected her to raise her hands for the blessing.”)
• Most of the written feedback was critical. evangelisch.de summarises it as too impersonal, too emotionless, no movement at the altar, no atmosphere, and music wie im Fahrstuhl, like in a lift.

The sharpest criticism came from Melitta Müller-Hansen, who is responsible for broadcasting in the Lutheran church in Bavaria. As evangelisch.de reports it, she called the prayer language ein einziges Geschwätz (“nothing but chatter”) and asked: Spricht die KI im Namen Gottes? (“Is the AI speaking in God’s name?”). Of the flat avatar voices she said: Die Kehle ist die Seele. (“The throat is the soul.”) The theologian Anna Puzio, who researches the anthropology of technology, took the other side and called the service eine coole Sache, a cool thing. On the wooden avatars she said: Die werden sich verändern (“They will change”).

Two further details matter for the argument later. The panel afterwards agreed, in the Kirchentag’s summary, that AI cannot replace people in worship, because a preacher draws on personal experience and a relationship with the congregation that a model built from thousands of other people’s texts does not have. Ralf Peter Reimann, the internet officer of the Protestant Church in the Rhineland, asked a less obvious question: whose texts is the model built from, and which voices are missing? And Simmerlein himself did not present the service as a replacement. He saw the chatbot as a possible assistant for preachers, and told evangelisch.de: Die Predigt ist ein Geschehen zwischen Prediger und Gemeinde - und das merken wir. (“The sermon is something that happens between preacher and congregation, and we notice that.”)

At the very end the moderator, Jürgen Pelzer, asked ChatGPT for a closing blessing. The Kirchentag report records the machine’s answer:

Natürlich kann ich euch einen Abschlusssegen sprechen. Bitte beachtet jedoch, dass ich als KI-Modell nicht über eigene religiöse Überzeugungen oder Autorität verfüge.ChatGPT, as recorded in the Kirchentag’s report “Deutschlands erster KI-Gottesdienst”, June 2023

(My translation: “Of course I can speak a closing blessing for you. Please note, however, that as an AI model I have no religious convictions or authority of my own.”) It is hard to state the problem more exactly than the machine did.

The Hamptons, early 2023: a sermon revealed afterwards

A few months earlier, and with a different design, Rabbi Josh Franklin of the Jewish Center of the Hamptons in New York preached a sermon written by ChatGPT. The Jewish Chronicle reported it on 16 January 2023, calling him the first Jewish leader to deliver a sermon written entirely by AI. The sermon was on the Torah portion Vayiggash. Franklin began by telling the congregation that he had plagiarised what he was about to say.

When he finished, people guessed. According to the JC, the guesses included other local rabbis, Franklin’s father, and the late Rabbi Jonathan Sacks. Then he told them a machine had written it, and they applauded. His reply, as the JC reports it:

You're clapping, I'm deathly afraid.Rabbi Josh Franklin, quoted in The Jewish Chronicle, 16 January 2023

Franklin told the JC that he did it to teach his congregation about the limits of a robot rabbi, and he named those limits. ChatGPT might be really great at sounding intelligent, but the question is, can it be empathetic? And at least not yet, it can't. And: AI doesn’t have nefesh, it doesn’t have a soul, it can mimic emotions, but not spirituality. He also said that he could spot the text as a machine’s, even if they had not: there were, in his words, distinctive markers that aren't how many rabbis would speak. His summary of his own job was: Being a rabbi is not about what I have to say on any given subject, it's about addressing the community's spiritual needs, which AI can't do yet.

Notice the two things this case shows at once. The congregation could not tell, and guessed it was the work of one of the most admired religious writers of their time. And the rabbi told them only afterwards, as a lesson. The experiment worked as a teaching tool precisely because, while he preached, they did not know.

Lucerne, 2024: an AI Jesus in the confessional

The third case is not a sermon, but it is the fullest test so far of a machine speaking in a sacred role. From 23 August to 20 October 2024, the Peterskapelle, a Catholic chapel in Lucerne, ran an art installation called Deus in Machina, made with the Immersive Realities Research Lab of the Lucerne University of Applied Sciences and Arts. An AI avatar modelled on traditional images of Jesus sat behind the lattice of the chapel’s confessional and answered visitors’ questions in close to 100 languages. The Catholic Church in the city of Lucerne’s own page describes the space as for four ears only, jedoch ohne Beichte, “but without confession”. It records more than 900 such conversations. The same page notes that world media interest was fuelled partly by the false report that the avatar heard confessions.

The Guardian’s report of November 2024 adds the numbers and the doubts. It gives more than 1,000 people who used the installation, a higher figure than the parish’s count of conversations. Visitors were told not to share personal information and had to confirm that they used it at their own risk. The theologian running it, Marco Schmid, told the paper: It’s not a confession. Of more than 230 people who gave feedback, he said two-thirds had found it a “spiritual experience”, which surprised him. Others told the church they could not talk to a machine. A local reporter found the answers, at times, trite, repetitive and exuding a wisdom reminiscent of calendar cliches. Schmid said Catholic colleagues objected to the use of the confessional, and Protestant colleagues to the use of the image.

What stayed with Schmid was the risk. The church had tested the avatar with 30 people before opening, and kept staff close by afterwards. Even so:

But of course we could never guarantee that he wouldn’t say anything strange.Marco Schmid, quoted in The Guardian, November 2024

That uncertainty is why he would not install it permanently: To put a Jesus like that permanently, I wouldn’t do that. Because the responsibility would be too great.

Three ways a congregation can be asked

Put side by side, the three cases are not really about whether the text was good. Many in Fürth found it flat. The Hamptons congregation mistook theirs for Jonathan Sacks. Lucerne got both reactions. What distinguishes them is what the people in the room had agreed to.

• Fürth told people in advance. The service was announced as an AI service, with a motto that said so. Everyone in the pews had chosen to come and see a machine lead worship. Even so, they stood for the creed.
• The Hamptons told people afterwards. For the length of a sermon, a congregation received a machine’s words as their rabbi’s. Then he told them, as a teaching point.
• Lucerne asked each person to opt in. A warning not to share personal information, a confirmation that they used it at their own risk, a clear statement that this was not a sacrament, and support close by.

None of the three is the case most clergy are actually asking about. That is the ordinary Sunday: the preacher uses a chatbot to draft some or all of the sermon, edits it, and preaches it without saying anything. Nobody has published an account of that from inside a congregation that I could open. What does exist is guidance written for exactly that case.

What written guidance says

I could not open a denomination’s own binding rule on AI-written sermons. The closest thing I found is Guidance for Ministers and Scholars on the Use of Generative AI from Virginia Theological Seminary and The General Theological Seminary, two Episcopal seminaries. It came out of a two-day conference on authorship and AI, and says it hopes Episcopal schools and dioceses will find it helpful. It is a permission with conditions:

Gen AI may be used judiciously and prayerfully for ministry tasks including sermon preparation, pastoral correspondence, and liturgy.Virginia Theological Seminary and The General Theological Seminary, Guidance for Ministers and Scholars on the Use of Generative AI

The conditions are specific. Where AI was used in public worship, it suggests a short general acknowledgment, offering the wording AI tools assisted in the preparation of today's service. The minister keeps full responsibility for the content, including its theology. Confidential pastoral material from counselling, hospital visits or confession must never go into a public AI platform. The congregation or its elected leaders should be told once a year how the church uses AI. Ministers with documented disabilities or learning differences, or writing in a second language, may use AI more extensively, and that use needs no public disclosure, since disclosure could effectively disclose the disability. Its first question for anyone writing with AI goes to the heart of the consent problem: Would my audience respond negatively if they knew Gen AI was used, and would they be justified in thinking this? Its principle is one sentence: Whatever tools we use, we remain the authors of our actions.

The opposite position has a written form too. The Living Church reported on 8 June 2026 on the Saint Dunstan Pledge for Preachers, written by the Rev. Emily García, an associate rector in Lexington, Massachusetts. The paper says about 150 people had signed. As the paper quotes it, the pledge reads: At no point in the brainstorming, planning, drafting, or writing of my sermons will I use ChatGPT, Claude, Copilot, Gemini, or any other generative AI of any type available to me. It goes on: I will use my own words. I will quote words written or spoken by other people, not produced by generative AI. The same report notes that the pledge still allows non-generative tools such as voice memos, spelling correction and text-to-speech. Neither document is a church law. They are the two positions most clergy are choosing between, in writing.

The case for letting it help

Put at full strength, the case for is not that a machine should preach. Almost nobody argues that, including Simmerlein, whose service was an experiment. It is three narrower claims.

Drafting is not preaching. Preachers have always written with help: commentaries, lexicons, other people’s sermons, the ideas of colleagues. The seminary guidance makes the comparison itself. It lists speech writers, copy editors, writing tutors, research assistants and conversation partners as help that already shapes a piece of writing, and notes that Authorship is shared. On this view, a chatbot is a faster version of a commentary, and what matters is what the preacher does with it. The Rev. Peter Levenstrong, quoted in the Living Church report, disagrees with the pledge while sharing its view of preaching: What kind of machines helped with the preparation is a smaller question than whether the sermon tells the truth, with authenticity and a willingness to listen for the Holy Spirit.

Access. For some preachers the tool removes a barrier rather than a burden. The seminary guidance names neurodivergent people and people writing in a second language, and calls this use important and good. A pastor who preaches every week in a language she learned as an adult, or who has dyslexia, is doing a different task from a fluent native writer, and the argument is that her congregation is better served if she can spend her effort on the substance.

Labour. A sermon every week is a heavy demand on top of visiting, burying, counselling and administration, and many ministers carry it alone or alongside another job. Outsourcing is also not new. Jack Franicevich, quoted in the Living Church, used to write copy for megachurch pastors, and he watched his employer lose clients as those pastors started using ChatGPT instead. If a congregation already accepts a preacher who uses bought material or a human writer, the question is whether a machine is really a different kind of help or just a cheaper one.

The case against

Put at full strength, the case against is not that the text will be bad. The Hamptons sermon was good enough to be taken for Jonathan Sacks. It is three claims about what a sermon is.

Authenticity: a sermon is an act, not a text. The Rev. Ruthanna Hooke, a professor of preaching quoted in the Living Church, puts it in theological terms: I am opposed to the use of generative AI in sermons because my conviction is that preaching is an incarnational and sacramental event in which the Holy Spirit works with and through our humanity to speak God’s Word. Müller-Hansen’s “the throat is the soul” and Franklin’s “AI doesn’t have nefesh” make the same point from a Lutheran and a Jewish pulpit. On this view the struggle with the text during the week is not a cost to be cut. It is where the sermon comes from. The seminary guidance, which permits AI, still asks that users first engage in productive struggle. Bird Treacy, a licensed preacher who signed the pledge, adds a communal version: using AI means that I get to have this conversation with a computer full of language instead of with colleagues and fellow believers.

Disclosure. Listeners assume, unless told otherwise, that the person in the pulpit wrote what they are saying. The seminary guidance draws the line at exactly that assumption: an administrative text where nobody is surprised by assistance is less ethically complex than a task where the recipient is assuming human agency and thought. A sermon is the clearest case of the second kind. The Hamptons case shows that a congregation cannot detect the difference on its own. If they cannot tell, only the preacher can tell them.

Consent and responsibility. A congregation that comes to hear its minister has agreed to be addressed by a person who can answer for what is said: someone they can question afterwards, who knows their dead and their children. Simmerlein said something similar in Fürth, about the sermon happening between preacher and congregation. In Fürth people chose an experiment. In Lucerne each person confirmed they accepted the risk. On an ordinary Sunday nobody is asked. And Lucerne showed the limit of responsibility plainly: the person in charge could not guarantee what the machine would say, so he would not leave it running. A preacher who reads unchecked machine text from a pulpit is in the same position, with fewer safeguards.

Where the answerability goes

This site has argued elsewhere, in Taking Its Word For It, that when a machine enters a chain of testimony, answerability does not disappear. It lands on the next human along. A sermon is that problem in its sharpest form. The congregation is not only receiving information. It is being told, by someone it trusts, what that person believes about God. The seminary guidance and the pledge both accept this. One says the preacher stays the author of whatever the tool produced. The other says the preacher can only be sure of that by not using the tool at all.

The churches’ longer texts on AI, read in A Pale Reflection, say AI is a human-made tool and that responsibility stays with people. They do not settle this narrower question. You can hold that AI is a tool and still disagree about whether a sermon is a task a tool should do. And whether a machine could ever mean what it says, the question behind Franklin’s “nefesh”, is the subject of Which Soul Do You Mean?

Questions a congregation could ask

This essay does not decide the matter. Reasonable people who take preaching seriously have come down on both sides, in writing, in the past year. But the cases above make some questions concrete enough to ask your own minister, or yourself if you preach:

• How much? Checking a translation of a Hebrew word, suggesting an outline, drafting a paragraph, and writing the whole sermon are different things. The seminary guidance and the pledge draw the line in different places; neither treats them all the same.
• Would we be told? The seminary guidance suggests a short line in the service and a yearly word to the congregation. The Hamptons shows what happens without one: nobody can tell.
• Who checked it? Everyone quoted here agrees that the preacher answers for the theology, the facts and the quotations, whatever tool was used.
• What went in? Nothing from counselling, a hospital visit or a confession belongs in a public chatbot. The seminary guidance says so without qualification.
• What is the week for? If the hours spent with the text are part of what makes a sermon true, time saved is time lost. If they are a burden that keeps a pastor from their people, time saved may be time given back. Congregations may reasonably want to know which one their minister believes.