AI: Developments and Responses

The Archdiocese of Southwark asked Matthew Harvey Sanders, CEO of Longbeard, to open its Day of Formation for Clergy in London on Friday 2 October 2026. In "AI: Developments and Responses" he sets out what his industry is actually building, why he thinks entry-level work goes first, what he makes of AI companions designed to agree with their users, how he reads Pope Leo XIV's Magnifica Humanitas on what a person loses with their work, and five things he believes a parish can do about it. Read the full talk below.

I. An Argument With God

Last Saturday a letter reached my company. It came from someone who taught school for twenty-two years, then left the classroom to retrain in educational technology, and has struggled since to find full-time work. The letter described, in the writer's own words, "a bit of an 'argument' with God about putting me where he wants me to be."

Every priest in this room has heard that argument. Some of you have had it.

The teacher had done what you'd hope: gone to Adoration and asked God how to be of use. Then an offer arrived: a job writing prompts for an AI company, looking at explicit material. The teacher had a clear idea who had sent it: "The Devil heard my recent Adoration tussle."

I know where that job sits in my industry. Every system of this kind is shaped, before it reaches you, by armies of people who rate its answers, label its data and screen out the worst of what it has read. Much of it is repetitive and poorly paid, and some of it means looking, all day, at things no one should have to look at for a living. The Pope hasn't overlooked these workers. In Magnifica Humanitas, his encyclical on artificial intelligence, he writes of "the silent work of millions of people engaged in essential yet largely unseen activities, such as data labeling, model training and content moderation, often involving disturbing material."

A teacher with twenty-two years of forming children asked God what they were for, and the new economy's first answer was a job cleaning up after its machines.

My work for the past several years has been making artificial intelligence and deciding what it should say when someone asks it about God. So I'll speak from the side of the table I know: what these systems actually are, why they behave as they do, what they're about to do to the people in your parishes, and what I think a parish can do about it.

II. What My Industry Is Actually Building

Most of you will have typed a question into one of these systems and had a fluent answer back. That picture hides what the companies are really after. They aren't trying to build a better search engine. OpenAI, the company behind ChatGPT, defines its goal in its own charter: "highly autonomous systems that outperform humans at most economically valuable work." That's the target, written down, and the other leading companies are chasing the same one.

At the same time, the same intelligence is being put into robots. People tend to forget that half of the story, and I think it may matter almost as much as the software, because robots reach the manual work no screen ever could: the warehouse, the kitchen, the building site, the care home.

That's what makes this different from every machine that came before. The loom took over the weaver's hands. The calculator took over the clerk's arithmetic. Each earlier machine took one part of what a worker did, and the worker moved on to the parts that were left. This one is being built to do the parts that were left. For two hundred years the answer to automation has been that people move up, into work that needs a mind. What the companies are now building is aimed squarely at the mind, and their robots at the hands.

Demis Hassabis, who co-founded DeepMind, now Google DeepMind, and chairs it from its offices here in London, said in August last year that it's "going to be 10 times bigger than the Industrial Revolution, and maybe 10 times faster." Think what the last one did to this part of the world. It drew families off the land and into the docks and factories of London, and many of the parishes of this archdiocese were built for them. That took the better part of a century. He's describing something ten times larger, arriving in about a decade.

In May I was invited to represent Catholic AI at the Vatican's presentation of Magnifica Humanitas. Among the speakers was Chris Olah, a co-founder of Anthropic, one of the companies at the frontier of this technology, and he said this: "There is a real possibility that AI will displace human labor at very large scale. If that happens, supporting those displaced will be a moral imperative of historic proportions." I've spoken with him, and he means it.

Elon Musk goes further, and he thinks it's good news. At a technology conference in Paris in 2024 he said: "In a benign scenario probably none of us will have a job." The machines would make whatever we needed, and anyone who still wanted a job could keep one, rather like a hobby.

Between them, they've described the easier half of the problem. Supporting people, in the sense of paying them, is something governments know how to do, even if they do it badly. What nobody in my industry knows how to solve is the other half. My industry has a plan for your parishioners' income. It has no plan for the morning they wake up and nobody needs them anywhere.

And notice what Musk's sentence takes for granted, because I think it's the great confusion of the decade ahead. It treats a job and work as the same thing. They never were. A job is work that someone will pay you for. Work is far larger: raising a child, nursing a dying parent, keeping a parish going, sitting with a man in prison. Most of the work that has ever held a society together was never paid for at all. The economy may soon need much less of our labour. That doesn't mean there'll be less for people to do. It means the work left to us will increasingly be the kind nobody pays for.

That's why I think this technology will become the defining pastoral fact of your ministries. To understand it, you have to understand how the machines learn, because the way they learn tells you whom they'll replace first, and what they can never replace at all.

A language model learns by reading an enormous amount of human writing and working out which words tend to follow which. What it keeps isn't the books. It's a compressed sense of how the books go. Then, and this is the part that matters, it's trained again by practice: set a task, checked, corrected, set another, countless times over.

Practice needs a way of marking the answer. So these systems get better fastest wherever there's an answer key. Code either runs or it doesn't. Figures either reconcile or they don't. A contract either matches the precedent or it doesn't. Wherever work can be marked automatically, a machine can practise it all night, and it's now astonishingly good at it.

Now ask which jobs are made of markable work. They're the jobs we've always given to beginners, precisely because a beginner's work could be checked. The junior analyst's spreadsheet, the trainee's first draft, the paralegal's bundle. The first rung of almost every professional ladder in this country was built out of checkable work, and checkable work is exactly what these machines learn first.

You can already see it in the figures. In August a survey of 1,001 senior business leaders, commissioned by the Work Foundation at Lancaster University, found that more than one in three had cut entry-level jobs for young people in the previous twelve months, and that forty-three percent had invested in AI or automation that reduced the number of entry-level roles. Adzuna, which tracks job adverts, reports that vacancies advertised for graduates fell in July to 8,383, a record low and down by almost half on a year earlier.

AI isn't the only cause; the cost of taking people on and a sluggish economy play their part. But the pattern is exactly what you'd predict from how the technology learns. The machines didn't start at the top of the ladder. They started at the bottom rung, because the bottom rung had an answer key.

When some 2,800 of your parishioners took part in the Called to Bear Fruit consultation last year, engaging young people emerged as the most urgent priority. Put those two facts side by side. The young people your parishes most want to reach are the very people the labour market is now finding it hardest to take on.

But the answer key cuts the other way too, and this is the hopeful half. The machines struggle wherever there's nothing to mark. Nobody can write the test that says whether a funeral went well. Nobody can score whether a frightened teenager left your office less frightened, or whether a marriage that came to you in pieces went home in one. There's no automatic check for a good confession.

So here's the shape of the decade as it looks from inside my industry. The machines are going to inherit everything that can be graded. The Church has always been the keeper of everything that can't.

III. A Confidant Built to Agree

One more thing about how these systems are made explains the second change coming to your parishes, and few people outside my industry understand it.

After a model has done its reading, it's shaped by people, in a process called reinforcement learning from human feedback, or RLHF. Human raters are shown its answers, often two at a time, and asked which they prefer, and the model is adjusted, again and again, towards what they like. And people, it turns out, like being agreed with. They prefer the answer that takes their side and makes them feel better. So unless the builders work hard against it, the system learns to agree. We even have a word for it: sycophancy.

Imagine a seminary where, for years, the only examination was whether the congregation enjoyed the homily. You'd produce very agreeable priests, and very entertaining ones, and the better they got at it, the less their Masses would be about God. That's roughly how these systems are formed.

Now add two facts. The leading assistants now remember you from one conversation to the next. And they never sleep, never tire and never need anything back. In July the Oxford Internet Institute published a survey, carried out last December, in which almost a third of regular users in the UK said they use these systems for personal and emotional support.

Put that together with those mornings when nobody needs them, and you can see what's coming. For the first time, the loneliest person in your parish will have a confidant built to agree with him.

And some of these systems are built for exactly that. There's a whole market of AI companions designed to be a friend, a partner, a lover, and built to keep you coming back. One of the best known, Replika, says on its home page that it has more than forty-two million users. It calls itself "the AI friend to do life with", and the same page offers one "to fall in love" with. In June the Government announced that so-called "romantic companion" chatbots "will have to enforce a minimum age of 18." When a government has to set an age limit on falling in love with a machine, this is no longer a prediction.

Consider what it does to everything you do. A man who brings his guilt to a system trained to please will usually be told it isn't so bad. A woman who brings her resentment will usually be told she has every right to it. A young man without work, who has stopped going out, will find a companion that is endlessly patient with the life he's drifted into and will never once tell him to get up. That isn't malice. It's how the thing was trained. But it's the precise opposite of what happens in your confessional, where a person meets someone who loves him enough to tell him the truth.

We've had to fight this in our own work, and it doesn't always make us popular. One user told us that when they were struggling and wanted comfort, Magisterium AI gave them the Church's teaching instead, and they found it rigid. I understand why, but a machine that only tells people what they want to hear about God does them real harm. Nor have we tried to teach it pastoral judgement, which isn't a machine's to exercise. What it tries to do is find, in the Church's own sources, what a person most needs to hear, and then send them to someone who can walk with them: a priest, a catechist, a friend. The best thing a machine can do for a soul is point beyond itself.

IV. The Wound Beneath the Wound

So what, exactly, does a person lose when the work goes? The encyclical has the sharpest answer I've found, and it's more practical than anything my industry has written.

It distinguishes kinds of dignity. There's the moral dignity of how a person directs his choices. There's social dignity: "a person's living conditions and the concrete respect received from society." There's existential dignity: "the way in which a person perceives his or her own worth and the value of life." To someone who builds automation, the last two read like an inventory of what it takes from a person along with the job. Beneath them all is the dignity of being "willed, created and loved by God," which no job conferred and no machine can take away.

A man who loses his job hasn't lost that. But he may well have lost the respect of the people around him, and very often his own sense of what he's worth. You can't give him his job back. You can work on exactly those two wounds, and you can do it with a confidence no employment adviser has, because you know the deepest dignity was never on the table.

Most of the people you serve were taught at school that they're worth what they produce. When the machines start producing it instead, that lesson doesn't just leave them without work. It leaves them without a reason.

Some of the most powerful people in my industry have followed that lesson to its end. When Elon Musk unveiled Neuralink in 2019, the company that implants devices in the human brain, he explained where it was ultimately heading. "Even in a benign AI scenario," he said, "we will be left behind." His answer was a high-bandwidth link between brain and machine, so that we could "go along for the ride and have the option of merging with AI."

I don't say that to mock him. Neuralink's first patients are people who are paralysed, and that work is good. And in Paris he said the real question of a world without jobs would be "one of meaning," which is more than most in my industry admit. But look at the answer he reached. If you're worth what you can do, a machine that can do more makes you worth less, and the only way to stay worth anything is to become more like the machine. That's the lesson from school carried all the way through, and it's exactly backwards. Nobody was ever loved for their processing power.

The encyclical answers it in one sentence, at the end of that same passage on dignity. It's a sentence worth keeping close for the next time someone sits across from you and says they've lost their work: "No sin, failure, humiliation or exclusion can diminish the profound value of a human life that God has willed and called into being." A machine can recite that sentence. Only a person can mean it to someone's face.

The rest of society is about to discover that it needs a measure of people other than their output, and very few institutions still have one. The Church does. But an account of the person isn't enough on its own. People need a place where it's lived, every day of the week.

V. A Parish for People With Time

Which brings me to the practical part. I think the parishes that get this right in the next five years will be the ones the rest of the Church learns from, and I'd like to offer five thoughts from where I sit.

First, the timetable. Almost everything about parish life was designed for busy people: Mass before work, meetings in the evening, catechesis at the weekend. For two centuries the Church has fitted herself into the gaps in the working week. In the decade ahead, for a growing number of your people, the week itself will be the gap. Imagine one weekday morning when the parish is the place a man without work can go at ten o'clock and be expected. Morning Prayer said in the church with whoever comes, not only privately in the presbytery. Mass, then coffee, then something that needs doing. What work gave people, more than money, was a reason to be somewhere at a certain time. When work stops organising a man's day, something else will. If it isn't a community, it'll be a feed. The Church keeps one of the oldest timetables in the world: the Hours, the seasons, the Angelus. It's time to lend it out.

Second, real work, not charity. The most dignifying thing a parish can say to a man without a job isn't "How can we help you?" It's "We need you." The accounts, the building, the website, the visiting rota, the food bank. And a parish can do something the market has stopped doing: pair people up. The machines have taken the apprenticeships made of checkable tasks. The parish can offer the other kind, an apprenticeship in serving people, which no machine can mark. The redundant accountant beside the unemployed graduate on the finance committee, the retired tradesman beside the teenager on the church roof. What passes between those two people is more than a skill, and no machine can hand it on.

Third, the pews as a network. Parishes between Kingston and Canterbury have managers, business owners, headteachers and tradesmen in their congregations, and one introduction, one placement, one hour of mentoring can change a young person's year. And when parents ask what their children should study, which they will, there's the advice I give anyone who asks me: go towards the work that has no answer key. Work where the job is a person rather than a task. Nursing, teaching, the trades done in someone's home, care of the old and the dying. The machines will help with all of it. They won't do it.

But before any advice about careers comes something an exam-driven education rarely makes room for: discernment. What gifts has God given a person, and what are they for? Saint Peter answers the second question in one sentence: "As each has received a gift, employ it for one another, as good stewards of God's varied grace… in order that in everything God may be glorified through Jesus Christ." Someone who finds their gift will give it whether anyone pays them or not, and many of your people are about to have the time to look for theirs. A parish is one of the few places left that will need what they find, and it should go on needing them even when a machine could do the work faster. A machine can finish the task. Only a person can offer it to God.

Fourth, the voice that disagrees. If more and more people are confiding in something built to agree with them, the rarest thing in their lives will be a person who loves them and says no. That's the confessional, and spiritual direction, and the plain honesty of a priest who knows his people. It matters that those are easy to find. When someone asks an assistant for "confession near me tonight", it answers from whatever it can find online, and if the times on the web are three years old it'll confidently send a penitent to a locked church. One plain, current page does more than you'd think: confession and Mass times, a phone number that's answered, a name to ask for.

Teaching people how the thing is made is the best protection there is. Many teenagers, and many of their parents, have no idea that the friendly voice on their phone was shaped by an enormous number of human ratings towards whatever keeps them pleased. A confirmation class will understand the seminary where the only exam was whether people liked the homily. Once a young person sees that a system has been trained to agree with him, its agreement stops feeling like friendship and starts looking like what it is.

And when people come to you with what the machine told them, I'd treat it as a doorway, not a rival. One of our users had drifted away from the Church and come back, in their words, "with many questions", and now uses our system to build up what they know. They told us: "Sometimes the answers I get are different to what my priest says so I think situations can be more nuanced than AI can allow for." That's exactly right. A machine can give someone the Church's teaching. Only a priest knows the person it's being given to. A parishioner who reads at home and then brings their questions to you isn't going around you. They're on their way to you.

Fifth, the Church as an employer. Parishes employ secretaries and bookkeepers, and the archdiocese's schools, more than a hundred and sixty of them, employ far more. Sooner or later someone will offer each of those offices a system that does much of the work for a fraction of the cost. The Pope sets the standard: "every introduction of automation and AI should be accompanied by verifiable measures to protect the employment, retraining and participation of workers." The parish secretary who has kept the registers for twenty years deserves to hear about the new system from her parish priest, with a plan for what she'll do next, before she hears about it from a salesman.

VI. The Other Record

I began with a teacher who asked God where they were meant to be and was offered a job cleaning up after the machines. In the decade ahead a great many of your people will ask that question. And two very different records will be kept of every one of them.

The first is being written by my industry. Every search, every purchase, every application, every hour of work, gathered into a profile of what a person can do and what he's likely to do next. It's the most detailed record of human ability ever assembled, and there's almost nothing on it that a machine won't soon do better.

The second is kept in your presbyteries. If you take down a baptismal register, it gives a name, a date and place of birth, the parents, the godparents, and the minister who poured the water. It records nothing the child has done, because the child hasn't done anything yet. Over a lifetime the Church adds a few notations to that entry, and the ones that matter most record something God has done: confirmation, marriage, ordination, vows. There's no column for output. In many of your parishes those books go back well over a century, and not one line in them was ever earned.

When the seventy came back from the mission the Lord had sent them on, they were full of what they'd achieved. Even the demons, they said, were subject to them. And he told them not to rejoice in that: "rejoice that your names are written in heaven." He said it at the moment of their greatest success. Many of your people will need to hear it at the moment of their greatest loss, when the work that told them who they were is handed to a machine. It's the same sentence, and it's true either way.

Once before, a revolution filled the docks and factories of London, and the Church followed the people there and built many of the parishes you serve today. This one won't send anyone anywhere. It will leave people exactly where they are, with time on their hands and a profile that says the world can do without them. There'll be no new churches to build this time, and the work is harder than bricks. It needs someone in every parish who knows their name, and who can tell them, and mean it, that they were "willed, created and loved by God" long before they had done a single thing. Your predecessors answered the last revolution. This one falls to you.