The Oxford University Catholic Chaplaincy asked Matthew Harvey Sanders, CEO of Longbeard, to speak in its series Faith and the Rise of AI at The Old Palace in Oxford on Tuesday 3 November 2026. The most valuable thing the students would take from Oxford, he told them, won't be what they know, since the machines will always hold more, but who they've become. In "After the Last Exam" he starts from a test that may be "the last academic exam we need to give to models", and sets out the Church's four areas of formation, the value of work done with the hands, and Newman's "definite service". Read the full talk below.

I. Ten Times Bigger, Ten Times Faster

"It's going to be 10 times bigger than the Industrial Revolution, and maybe 10 times faster."

Sir Demis Hassabis said that to the Guardian last year. He co-founded DeepMind, the London laboratory that became Google DeepMind. He shared a Nobel Prize for teaching a machine to predict how proteins fold, and he's now Alphabet's chief scientist. In my experience most people have never heard that sentence. Two of its words deserve a closer look, because together they describe the decade you're about to walk into.

Start with "bigger". The first Industrial Revolution mechanised muscle. It took the strength of a man or a horse and put it into an engine. What it couldn't reach was the mind, so the mind became the place to retreat to. For two hundred years the advice to a clever child has been to get out of the field and the factory and into the office, the laboratory, the courtroom or the surgery. Universities like this one were the road.

This revolution has no such place to retreat to. The machines are reaching the mind first, and, as I'll show you in a moment, the hands right after. Law, medicine, finance, consulting, research, the civil service: the professions this university has supplied for generations are made of exactly the kind of work these systems are being built to do. That's what "bigger" means.

Then "faster". The first revolution took most of a century to remake Britain, and even at that pace it was hard on ordinary families. Robert Allen, who taught economic history here at Oxford, gave one stretch of it a name: Engels' pause. On the figures he cited, output per worker in Britain rose by 46% between 1780 and 1840, while real wages rose by only 12%. Others have argued since that the gap was smaller. Either way, the people who lived through the greatest economic transformation the world had seen waited a long time to share fully in its fruit.

Slow as it was, that revolution gave people one mercy: time. The adjustment happened across generations. A handloom weaver's son could become a mechanic, and his grandson a clerk. Now run it ten times faster, and most of a century shrinks to less than a decade. The adjustment has to happen inside a single working life. The industries you're preparing to enter may be reorganised before you've finished your first job in one of them.

If Hassabis is right, much of what's coming could arrive while many of you are still in your twenties: the years in which people have always learned a profession or a trade, found their feet, and often started a family. Your parents could plan a career on the assumption that the ground under it would hold for forty years. You'll have to plan on the assumption that it won't.

I don't say this to frighten anyone. You can discount Hassabis a little, if you like, because his company sells AI, and you can discount me for the same reason. But he's also a Nobel laureate describing what he sees from inside the laboratories. I see the same thing in my own work, and I think he's right.

II. The Arithmetic

You'll hear the reassurance that the employment figures don't show much of this yet. In some ways they don't, and I wouldn't build an argument on them either. Short-term indicators lag, they're revised, and they measure a world that's already passing. Last quarter's figures can't tell you where a ten-year trajectory ends. So I'd rather put the case to you as a matter of logic, and let you test it.

Begin with what the machines can already do. Early last year two research groups, the Center for AI Safety and Scale AI, published a test called Humanity's Last Exam. Nearly a thousand specialists from more than 500 institutions in 50 countries wrote its questions. They drew them from the frontier of knowledge in their own fields, and anything the best models could already answer was thrown out. When it came out, no model scored above 10%. Today the best score on the independent tracker Artificial Analysis is 61.4%. The climb from one to the other took less than two years. Its makers saw where this leads. It may be, they wrote, "the last academic exam we need to give to models". Once the machines answer nearly all of it, there's little left to learn from setting them exam questions whose answers we already know. The test that matters becomes whether they can answer questions nobody has.

That test has begun. In September, OpenAI announced that one of its models had solved a version of the Navier-Stokes problem, one of the seven Millennium Prize Problems, and had checked the proof by machine. The Clay Mathematics Institute, which offers the prize, welcomed "the announcement that the Navier-Stokes problem has apparently been settled". But the written proof hasn't yet been checked independently. Researchers at King's College London and Cambridge have since argued that the written and machine-checked versions make different mathematical claims. Whatever mathematicians conclude, a machine is now doing mathematics at the frontier of what's known.

And not only at the frontier. Ken Griffin founded the investment firm Citadel, and for years he was one of AI's sceptics on Wall Street. In May he described something else inside his own firm: "work that we would usually do with people with master's and PhDs in finance over the course of weeks or months being done by AI agents over the course of hours or days." He went home one Friday, he said, "actually fairly depressed by this, because you could just see how this was going to have such a dramatic impact on society."

Now the logic. Firms compete. That isn't a moral judgement on the people who run them. It's the condition they work in. In America a full-time employee costs a company about $9,000 a month on average, counting pay and benefits, and a manager or finance professional more than $14,000. Firms like Griffin's advertise starting salaries of up to $300,000 a year for new PhDs, before bonuses. A top-tier AI subscription costs about $200 a month. Serious work with AI agents costs more than that in computing power, sometimes many times more, but the gap is still enormous. Suppose the machine can do a large share of a person's work, and do it in hours rather than weeks. A firm that keeps paying for the slower and more expensive way will be undercut by a competitor that doesn't. Its owners may be the kindest people on earth. They'll still lose, and when they lose, everyone they employ loses with them. Nobody has to want this outcome for it to happen. The market only has to allow it.

You don't have to wait ten years to see this at work. Last week Microsoft offered voluntary buyouts to 60,000 of its employees, more than a quarter of its global workforce, including staff here in Britain. This isn't a firm in trouble. Its shares reached an all-time high in October. So look at where the money is going. Microsoft, Alphabet, Amazon and Meta expect to spend more than $700 billion between them this year, most of it on data centres for AI. A buyout on its own doesn't prove that a machine is doing anyone's job. But a company at its peak is offering to pay a quarter of its people to leave, while it and its rivals pour money into machines. That's what it looks like when the market allows it.

Then the machines get bodies. On the first of September Elon Musk told the G20's innovation ministers that within ten years there'll be well over a billion humanoid robots, each about five times as productive as a person. He called that forecast conservative. Discount him as heavily as you like. A robot doesn't need sleep, or a weekend, or a pension. Some can already change their own batteries in about three minutes and go straight back to work. Set a machine that works around the clock, stopping only to swap its battery, against a person who works forty hours a week, and the arithmetic isn't close.

Musk drew the conclusion himself, in February. "Corporations that are purely AI and robotics," he said, "will vastly outperform any corporations that have people in the loop." And a little later: "this will happen very quickly."

You might expect that from Musk. You might not expect it from Bill Gates. Last month he told The Verge he isn't a doomer. But he also said: "There isn't some mountaintop of cognition that the AI does not achieve." And of the physical world: "if we don't choose to massively slow down or stop, then we will have robots that can build houses, grow food, do surgery, and none of the things that human labor provides to the economy will be necessary for most goods and services." His big concern, he said, is the road there: a transition of about twenty years, with "winners and losers and a lot of disruption".

The usual answer is that new jobs will appear, as they always have. Some will. But consider why they always appeared before. The machines couldn't follow us into the new work. When the loom was mechanised, the weaver's son went to do something the loom couldn't: keep the accounts, mend the engine, teach the next generation. This time the machine can follow us into most of the work we invent. It still needs far more examples than we do to learn something new. But its makers release a better version every few months, and it takes a person years to retrain for another profession.

Others will say: then regulate it, tax it, slow it down. In places we may need to. But the logic that binds a firm binds a country too. No nation can hold the line for long while its competitors automate. Wise policy can soften the road and share out what it produces. It can't make the road go away.

I may be wrong about the timing. I don't think I'm wrong about the direction.

So I'd put the question to you differently from the way you'll hear it at most careers events. Not "which jobs will be safe?" Very few will be, for very long. The real question is what a human being is for, once a machine can do the work. Gates sees that too. In the age of abundance, he said, "Humanity... will have to find a way to organize and have purpose".

That isn't a new question. It's among the oldest there is, and the Church has been answering it for two thousand years.

III. What We're Building

More and more, people will take that question to a machine first, along with every other hard question they have. Many already do. So it matters enormously what the machines draw on, who decides what they say, and where they send people next.

My company, Longbeard, exists so that when someone asks a machine about the things that matter most, the Church's answer is there: whole, accurate, traceable to its sources, and open to everyone. Everything we've built follows from five decisions.

The first was to make sure the Church's memory serves everyone, and not only the laboratories. The laboratories need books. In September the Guardian reported that historical texts from Oxford's Bodleian Library, digitised through a project with OpenAI, had been used, according to internal documents, to "populate the OpenAI training set". Anthropic, it reported, has spent tens of millions of dollars buying books and slicing off their spines so the pages can be scanned. The great libraries of the world are becoming raw material for a handful of companies, and the Church holds one of the greatest of them, much of it never scanned at all.

So in Rome, the Pontifical Gregorian University and Longbeard set up what we call the Alexandria Digitization Hub. Its robotic scanners work through the holdings of the pontifical university libraries, and they're gentle enough to turn the pages of fragile books without damaging them. No spine is cut. The books stay where they are, part of the Church's patrimony, and anyone with a free Magisterium AI account can read what we've scanned.

You could fairly ask why we don't simply let the big laboratories do it. They'd be faster, and they'd do it for nothing. But nothing is the problem. Whoever holds a collection sets the terms on which everyone else may use it, and we've set ours: open to all. Pope Leo puts the principle well in Magnifica Humanitas: "truth is a common good and not the property of those with power or influence."

The second decision was to make that memory readable. A scan is only a picture of a page. A person can look at it, but a computer can't search it, so much of what the world's libraries have already scanned is, for practical purposes, still lost. Vulgate AI, our library platform, turns those images into text that a scholar can search and a system can retrieve. A monastery's archive or a diocese's records can then be consulted in seconds, rather than discovered by accident. It's the least glamorous thing we do, and nothing else we do would work without it.

The third decision was to make the machine answer from the sources rather than from memory. There's an easier way to build a system like ours. You take a general model, pour the Church's documents into it, and let it answer in a confident voice of its own. It's fluent, and far more impressive in a demonstration. We decided against it. Ask Magisterium AI a question, and before the model composes anything, the system searches more than 120,000 documents for the passages that speak to it. Only then does the model write, and its job is narrow: to set out what those texts say, with every point carrying a citation you can open and read in context.

Not every fact lives best in prose. General models are especially likely to go wrong on small, hard details: a person's formal name, the date they were born, the place they came from. So Magisterium AI can also use tools and consult structured data, records kept for exactly those facts, and set them alongside the written sources. The point is to assemble the right context first, so the answer is as faithful as it can be.

It also has to weigh its sources. A conciliar decree, a papal encyclical, a theologian's commentary and a parish newsletter don't carry the same weight, and a general model answering from memory can't reliably tell which of them it's drawing on. An answer that averages a council with a comment thread isn't the Church's answer. It's nobody's. Ask why a good God permits suffering, and what you should receive isn't the machine's opinion. It's St Augustine's, and St Thomas Aquinas's, and St John Paul II's, each with the passage it came from, and each marked for what it is: a Father, a Doctor, a pope.

Some of you will have the obvious objection already. Isn't this still asking people to trust a machine? Yes, in the way you trust a good index or a footnote: to send you somewhere you can check. By our measure, a summary that leaves you no reason to open the source has failed.

The fourth decision was to build it to let people go. Much of our consumer technology is designed to keep you. Its success is measured in minutes on the screen. We've tried to build the opposite. When someone has gone as far with a machine as a machine can take them, the system should hand them on to people who can walk the rest of the way with them: a parish, a priest, Christians nearby. That's what Catholic Intelligence, our platform for parishes and dioceses, is for. It holds the local details: the parishes, the Mass times, the ministries that are active. So it can send each person to the place most likely to help them. Someone anxious about walking into a church for the first time may be better served by a quiet weekday Mass than by a packed charismatic one. And a bishop can see, anonymised and in aggregate, what his people are actually asking, so he can focus his teaching and pastoral care where they'll do the most good.

The fifth decision is the newest, and it's the one I'd most like to discuss with you afterwards. Almost every AI system people use is adapted from a small number of models made by a small number of companies, trained largely on the open internet and tuned to their makers' goals. When you adapt one, you inherit its idea of what a person is and what a person wants. So we're training a model of our own from scratch. It's called Ephrem. It's compact enough to run on a family's own computer, so the questions people ask in private stay with them, at home. It isn't released yet, it's still research, and our aim is 2027.

Why does it matter what a model is trained on, if the answers come out right? Because no model is neutral. Someone decides what it should value, and that someone is a company, deciding for a very large number of people at once. The most careful attempt I know of is the constitution Anthropic published in January for its model, Claude. It's a serious and thoughtful document. Its primary author, Amanda Askell, studied philosophy here at Oxford, and I met her at the Vatican in May, at the launch of Magnifica Humanitas. Its authors say they don't want "to assume any particular account of ethics". If there's a "true, universal ethics", they hope their model will be good by it. Where neutrality runs out, they aim for "value judgments that wide swaths of relevant stakeholders can feel reasonably comfortable with". And they're candid that they have "a commercial incentive that might affect what dispositions and traits we elicit". For a company serving the whole world, that's an honest answer. It isn't the Church's. The Church holds that there is a true, universal ethics, and that it has been handed on to us. It doesn't settle questions of value by what the most people can live with. We want a model formed in that inheritance from its first day of training: one that tells the truth when it isn't welcome, and sends you towards other people instead of keeping you for itself.

What do we hope all this will accomplish? Two things. The first is that the Church's memory becomes open to anyone. Think of a seminarian in Kerala, a catechist in Kinshasa, a nurse in Manila who has never been near a great library. Each of them should be able to ask what the Fathers taught, in their own language, and be shown where they taught it. Magisterium AI already has users in more than 190 countries. The second is that seekers find something solid to push against. People who would never ring a presbytery bell will put their hardest questions to a machine. Often they'll find that their objection was raised centuries ago, and that a saint or a scholar answered it.

I've said in public that I wish Magisterium AI didn't need to exist, and I meant it. In a better world, every seeker would have a friend who knew the faith well enough to answer, and a priest with time to sit down. We built it because that world isn't the one we have, and because the questions won't wait for it.

But all of it has a limit, and I want to be plain about that, because it's the most important thing I'll say tonight.

A machine can play a real part in forming someone. It can tell the stories of the saints, and stories have changed lives before. In a Jesuit house the story is a familiar one. Ignatius lay at Loyola with a leg broken by a cannonball. The house had no romances, so he read a life of Christ and a book of saints' lives, and one thought kept coming back: "St. Dominic did this; I, too, will do it." The book played its part. What moved him was a man who had lived it.

A machine can tell you about a witness. It can't be one. "The Christian faith is not a 'religion of the book'," the Catechism says. Christianity is the religion of the Word of God, "not a written and mute word, but incarnate and living." We can put the whole memory of the Church in front of anyone who asks. What we can't build is what that memory is for: a life in which the faith is lived. That takes a person, and no machine is one, ours or anyone else's.

IV. Who You Become

That brings me to you. No laboratory can build a witness, but each of you can become one.

If the arithmetic I've set out is even roughly right, the most valuable thing you take from this place won't be what you know. The machines will always hold more. It'll be who you've become.

Few institutions have thought as long about forming people as the Church. In 1992 John Paul II set out four areas of formation for its priests: human, spiritual, intellectual and pastoral. He called human formation "the basis of all priestly formation". Most of you aren't seminarians. But in 1988 he'd written that lay people need much the same: spiritual and doctrinal formation, and "the cultivation of human values", all so that each of them can discover a vocation and live it. For most of you, the fourth area will take the form of service.

Human formation is character. It's becoming someone other people can rely on: keeping your word, telling the truth when it costs you, staying with someone in their grief when you don't know what to say. John Paul II wanted a priest's personality to be "a bridge and not an obstacle" for others. When machines can produce warmth on demand, people will learn to prize the real thing, and to tell the difference.

Intellectual formation isn't storing information. It's judgement: knowing how you know a thing, noticing when a confident answer is wrong, holding an argument in your head long enough to test it. For the rest of your lives, machines will write the first draft. Whether you can tell a sound draft from a plausible one depends a great deal on what you do here. Pope Leo calls for "rhythms that incorporate silence, in-depth study, reading and judicious analysis, for without these elements inner freedom may be compromised."

Spiritual formation is the one the world will most tempt you to skip, and it's the one that holds the other two together: prayer, the sacraments, the habit of silence before God. When your work no longer tells you who you are, you'll need to know whose you are.

And service, the fourth. For a priest it's pastoral care. For everyone else it's the habit of putting what you've been given at the disposal of other people. It's worth forming early, because the world you're entering will hand you extraordinary leverage. A single graduate directing a handful of AI agents may soon do what once took a whole department. What you do with that power will depend on whether you learned, before you had it, to ask whom it's for.

Let me add one thing that may surprise you, coming from someone in my trade. Don't let all of your formation be abstract. While you can, get as wide an exposure as you can to different fields and kinds of work, including work done with your hands. Cook for a crowd. Build something. Work a harvest, a hospital ward or a workshop. Not because a trade is a safe harbour. I've just told you the robots are coming for those too. Do it because some of you will discover, to your own surprise, that you love it, and that it tells you something about what you're for that no seminar ever could.

That's what a vocation is. It isn't a passion you possess. It's a call that comes from outside you, usually through somebody else's need. And you can't hear it if you've never been anywhere near the need.

The machines will probably take most of the jobs we know. That doesn't leave you with nothing to do, because a job and a vocation aren't the same thing. Becoming a saint takes a whole life. So does building the families, parishes and communities in which saints are formed. That will be one of the great undertakings of your generation, and one of the hardest. Nor does the market get the last word. Firms will automate whatever competition forces them to. But a family, a parish or a town doesn't have to hand its work to a machine just because the machine could do it. Not when one of its own has the gift for it, and wants to give it. How much of our common life we give to machines is for you and your communities to decide. That's why it matters so much that each of you finds your calling. In a world where machines do most of the paid work, the reason to serve can't be money. It has to be love, of God and of the people around you.

V. Imagine It, Then Work Backwards

I want to end with hope. "10 times bigger… and maybe 10 times faster" could describe a catastrophe. It could just as easily describe the best world human beings have ever lived in.

Gates made the point himself. In parts of Africa today, he said, "most people never meet a doctor or almost half the kids are malnourished". Imagine 2040 as it could be. Every child, anywhere, eats enough, and a doctor's knowledge is within reach of every family. The drudgery is done by machines, and the time it frees is given to the people who need it: the old who are lonely, the children who need their parents, the sick who need someone to sit with them. Families own the machines in their homes, instead of renting them from a platform. A girl in Kinshasa reads Augustine in her own language and wrestles with his arguments. The work that remains is chosen, done well, and offered. Sunday is kept, and the parishes are full of people with time for one another.

Life could be better than it has ever been. But none of it happens by default. The same machines could just as easily give us a world of idle, distracted and lonely people, owned by whoever owns the machines. Which world we get depends on choices being made now.

In my industry, when we set out to build something difficult, we often start from the finished thing and work backwards. You picture it in detail. Then you ask what had to be true five years before, and ten, and what has to begin this year. Try that with the world I've just described. Work backwards from 2040, and the problems worth a life come into view. Who will own the machines? Will Gates's twenty years of transition be merciful or cruel? Will the young be formed, or only kept busy? Will the wisdom of the Church be present in the systems people consult?

Those are the right problems, and they aren't where the money and the talent usually go. In 2011 Jeff Hammerbacher, who'd been one of Facebook's early employees, told Bloomberg Businessweek what he made of his peers across Silicon Valley. "The best minds of my generation are thinking about how to make people click ads," he said. "That sucks." My generation built much of what you carry in your pockets, and a great deal of it was designed to hold your attention rather than to serve your good. Yours can choose differently. The right problems deserve the best minds and the best hearts of your generation, and the money, the time and the institutions to match. Longbeard is our attempt to work on one of them. You'll find yours. Some of you will work on them in laboratories and government departments, some in hospitals, workshops and parishes, and some in your own homes, which may turn out to matter most of all.

St John Henry Newman wrote in 1848: "God has created me to do Him some definite service; He has committed some work to me which He has not committed to another." And a few lines later: "I am a link in a chain, a bond of connexion between persons."

Everything I build can be copied a million times overnight. You can't be. The work committed to you has been committed to no one else, and no machine will ever do it in your place.

So picture the world that could be. Work back from it. Find the part that's yours.

And then go and do it.

Thank you.