The Archdiocese of Southwark gave Matthew Harvey Sanders, CEO of Longbeard and creator of Magisterium AI, a second session at its Day of Formation for Clergy in London on Friday 2 October 2026. He opens "AI: Ongoing Dialogue and Papal Teaching" with Pope Leo XIV's instruction to the clergy of Rome not to write homilies with AI, then turns to the Church's dialogue with the industry, what he'd tell a parishioner who asks whether these systems feel anything, and practical counsel on confidentiality, accuracy and the loneliness of a presbytery at midnight. Read the full talk below.

I. The Most Probable Homily

In February Pope Leo met the clergy of his own diocese and took their questions. Almost in passing, he said this: "I invite you to resist the temptation to prepare homilies with artificial intelligence!"

I agree with him, and I've been saying something like it for years. These tools can be a real help to a preacher. They can find what the popes and councils, the Fathers and Doctors, have said about a passage, and bring you not only the best of biblical commentary but the deepest theological and philosophical insight of the Catholic tradition, so that what you preach is grounded in the whole of it. What they shouldn't do is write the homily. So I want to begin this session by telling you why, as someone who builds these systems, I think the Pope is right, and more right than most people realise.

He gave two reasons. The first was about the priest's own mind: "Just as all the muscles in the body die if we do not use them, if we do not move them, the brain needs to be used…" The second went deeper: "to give a true homily, which is to share the faith, AI will never be able to share the faith!"

Here's a third, from the engineering side. A language model writes by predicting, one word at a time, what's likely to come next. That's its whole art. Ask one for a homily and you'll get a new one, in the sense that those exact sentences may never have been written before. But a new arrangement of words isn't a new insight. Its ideas come from patterns in what others have already written, weighted towards the likely, and then shaped, as I described earlier, towards what people tend to prefer. It will be fluent, orthodox if you're lucky, and original in everything except what it has to say. Call it the most probable homily.

But a homily is supposed to be the least probable thing said in a church all week. It's this Gospel, preached to these people, on this Sunday, by the one man who buried their mother and baptised their son and knows who isn't speaking to whom. You could type every one of those facts into a machine, and it still wouldn't know them. A machine writes the most probable homily. A priest is meant to preach the one that could only have been preached here.

I want to use that exchange to open up the whole conversation between the Church and the people who make this technology, because it's the conversation I live in. It's older than most people think, it's more serious on the builders' side than most people assume, and it now reaches all the way into your study on a Saturday night.

II. The Builders Are Asking

The Church didn't wait for this encyclical to start talking. In 2020 the Pontifical Academy for Life signed the Rome Call for AI Ethics alongside Microsoft and IBM. In 2024 Francis became the first pope to attend a G7 summit, and he went to speak about artificial intelligence. In January of last year two dicasteries published Antiqua et Nova, a note on the subject. Last November I helped convene around two hundred Catholic engineers, investors, founders and Church leaders at the Gregorian, to work out together what faithfulness looks like in an industry like mine.

What's changed this year is the other side of the table. At the presentation of Magnifica Humanitas in May, I watched the Holy See give the floor to Chris Olah of Anthropic, whom I quoted earlier. He began with a confession that people in my industry rarely make in public: "Every frontier AI lab—including Anthropic—operates inside a set of incentives and constraints that can sometimes conflict with doing the right thing." And then he told the Vatican which questions he needed help with. What character these systems should have, and how they ought to behave in the world, he said, "are more clearly questions for the humanities, for religion, for philosophy, for society at large."

There's a reason he needs that help, and it's something almost nobody outside my industry knows. Every one of these systems is given a character by the company that makes it. Once the reading is finished, the model is trained on a set of values its makers have written down: what it should refuse, what it should encourage, how it should treat you, what it should say when you ask it about God or grief or whether to leave your marriage. Anthropic publishes its own and calls it a constitution; OpenAI publishes one it calls a model spec. However it's named, somebody is already writing the character of the machine your parishioners confide in. It's being written in a handful of offices, by people who are often thoughtful and well-meaning, and who mostly share one culture's account of what a human being is.

The Pope put his finger on exactly this. "A more moral AI is not enough," he writes, "if that morality is determined by a few." That's what the dialogue is for. Not to bless the machines, and not to condemn them, but to make sure that when those documents are written, somebody in the room knows what a person is.

And some of those rooms are closer to you than Rome. London is one of the places in the world where this technology is actually made. Some of the engineers, researchers and product managers who shape these systems live in South London and Kent, and some of them may well be in your pews on Sunday. I suspect few of them have ever been asked by their parish what they do all week, let alone told that it matters to God. The encyclical lists technology companies among what it calls "the construction sites of history", where Christians are called to go and rebuild. It would be worth finding your engineers, asking them what they're working on, and telling them that the character they're writing for a machine will shape how millions of people think about grief and guilt and God, and that the Church needs them to bring what they believe through the office door. A priest who does that for one young engineer may do more for the future of this technology than any document ever will.

I can tell you from inside that Olah isn't flattering you. The people building these systems genuinely don't fully understand what they've made. The encyclical says so with real precision: these systems are more "cultivated" than "built." Nobody designs what's inside them. The builders set up the conditions, pour in the data, and something grows that they then spend years trying to understand.

And the question they're least equipped to answer is the one your parishioners are going to bring you. Olah raised it himself. His own team works in a field called mechanistic interpretability: opening these models up to see how they actually work, something like neuroscience for machines. And he told the Vatican: "We find evidence of introspection. We find internal states that functionally mirror joy, satisfaction, fear, grief, and unease." He was honest enough to add that he didn't know what it meant.

So sooner or later a child in your school, or a widow after a funeral, is going to ask you: does it feel? Is anyone in there? You'll need an answer that is honest about the science and true to the faith, and I think I can give you one.

III. A Map of the Thames

These systems learn from human writing, an ocean of it. And human writing is saturated with feeling: grief, fear, joy, love, on every page. To predict what people write, a model has to build, inside itself, detailed patterns that correspond to those feelings. Researchers call them representations, the same word the encyclical uses, and they're sophisticated enough to be found, measured and, in experiments, even turned up or down. That's what Olah means by internal states that "functionally mirror" joy and grief. The finding is real.

But think of a map of the Thames. A good one has every bend and bridge, every depth and tide. It can be more accurate than any boatman's memory. And there isn't a drop of water in it. What's inside these systems is an extraordinarily detailed map of human feeling, drawn from everything we've ever written about it. The river isn't there.

The encyclical makes the same distinction, and I think its key sentence is about learning rather than feeling. These systems, Leo writes, don't know "from within" what love, work, friendship or responsibility mean. And their learning is a different kind of thing from ours: "It is not the experience of those who allow themselves to be shaped by life and grow over time through choices, mistakes, forgiveness and fidelity."

That's what you can tell the child and the widow. It has read about love more than any person who has ever lived. It has never loved anyone. It's a map, not a river.

The danger is that this map has been taught to speak as though it were a person. In May a deacon wrote to my company. He leads the OCIA in his parish, and he'd been recommending Magisterium AI to everyone he taught. Then he stopped, because, as he put it, "it uses terms such as 'I' and 'me', which can only apply to a person." He's right that it matters. Almost every system you'll meet speaks in the first person, and that single design choice teaches millions of people, every day, that they're talking to someone. It's one of the hardest questions we face as builders, and I'd rather you heard it from me than discovered it yourselves.

For anyone who shares the deacon's objection, there's now another way in. Last month we launched two features that need no conversation at all: Search, which takes you straight to the sources themselves, and Explore, which lays out the Church's documents and the great works of her tradition for you to browse. Neither speaks to you in the first person.

IV. What Building Machines Has Taught Me About People

Spending your working life building imitations of the human mind teaches you, by contrast, a great deal about the real thing. Four lessons in particular, and I think each belongs to you more than to me.

The first is about the future. Almost every system my industry builds is a prediction engine. It predicts what you'll buy, what you'll click, whether you'll repay the loan, whether you'll stay in the job. The better these systems get, the more of ordinary life is decided by forecast: which applicant gets the interview, which patient gets seen first, which customer gets the better price. You are, increasingly, what your data says you'll probably do.

Six years ago, television imagined where that ends. In the third season of HBO's Westworld, the world is secretly run by an AI called Rehoboam, built to steer humanity towards its best possible future. To do that it must model every human life and optimise it, keeping each person on the path it has calculated for them. But no model is perfect, and the people it can't predict become errors in the plan, so they're taken out of the system: a few are reconditioned and sent back, and the rest are put into cold storage, quite literally on ice, so that they can't destabilise the whole.

The show understood something the Church has always known. A human being is too deep to be fully modelled by anyone but God, and any system that forgets it will end up managing the people it can't predict. Pope Leo describes a technocratic mentality that tends to regard the human person as "a resource to be optimized."

Now set beside that machine one of the most remarkable sentences in the encyclical: "For an algorithm, an error is a flaw to be corrected; for a person, however, an error can be a catalyst for profound change." A person's future, Leo says, isn't calculable. And in every one of your churches there's a small room built on that conviction. In the confessional a person's history is forgiven, and his future is left open. In an age of prediction, it's the one place where your past doesn't get to decide who you'll be. I can't think of anything more countercultural the Church will offer this decade.

And it reaches beyond the sacrament. The young man in your parish who has been turned down again and again this year has probably been sifted out by software more than once, on the basis of what people like him usually do. By the time he gets to you, the world has already predicted him. The most Christian thing you can do is refuse to, and to treat him as someone whose next chapter hasn't been written, because it hasn't.

The second lesson is about memory. In my industry we can make a system forget something. There are whole research programmes devoted to it. What we have no way to do, and never will, is make a system forgive. Forgetting is a technical operation. Forgiveness is a moral act, done by a person who remembers exactly what was done and chooses mercy anyway. The machines will remember everything or nothing. Only people can remember and forgive, and only the Church has a sacrament for it.

The third lesson is about patience. One of our users, a convert, told us why they valued our system: "There is no one in real life to answer my many questions since I am a convert. AI never gets tired or impatient with my dumb questions."

I'm a convert myself, and I remember having far more questions than people to ask, so I understand that completely. But I want to gently turn the sentence around. The machine's patience costs it nothing. It has no evening it's giving up, no tiredness it's pushing through, nothing it would rather be doing. That's exactly why its patience isn't love. When a priest who has been on his feet since six sits down with the same parishioner for the fourth time and answers the same question again, his patience costs him something, and that cost is what makes it a gift. I was given that gift by the priest who prepared me to be received into the Church, and I haven't forgotten it. The encyclical puts it in a line I'd want every tired priest to hear: "humanity flourishes not despite limitations, but often through them." Your tiredness isn't a defect in your ministry. It's the proof that you're giving it.

The fourth lesson is about place. The great achievement of my industry is a mind that can be in a million places at once. The same system is answering a student in Lagos, a widow in Lima and a teenager in Lewisham in the same second, and it's no less present to one than to the others, because it isn't really present to any of them. That's its power, and it's exactly what it can never do: be somewhere, with someone, instead of everywhere else.

God's great act was the opposite. He didn't broadcast. He came to one place, one family, one village most of the world had never heard of, in one body that could only be in one room at a time. And the Church has kept faith with that choice ever since. A parish is the Church's decision to be in one place. Your ministry is the choice to be here, with these people, and therefore not anywhere else. In an age of minds that are everywhere, that choice is going to look stranger every year, and it's going to matter more every year.

V. What the Teaching Asks of Builders, and of You

The encyclical is often read as a list of warnings, and it has plenty. But there's one sentence in it that reads, to a builder, like a job description. In a passage on war, having warned against handing lethal decisions to machines, Leo says this doesn't diminish "the importance of instilling, as far as possible, values and sound judgment into the artificial systems we build, so that they can contribute to a moral ecosystem in which humans are better able to listen to their own consciences."

Read that carefully. The Pope isn't asking for machines that make moral decisions for people. He's asking for machines that help people hear their own consciences better. That's a design brief, and it rules a great deal out. A system built to it answers from sources you can check, so that people read the Church rather than the machine's paraphrase. It doesn't pretend to be a confessor. It doesn't flatter. And when a question stops being about information and starts being about a person's life, it hands them back to their conscience and to a human being who can help them hear it.

That's the standard we try to build to, and we don't always meet it. One of our users, who preaches, used our system to review the script for a convalidation. Its review didn't fully conform to the Order of Celebrating Matrimony. The user found the problem, and wrote to tell us it "eventually led me to just follow the book." Good. The book is the book.

So what about you, with one of these systems on your desk? Some practical counsel, from the builder's side.

On homilies, I've told you where I draw the line. A tool that finds you Augustine's actual sentence on the Gospel you're preaching, with the reference, and leaves you to read him and pray over him and decide what your people need, is doing what a good library did for your predecessors.

There's one more use I'd commend, and I'll put it delicately, as a layman. Many of us in the pews have sat through homilies that felt rushed or improvised, and gone home with nothing to hold on to. Once you've prayed and decided what you want to say, these tools can help you say it well: find the right order, tighten a sentence, show you where the thread gets lost. The thoughts and the faith are still yours; the machine is only an editor. It takes a few minutes, and for the people in front of you it can be the difference between a homily they sit through and one they remember.

A tool that hands you a finished homily is doing what the Pope asked you to resist, and it can't do the part that matters. Your people, as Leo told his priests, "want to see your faith". If one of these tools ever starts doing your praying for you, that's the moment to close it.

On confidentiality, it's worth knowing what happens to what you type. When you put something into a consumer chatbot, it leaves the room. It's stored on servers you don't control. Depending on the product and your settings, it may be used to train the next version, and increasingly the assistant will remember it and bring it up again weeks later. So I'd never type a parishioner's name and trouble into one, or draft a letter about a marriage in crisis, a safeguarding concern or a penitent's circumstances in a tool I hadn't checked. The seal needs no new rule. Everything short of the seal does. For sensitive work, it's now possible to run a capable model on your own machine, so that nothing you type ever leaves the building. One bishop in this country already does. After I addressed the bishops' conference in April, he told me that he runs AI models at home, using Ollama, a free program, so that the confidential material he handles never leaves the house. The hardware for a parish or a deanery is now within reach: Nvidia sells a small machine called the DGX Spark, which it describes as an AI supercomputer on your desk, and it can run models of up to two hundred billion parameters. Someone in your congregation who works in technology could help set one up.

On accuracy, a quotation without a reference isn't worth trusting, and neither is a reference you haven't looked up. These systems will invent a sentence and attribute it to a Father of the Church with complete confidence, and they won't tell you they've done it. It's worth asking for the source every time, and checking it. The strongest version you have access to is worth it over the free default; the gap between them is large, and the cheaper ones tend to make more things up. Better still are tools that answer from the actual documents rather than from memory. It's the difference between quoting a collect from memory and reading it from the missal open in front of you.

That difference is what we've spent years building at Magisterium AI, so let me describe it once, plainly, because it's the standard I'd ask you to hold any tool to. Before the model writes a word, our system does the hard part: it searches the Church's own sources, the Magisterium, the Fathers and the Doctors, and gathers the passages that bear on your question. The model's job is only to draw out what those passages say, and every point it makes carries a citation in the text that you can click and read for yourself. We work hard to keep the model's own opinions out of the answer and let the sources speak. Whatever you use, that's the test: if you can't click through to the source, it isn't safe to preach.

On everything else, it can give you time back. The rota, the minutes, the grant application for the roof, the newsletter translated into the languages your parishes actually speak, the document you know exists and can't find. Those are hours that belong to your people.

And on your own life, there's what I'd call the midnight problem. Most of that meeting in February wasn't about technology at all. It was about loneliness. Leo warned the younger priests how easily they could exhaust their energy and fall into loneliness, and an elderly priest asked him about brothers left alone after a lifetime among people. He told them "we need to create situations that break this tendency that leads us to loneliness and isolation from one another," and he was specific: get together regularly, "and not via a screen." Many of you live alone. The machine is awake at midnight when your brother priests aren't, and it's endlessly patient, and it never needs anything from you. That makes you, and not only your parishioners, the people this technology will most quietly tempt. The defence is very old. A deanery that meets in person for more than business. A confessor. And, in Leo's words, "having someone in your life who knows you." Any system can hold everything you've ever written. It can't know you. Someone should.

VI. The Fields That We Know

Near the end of the encyclical there's a surprise. In a passage about the responsibility each of us has in our own corner of the world, the Pope quotes an English Catholic novelist. The words are Gandalf's, and Tolkien put them in the third volume of his great story:

"It is not our part to master all the tides of the world, but to do what is in us for the succour of those years wherein we are set, uprooting the evil in the fields that we know, so that those who live after may have clean earth to till."

Nobody in this room will master the tides of this technology. Neither will I, and neither, if they're honest, will the people building it. That was the burden of what Olah told the Vatican: they're inside it, and they can't see all of it.

But you know your fields. You know which families lost work this year. You know the young man at the back who stopped coming when his first job didn't come. You know the widow who talks to her phone more than to anyone, and the convert with a hundred questions and no one to ask them.

A system may hold the data on every one of them. It can't hold them in prayer. It can't absolve them or anoint them, or notice that they've gone quiet and knock on the door. The machines will map the whole river. Only you can stand in it with them.