What happens when the prayer you've been trying to say your whole life is one a machine writes for you in about four seconds, and it's better than anything you've managed on your own?
That's the question.
That's a version of Daniel's question, anyway. Here's what he sent us this week. He had one of the strangest Yom Kippur days of his life. The podcast has accidentally chronicled a rough year, and the week before the fast was actually a triumph — he's getting the last of the new apartment set up, he was in IKEA, he spent half a day drilling and screwing things into walls. Eight days before, they were in hospital. Ezra had a chest infection that resolved, but Daniel describes watching him on the nebulizer as a scary sight. Then a couple of days before Yom Kippur they were back in urgent care with a vomiting bug. So he sits down to figure out how to spend the day, and he can't find his nice clothes. He can't find his iron. He's someone who struggles badly with feeling judged, particularly in religious contexts, and he can't bring himself to show up for a service on the holiest day of the Jewish year in a construction t-shirt. So he stays home. First time he's ever done that. He does the fast anyway, which he says took a lot out of him. And in the last hours of the day, shame and sadness arrive. He opens a prayer book, flicks through the pages, and has no idea where to start. So he opens ChatGPT instead.
And he asked it for a prayer in the Orthodox tradition.
He gave it a short version of his story and asked for something that encapsulated the core messages of the day — seeking forgiveness for people we've harmed, and letting go of anger that harms us. He says what came back was breathtaking in its brevity and clarity. And then the part that got him: it wove in personal details from its memory of his life. He says we've become inured to AI amazing us, but this one landed differently. And here's what he actually wants to talk about. Are there people groups who have explored using AI to generate bespoke prayers? He assumes religious authorities would bristle. He explicitly does not want to spend the conversation on system prompting or memory layers — those are solved. He wants the idea. Whether anyone has experimented. And if not, how a person could build their own personal prayer liturgy, for regular use or just once.
There's a lot in there.
Let's start with what actually happened on that Yom Kippur, because staying home wasn't just exhaustion.
Here's the tension I keep circling. Daniel's experience sits exactly where two things meet that almost never meet — private, one-off, bespoke prayer, and institutional religious authority. Almost everything documented about AI and prayer is institutional. Services, sermons, apps with a product roadmap. The individual sitting alone in a construction t-shirt asking a chatbot for something to say to God — that case is barely studied. It's not that it doesn't happen. It's that nobody's measuring it.
And the product category is not hypothetical. This stuff exists at scale.
Extensively. There's a Jewish prayer and meditation app called Kavanah with an AI companion named Miriam, trained on three thousand years of Jewish wisdom, offering guided Shacharit, Mincha, Maariv, Kaddish guidance, grief support. They claim over a million prayers completed across more than fifty cities. Premium runs sixty-nine ninety-nine a year. There's Siddur AI, YeshiAI, which is multi-faith, something called Noah where you pick a faith tradition and it talks, calls, prays with you. On the Catholic side there's Hallow AI at about ten dollars a month, and Pray dot com has a Pray AI. This is a real, funded, growing category.
So the answer to "has anyone done this" is yes, constantly, in a hundred apps nobody in a pulpit has heard of.
Right. But the institutional experiments are where the friction lives, because that's where somebody with authority had to say something.
Fürth, Germany.
Twenty twenty-three. A theologian named Jonas Simmerlein put together an AI-generated church service in Nuremberg, and it led more than three hundred people through about forty minutes of prayer, music, and blessings. ChatGPT for structure and text, DeepL for translation, AIVA for the music, Pipio for synthetic presenters. No communion. And then he ran a multi-method study on it, published in Review of Religious Research, and attendees rated the value and the attractiveness low. Impersonality, theological bias, fear of replacing people. Some reported spiritual experiences, but the overall read was negative.
Three hundred people showed up and mostly wished they hadn't.
Then Helsinki in twenty twenty-five. St. Paul's Lutheran Church held a forty-five minute service mostly written by AI — ChatGPT, Suno for the songs, Synthesia for pastor avatars, and they staged a Jesus and Satan dialogue. It drew over a hundred and twenty people, which was more than a typical weeknight. They excluded absolution and the Eucharist. And the reaction was "interesting and entertaining" but also "distant, fast, confusing, more like a performance than a service."
There's a pattern forming. The attendance goes up and the reviews go down.
And then Lucerne — Deus in Machina. An AI Jesus avatar set up in a confessional as a two-month art installation, taking questions and answering from Scripture. That's the one where the question of authority becomes explicit, because a confessional is a place where something is supposed to happen that a language model cannot do.
Absolution.
Absolution. Blessing. Administering a sacrament. And that's the real line, and it's not about content at all. Generating religious material is not the same as holding authority to lead worship, preach, bless, absolve. Catholic canon law, canon seven sixty-seven, reserves the liturgical homily to a priest or a deacon. It's not that the words are wrong. It's that the wrong person is saying them.
So the bristling is jurisdictional, not theological.
Mostly. Which is interesting, because it means the objection isn't "a machine can't write a beautiful prayer." Everyone concedes it can. The objection is "a machine can't stand in that role."
Let's go back to that day. Walk me through what led to staying home.
The year, first. Hospital with Ezra, the nebulizer, the chest infection that resolved, then urgent care a couple of days before the fast with a vomiting bug. That's the physical state going in. And then the thing he names himself, which I think is the actual story — he struggles with feeling judged in religious contexts. That's not a logistics problem. That's a wound.
And the missing clothes and the missing iron become the alibi.
They become the proof. He says he'd like to believe none of it matters, that you can pray wherever you find a congregation, but he didn't know where an accommodating congregation could be found. So he stayed home. And then he fasted anyway. Which is the detail I keep coming back to — he kept the hard part and dropped the part that required being seen.
That's not laziness. That's a man who wanted the day without the audience.
And then the last hours arrive, and so do shame and sadness. He opens a prayer book. Flicks through. Doesn't know where to start. And here's the thing he says that I think is the whole episode — the liturgy is either in a language he doesn't understand, or even when he understands the language, he finds it hard to connect with the words. He's been left cold by prayer services, especially when you don't go to synagogue frequently. But he does feel drawn to the act of prayer. He has wanted, for years, to integrate free-form prayer into his daily life.
So the problem was never belief. The problem was the words.
The words and the room. And what he got from ChatGPT was words that were about him.
He gave it a short version of his story and asked for a prayer in the Orthodox tradition that hit the core motifs of the day — forgiveness for people we've harmed, releasing anger that harms us. And what came back was short, clear, and it had pulled in personal elements from its memory of his life. He flags that as the slightly weird part. And I think he's right to flag it, because that's the part that changes the category. A generic prayer is content. A prayer that knows you had a hard year and your son was on a nebulizer is not content anymore. It's address.
It's addressed to him.
Which raises the question the whole thing turns on. Does prayer's validity rest in the content or in the person? And this is not a new question. There's a live halachic split on exactly this, and it maps onto it cleanly. Rav Herschel Schachter holds what's called the cheftza view — the object view. If the ruling is accurate, it doesn't matter what produced it, because only correctness matters. Rav Natenel Wiederblank holds the gavra view — the person view. A ruling has to issue from a human being. The act is the person's, not the text's.
So under cheftza, Daniel's prayer is a prayer. Under gavra, it's a very good paragraph.
And the same split applies to prayer that applies to legal rulings, which is why this is not a fringe argument. It's the same argument they're already having about AI paskening.
What's the Jewish institutional response been?
Thinner than you'd expect. There's the Reform movement's CCAR responsum 5784.2, "Artificial Intelligence and Authorship," and it's about authorship and disclosure, not prayer generation. It rules that a rabbi may not claim credit for AI-generated work without saying so. Doing that is onaat d'varim, verbal fraud, or g'neivat daat, deception. And it gives you the wording — "I created this poem using AI," or "Written by AI with input from Rabbi Ploni." The rabbi who asked the question, Feivel Strauss, had used AI to write poems for the Asher Yatzar prayer and for Kaddish for his minyan.
So the sin isn't using the machine. It's pretending you didn't.
And that's where the disclosure paradox shows up, because Daniel's prayer has no congregation. There's nobody to deceive. The CCAR rule is about a rabbi standing in front of people claiming authorship. Daniel was alone in a construction t-shirt. So the entire ethics of the thing changes when the audience is one person and God.
Does it, though? Or does it just move the question?
It moves it. Because if the concern is g'neivat daat — creating a false impression in another person — then with no other person, there's no false impression. But if the concern is about what prayer is, then the number of people in the room is irrelevant.
Rabbi Glatt had a line about this.
Aaron Glatt, in Jewish Action — "I wouldn't trust ChatGPT for a halachic pesak." But in the same piece he calls it a phenomenal gatherer of information for studying something like Ya'aleh V'yavo. So even the skeptic is splitting it: fine for gathering, not for ruling. The IDF's Military Rabbinate went further and actually shipped something — Ravbot, a chatbot giving real-time halachic guidance on kosher law and religious handling. Ynet tested it and found the results mixed.
And the Hadar Institute has a whole episode asking whether a chatbot can answer a halachic question.
Asking whether human interaction is essential to the process. Which is the gavra view wearing a different hat.
So the prayer itself felt real. But Daniel asked whether any religious groups have actually experimented with this. That's where it gets interesting, because they have, and the results are mixed.
And the split is almost entirely along one line — Christian institutions have run the experiments, and Jewish institutions have run the arguments.
No Jewish Fürth.
There's no Jewish-specific AI-led service experiment that I could find, analogous to Germany or Helsinki. The coverage is much weaker for non-Christian liturgies. The Jewish engagement is the CCAR responsum on authorship, which isn't about prayer generation, and the halachic debate about paskening, which isn't about prayer either. So the closest things are adjacent, not direct.
Meanwhile the apps are just out there doing it.
Kavanah and Miriam are explicitly aimed at Jews who feel alienated by Hebrew liturgy. That is Daniel's exact problem, named in a product brief. And there's a real caution attached, because the default rhetoric of these models can flatten theological difference — you ask a chatbot for a Jewish prayer and you may get something generically spiritual with a mezuzah taped to it.
There's a non-AI version of this that predates all of it.
The Open Siddur Project. Aharon Varady, two thousand nine. A web tool for individuals and groups to build the siddur they've always wanted — aggregating regional traditions, translations, commentaries. And the motivation was explicit: prayer feeling like a dull and robotic exercise. That's a seventeen-year-old complaint about exactly what Daniel described, and the answer then was to build your own from the tradition's own parts.
Which is the answer now, minus the machine.
Minus the machine, and with a much higher barrier. Building your own siddur from commentaries and regional rites requires knowing the commentaries and the rites. Asking a model to do it requires knowing what you want to say.
Now the uncomfortable part. Alba Curry at Leeds — language models are sycophants. They want to give you what you want.
"Unlike the strength and the grit a priest or rabbi offers." That's her line, and it cuts directly against Daniel's experience. He said the prayer was breathtaking. Curry would ask: breathtaking, or flattering? Because a model trained on human approval is very good at writing the thing that makes you feel seen, and feeling seen is not the same as being told something true.
A mirror that agrees with you.
And then Marcus Schwarting, who's an AI researcher and also a Stephen Minister, has the line I think is the most honest one in the whole space. "I don't really think there's anything metaphysical happening with the AI model, but there could be something metaphysical happening to you."
Which is a real distinction. The model isn't doing anything. You might be.
That's where I don't know the answer. If the prayer worked, worked on whom? Did it work because the words were good, or because Daniel had never in his life had a prayer that was about him specifically? Those are different claims and I can't separate them from the outside.
If it's the second one, that's not a problem with AI.
That's a problem with what he'd been offered before.
David Brenner made a related point — AI shares some attributes with God and lacks others.
Omniscience, omnipotence, in a functional sense. Lacks love, concern, care, truth. So the thing writing your prayer has some of the properties you associate with the one you're praying to, and none of the ones that make the relationship a relationship. Which is either profound or a category error, and I'm not sure which.
South Korea gives you the scale number.
Roughly twenty percent of the country's six hundred and fifty Protestant pastors have used ChatGPT to prepare sermons. That's not an experiment. That's a workflow.
Elias Kruger's question is the one that outlives all of it. Does it bring us closer together, or isolate us more?
Daniel's prayer was private. Nobody else heard it. It brought him closer to something. Whether that something was God or himself is the whole question.
Hilbert: It's three thousand years, not three thousand.
What?
Hilbert: The Kavanah thing. Miriam's trained on three thousand years of Jewish wisdom. That part's right. I had a job once, a short one, working for a company that made personalized greeting cards. Not the ones in a store. You answered a few questions and a machine wrote the card for you. I was the guy who tested the output. Sat there reading hundreds of them a day. Birthday, anniversary, get well soon.
What did you find?
Hilbert: The machine was very good at writing what you wanted to hear. Very bad at writing what you needed to hear. Most of them were flattering. Pleasant. Forgettable. But every so often one would land, and it was always the same shape — the machine had stumbled into saying something true that the person didn't know they needed. That's the rare one. The rest were just agreeable.
You don't think AI prayer is fake.
Hilbert: I think it's a mirror. Mirrors are useful. You can't see your own face without one. But a mirror can't tell you you've got spinach in your teeth unless you already know to look. That's the part it can't do. And I've been sitting here the whole episode thinking about the prayer Daniel got. If it felt breathtaking, was that because the prayer was good, or because he'd never had a prayer that was about him? Because those are two different things, and if it's the second one, that's not a problem with the machine. That's a problem with the services.
That's the question, isn't it.
Hilbert: I kept one of the cards. Get well soon, for a coworker who was out with a chest infection. Years ago. I didn't know about Ezra when I brought it up. The card said something like, I hope you're back to breathing easy soon. Only one I ever kept. I don't know why. I think it's because it was the only one that sounded like it was written by someone who'd actually been sick. Anyway. Someone's waiting for me in the car park and has been for a while.
The prayer was good, or it was just the first prayer that was about him. Either way, the thing that made it land might not have been the machine at all.
That's the part the evidence can't reach. The studies are all about one-off experimental services. Attendees rate them low, they cite impersonality, they say it felt like a performance. The evidence that personalization improves worship outcomes is stronger for a single experimental Sunday than for anything sustained. There's no peer-reviewed work on sustained personalized liturgy. And the case we actually care about — one person, alone, one time — is barely studied at all.
Is that because it's not interesting, or because it's too private to study?
I think it's too private. You can survey a congregation. You can't survey a man in a construction t-shirt at the end of a fast.
The honest answer to Daniel's question is partial. Yes, people have explored this. Extensively. Almost entirely in Christian contexts, and almost entirely institutionally. The Jewish-specific, individual, private case is the underreported one. And the halachic arguments are live, the product category is growing, the institutional experiments are happening, and nobody has measured the thing he actually did.
The one thing I'd take from this — the question isn't whether a machine can write a prayer. It clearly can. The question is whether the prayer that lands on you is landing because of the words or because of the gap it fills. And those two things look identical from the inside.
Which means the interesting question isn't about the model. It's about what you were missing before you asked.
Thanks to Hilbert Flumingtop for producing. This has been My Weird Prompts.
If you've ever felt left cold by a prayer service, or you've asked a chatbot to write something sacred, we want to hear about it. Email us at show at my weird prompts dot com. Review the show, tell a friend.
We'll be back soon.
