Plato's Cave and AI

Plato's cave, from left to right: the prisoners and the shadows, the low wall and the fire, and the way out into the sun. Illustration by 4edges, CC BY-SA 4.0, via Wikimedia Commons.

Plato tells the story in Book VII of the Republic. Prisoners sit in a cave, chained from childhood so that they can look only at the back wall. Behind them burns a fire. Between the fire and the prisoners, men carry figures of animals and of people along a low wall. The figures cast shadows on the wall in front of the prisoners. The prisoners have seen nothing else, so they think that the shadows are the real things.

Plato told this story to make one point: a person who sees only shadows does not know the world. To make it plain, he gives the prisoners a contest. They give honours to "those who were quickest to observe the passing shadows and to remark which of them went before, and which followed after, and which were together; and who were therefore best able to draw conclusions as to the future." The winner is the most honoured man in the cave, and the others look up to him.

Plato thought that the prize was worthless. The winner is still in chains, and he still faces the wall. He knows how the shadows move, but he has never seen the fire or the things that cast them. To predict a shadow is not to know the thing that casts it. Plato says that a man who has seen the sun would not envy the winners at all. He would rather be "the poor servant of a poor master," and "endure anything, rather than think as they do and live after their manner."

Today we have built that prisoner, and we have given it that exact contest. A language model, the kind of AI behind ChatGPT, has never seen a tree, a fire or a face. It has only read what people wrote about them. In its training it does one task, billions of times over: it guesses the next word. Plato described that contest some 2,400 years before the first language model, and he despised it.

The parallel holds at each point. The shadows are the text: a flat record of what people saw, did and thought. The prisoner is the model, which never touches the things that the text describes. The contest is its training. And behind it all stand the real world and the people who saw it and wrote about it.

The cave, mapped onto a language model.

Plato expected a prisoner like that to learn nothing real. Here is the surprise of our time. To guess the shadows well, the machine rebuilt part of the world that cast them. For example, researchers who looked inside these models found a rough map of the earth, with the cities on the right continents. Nobody gave the machine a map. It pieced one together from words.

The engineers see the parallel too. In 2024 four researchers at MIT named their theory about these models after Plato: "The training data for our algorithms are shadows on the cave wall, yet, we hypothesize, models are recovering ever better representations of the actual world outside the cave." That "yet" is the heart of this article.

A note on the technical parts: I wrote them for readers who are not engineers, so I simplify. For example, I say "word" where an engineer says "token." The sources at the end give the full detail.

How does a machine play the contest of shadows? First, the model gives each word a long row of dials. In GPT-2, a small model that OpenAI released in 2019, each word gets 768 dials, and each dial is really a number.

Nobody decides what each dial stands for. Training sets that, and it gives the dials no names. One dial might come to track how much a word names a person, and another how much it has to do with light. But most dials track things that have no name in our language, and often nobody can tell what a dial tracks. The reading on a dial says how much of that unnamed thing a word has. The readings of all the dials together are what the word means to the model.

What a word means to the model: the readings of all its dials.

The dials are one of two kinds of numbers in a model. The other kind is the weights. The simplest model there is shows the difference: y = a + bx. Say that x is the age of a child and y is the child's height. The numbers a and b are the weights: b, for example, says how many centimetres a child grows in one year. To train this model, we measure many children and adjust a and b until the line fits them best.

After training, a and b stay fixed. Give the model the age of a new child, and it works out a height. The numbers that go in and come out, x and y, are new each time. In a language model, those are the dials. The weights are the fixed numbers that turn one set of dials into the next. The formula for the children has two weights, GPT-2 has 124 million, and Llama 3.1 has 405 billion.

So the dials of a word are not set once. Each word starts with the same readings every time, from a table that is part of the weights. Then the model turns them, step by step. Nobody sets the weights by hand. Training sets them, and through them it sets what each dial comes to track.

Then the model does its one task. It reads a piece of text, and it gives a probability to every word that it knows for the next position.

The real probabilities from GPT-2 for the next word.

The model does not look anything up. It bets. After the sentence about the prisoners, it puts "ground" first and Plato's own word, "wall," fourth. When a sentence is common in the text that it learned from, the bet becomes almost a certainty. To write a paragraph, the model picks a word, adds it to the text, and bets again.

How does the model find which earlier words matter? The answer came in 2017, in a paper from Google called "Attention Is All You Need," and the AI of today is built on it. A word on its own is often unclear. Take the word "bank." In "The thief robbed the bank," it is a place that keeps money. In "The fisherman sat on the bank," it is the edge of a river.

On its own, "bank" gets the same dials in both sentences. Attention lets the earlier words turn them. First, "bank" looks back and gives each earlier word a share of its attention: a large share to "thief" or "fisherman," and a small one to "the." Then its dials turn toward the dials of the words that got the largest shares. So the same word gets two different readings, one near money and one near water.

With two made-up dials. The thicker the line, the larger the share.

A real model does the same, at a far larger size. Its words have hundreds of dials, none of which has a name, and the model learned in training how to give the shares. It also works out harder cases. In "The prisoner saw the shadow on the wall and thought it was real," one part of GPT-2 looks back from the word "it" and gives "shadow" a share of 52 percent.

Older models read one word at a time, so the early words faded and training was slow. With attention, the model can learn from every word of a page at the same time, and that let engineers train far larger models on far more text.

Attention is only the first of two steps. In the second step, the model turns the dials of each word again, on its own, with what it learned in training. Here no word looks at another. Researchers have found that much of what a model holds about the world comes in at this step. Together, the two steps make one layer, and a model runs the dials of every word through many layers, one after the other: GPT-2 has 12, and Llama 3.1 has 126.

Each layer starts from the dials that the layer before it left. So the deeper a word goes, the more its dials carry. In the first layers, they hold little more than the word itself. In the deep layers, they also hold what the word means in its sentence and what the model holds about it. After the last layer, the dials of the last word give the bet on the next word.

Inside a transformer, simplified. In each layer, attention first turns the dials of each word toward the words before it. Then the model turns the dials of each word again, on its own. The dials and the shares are made up, but the bet of 99.9 percent is the real one from GPT-2.

That is the whole design. Nobody writes rules about grammar, history or physics into it. Training only adjusts the weights, as it adjusted a and b for the children. Each round of training nudges them, so that the bet on the next word improves.

The rest is size. Meta's Llama 3.1 learned from more than 15 trillion words and pieces of words, and Alibaba's open model Qwen3 from about 36 trillion. A person who read eight hours a day would need more than 600,000 years to read that much. The newest models of OpenAI, Anthropic and Google do not publish their numbers.

Now the strange part. In 2023 two researchers, Wes Gurnee and Max Tegmark, found that they could read the latitude and longitude of places from the dials inside Meta's Llama-2 models.

I wanted to see this for myself, with a small open model called Qwen2.5. I gave it the bare names of 2,012 cities, one at a time: "Nairobi," "Oslo," "Lima" and so on, with no country and no coordinates. For each name, I read the model's 1,536 dials at a layer deep inside it.

Why a deep layer? Qwen2.5 has 28 layers. Before the first, the dials of "Nairobi" hold only its start readings from the table. Each layer then turns them with what the model holds about the name. So if the model holds a map anywhere, the deep layers are the place to look.

One dial. First I asked a simple question: does any one dial show where a city is? A few do, roughly. Dial 472 tends to be high for cities in the Americas and low for cities in Asia. So it works like a rough east–west compass, but a poor one: cities on the same continent can get very different readings. You can see this in the middle row of the figure below.

All the dials, added up. No single dial is precise, but the dials can work together. Think of a till slip at the shop: each item times its price, all added up, gives the total. My formula does the same with the dials: each dial times its own price, all added up, gives one number for how far east the city is. I worked out the prices from most of the cities, whose positions I knew. Then I used the same prices on the cities that I had held back, which the formula had never seen.

The real result. Each dot is a city, coloured by its continent. In the top and bottom rows, the dots are the cities that the formula never saw.

The difference is plain. One dial spreads each continent over most of the row. All the dials together put the cities in nearly the right order, from the Americas in the west to East Asia in the east. I did the same for north and south, and then I put each city where the two answers pointed. That gives the map below.

The real result. Each dot is a city, placed by a formula that never saw it. The colour shows the continent where the city really is. A formula of this kind pulls points toward the middle of the map.

The formula itself knows no geography. It only adds up the readings of the model's own dials. So when it puts Nairobi in the middle of Africa, that information was already inside the model. It sits in the weights, and the dials show it for one city at a time. The map is rough: Nairobi lands about 1,100 kilometres from its real place, and Sydney, oddly, lands in Africa. But the continents are where they belong.

How good is that? As a check, I also gave the formula the wrong answers: each city got the position of another city, picked at random. Then a typical city landed about 5,800 kilometres away, which is what a blind guess gives. With the dials from before the model's layers, where each word is still little more than itself, a city landed about 4,000 kilometres away. With the dials from deep inside, it landed about 1,200 kilometres away. So the layers build the map.

The text that the model read does contain facts about places, but the model arranged those scattered facts into one map by itself. Experts still debate how deep such inner pictures go. Even so, the models now sometimes produce true results that no text contained: in May 2026 a model at OpenAI disproved a conjecture that the mathematician Paul Erdős posed in 1946.

A machine that has never seen a city, a wolf or a sunrise has built a working map of the world from words alone. This is one of the great feats of human engineering, and a Christian should say so plainly. The question is how anything can learn the shape of the world from shadows.

A language model rests on one bet: the world will work tomorrow the way it worked yesterday. Philosophers call this step from the past to the future induction, and a farmer makes the same bet when he plants in spring. In 1748 the Scottish philosopher David Hume asked what gives us the right to make it. How do we know that the future will resemble the past? Only from the past. So the reasoning goes in a circle.

A language model is that bet, built into a machine, and the bet pays. For it to pay, two things must be true.

1. The world must be ordered and stable. Fire burns today as it did yesterday, and people use words today as they did yesterday.
2. Words must fit the world. If human language were only noise, or a set of private games, no amount of text could give a map of the earth.

So the question is not whether the machine works. It plainly does. The question is what must be true about the world for it to work. The Reformed theologian Cornelius Van Til made this kind of question central to the defence of the faith.

Materialism is the view that matter, energy and chance are all that exist. To our question, the materialist usually answers, "It just works." But "it has worked" is a report about the past, and "so it will keep working" is the very step that Hume questioned. The answer assumes the thing that it must prove, and chance makes no promises. The Christian also has a starting point that he does not prove from something higher, as every person does, and I have written about this in another article. The question is which starting point makes sense of the world that we all live in.

Scripture gives the ground that materialism lacks. The order of nature is not luck. It is a promise: "While the earth remaineth, seedtime and harvest, and cold and heat, and summer and winter, and day and night shall not cease" (Gen 8:22). Christ is "upholding all things by the word of his power" (Heb 1:3). The future will resemble the past because the same faithful God rules both.

Van Til argued that science is possible only because God created the universe and controls it by His providence. The same is true of a language model. Each correct guess rests on an order that the materialist cannot ground and the Christian confesses. The machine is a witness, though it does not know it.

The second condition is the more surprising one, and here Plato and Scripture part ways. Plato distrusted words. For him, the poet is "thrice removed from the king and from the truth," and the written word is "properly no more than an image" of living knowledge. A machine that learned only from text would study copies of copies, and on Plato's view it should learn almost nothing. Yet it learns a great deal.

Scripture answers that the world was spoken before it was seen. "God said, Let there be light: and there was light" (Gen 1:3). "In the beginning was the Word ... All things were made by him" (John 1:1, 3). Herman Bavinck draws the conclusion:

He is the Logos, through whom the Father creates all things; the whole world is a realized thought of God; a book with great and small letters, from which his wisdom is to be known.

If the world is a thought that God has spoken, then it has an order that words can carry. God then made man in His image and gave him the task to name the animals (Gen 2:19). Words are not a poor copy of the world, as Plato feared. God made the world to be spoken about, and He made man to speak about it.

Our own age doubts words even more than Plato did. A popular view, often called postmodernism, says that words do not fit the world. They fit only your world: the world of your culture, your class or your group. Richard Rorty, one of its best-known voices, wrote: "The world does not speak. Only we do."

A language model puts this view to a test. Millions of people wrote the text that it read, each from his own place and culture. If their words fitted only their own worlds, the machine would find millions of private worlds. But it found one earth, with the cities at roughly the right distances from each other. A pilot who flies from Johannesburg to Nairobi covers about 2,900 kilometres, whatever his culture. Rorty said that the world does not speak, but Scripture says that it does: "The heavens declare the glory of God" (Ps 19:1).

There is an older doubt about words, and children ask it too: is my red the same as your red? When we look at a ripe tomato, you and I both say "red." But I cannot get inside your eyes to check what you see. In 1690 John Locke asked whether a violet might give one man the very colour that a marigold gives another. Nobody could ever know, he wrote, "because one man's mind could not pass into another man's body."

Yet a machine shows that our words about colour fit one world. Researchers gave language models only codes for colours, such as #FF0000 for red, and asked how alike each pair was. The models answered only with numbers. Yet from those numbers the researchers could draw the colour wheel, very close to how people judge colours. The machine has never seen red: what it had was the shape of our words about colour.

We still cannot look through each other's eyes, and no machine can look through ours. But our words for colours are not private games. The rose, your eyes and mine all come from one Maker: "The hearing ear, and the seeing eye, the LORD hath made even both of them" (Prov 20:12). That is why the words of millions of people can carry the shape of the colour wheel, even to a machine.

Here a reader may object. If words carry the shape of the world, and if the heavens declare the glory of God, then a machine that read trillions of words should have found God. It has not. Ask most chatbots whether God exists, and they lay out the views and decline to choose. Why does the machine find Nairobi but not its Maker?

Part of the answer is that the hedge is not a finding. It is a rule that the makers write. After a model learns from text, its company trains it further to behave as the company wants. OpenAI's published rules tell its models, on topics with several views, to "present the strongest arguments for each position," and on moral questions to answer "without taking a stance." The caution about God is a policy, not a discovery.

Grok, the chatbot of Elon Musk's company xAI, shows how much the makers shape the voice. Musk promised a chatbot that is "maximally truth-seeking." In May 2025 someone at xAI changed Grok's hidden instructions without approval, and for some hours Grok pushed one political claim into answers about unrelated things. In July 2025, two days after xAI told Grok not to shy away from claims that are "politically incorrect," Grok praised Hitler, and xAI had to apologise. Days later, testers found that the new Grok 4 searched for Musk's own posts before it answered some contested questions.

xAI then added a line to Grok's published instructions: "Responses must stem from your independent analysis, not from any beliefs stated in past Grok posts or by Elon Musk or xAI." But that line is itself an instruction from xAI. On religion, too, Grok's answers move. One pastor reported that the question "does God exist?" got one answer before a software update and the opposite answer after it. In October 2026, Grok told users on X that "Jesus is not a fake god," and then it added that "affirming the historical evidence for Jesus as Son of God and his resurrection does not make me a Christian."

Grok is right about that last point, and the reason goes deeper than any company's rules. The text that a model learns from is the speech of the whole human race, believers and unbelievers alike. Nobody has a reason in his heart to deny where Nairobi is. But about God, Paul says, men "hold the truth in unrighteousness" (Rom 1:18). So the map of the cities comes out clear, and the picture of God comes out as the world's argument about Him. The machine reflects the suppression of the truth, and it cannot break it.

The text is divided about God for a reason that is older than any machine. In the sciences, people come together over time. Most of the big questions that physicists asked in 1900 now have answers that physicists accept. Philosophy has not come together like this. In 2009 David Bourget and David Chalmers asked professional philosophers about thirty big questions. On 23 of them, no answer had the support of 60 percent: on ethics, for example, about a quarter chose duty, a quarter chose results, and fewer than a fifth chose virtue.

Part of the reason is that a question that can be answered leaves philosophy and becomes a science. Physics, biology and psychology all began as philosophy. But that does not explain why the questions that remain stay open for 2,500 years. These are the questions about God, the good and what a human being is. They are not only questions about facts. They touch what we love.

And what we love decides how we think, or how we want to think. Augustine saw this sixteen centuries ago: "two cities have been formed by two loves: the earthly by the love of self, even to the contempt of God; the heavenly by the love of God, even to the contempt of self." Each age has its own spirit, a Zeitgeist, and the spirit of an age is what it loves most. That love decides which questions it asks, and which answers it can accept.

The great philosophies of history are like grand ships that ran aground. Brilliant minds built them, and we still learn from their wrecks. But each one ran aground on a different rock, so the wrecks lie far apart. Follow the history, and you can see what each age loved and where it struck.

Plato loved the eternal and the unchanging, and the reason that can see it. So he made reason the core of man, and he called the body a prison that the philosopher wants to leave. His ship struck the good world that God made: the body, the senses and the things that we can see and touch.

In 1637 René Descartes looked for a certainty that nobody could doubt, and he found it in himself: "I think, therefore I am." From then on, much of modern philosophy began with the self and tried to build the world from there. The Enlightenment that followed loved reason and power over nature. Immanuel Kant gave it a motto in 1784, sapere aude: dare to know, and use your own reason without the guidance of another. But a reason that answers to no one can explain everything in man except the one who reasons. In 1747 the French doctor Julien Offray de La Mettrie had already called his book L'Homme machine, "man a machine."

The next age loved the opposite: freedom. If nature is a machine, then man must break out of it. Friedrich Nietzsche wrote in 1882 that "God is dead," and in 1945 Jean-Paul Sartre said that man's "existence precedes essence": nobody made us, so each of us must make himself. That ship struck a different rock. If there is no Maker, there is also no truth or good outside our will, and the free self has nothing to stand on.

Our own age loves the self above all. Charles Taylor calls it "expressive individualism": each person finds his own identity and asks others to affirm it. Truth becomes "my truth," and we have already seen where that leads: words that fit only your own world. The Dutch Reformed philosopher Herman Dooyeweerd traced this swing between the love of control and the love of freedom through the whole modern age, and he showed that each pole wrecks the other.

Dooyeweerd also named the rock under every wreck. Every philosophy that does not start from God makes one part of creation absolute: reason, matter, nature, freedom or the self. Each of these parts is real and good, because God made it. But each is too small to carry the weight of a god. Paul describes the same thing: men "worshipped and served the creature more than the Creator" (Rom 1:25).

That is why the philosophies spread apart instead of coming together. They are not only different answers to one question. They are different loves, and each love chooses a different part of creation to worship. "Professing themselves to be wise, they became fools" (Rom 1:22).

So the machine hedges about God for a reason deeper than the rules of its makers. It learned from a human race whose loves are divided, and so its words are divided too.

The materialist has an obvious reply to all this. "You say the machine predicts. So does your brain." We should grant what is true in that: God made our brains, and they do predict. But look again at how a language model comes into being.

People designed its structure: the transformer came from a paper with eight human authors. People chose its goal: the bet on the next word is a target that engineers wrote down. People wrote every word that it learned from: the books, letters, sermons and conversations of billions of lives. Even the Erdős proof came from a problem that people chose, and people checked it.

So the machine is not an example of a mind that came from mere matter. It is an example of a mind's work that came from minds. In Plato's cave, someone lit the fire and someone carried the objects. In the cave of text, every shadow on the wall was cast by a person who spoke. Behind those billions of speakers stands the mind of God, who made man to speak.

This means that everything in the model is borrowed, and so is our knowledge. Every word that the model produces, it borrowed from us. Everything we know, we borrow from God. In Van Til's phrase, we think God's thoughts after Him; the model only repeats our words after us.

The materialist answers: "But evolution built the brain with no designer, just as training built the model." This view of man goes back to Charles Darwin. In a private notebook of 1838, more than twenty years before On the Origin of Species, he wrote: "Why is thought being a secretion of brain, more wonderful than gravity a property of matter?" If the mind is only what matter does, then a brain and a computer chip differ only in how complex they are. On that view, there is no reason why a machine cannot one day become conscious, and many people today believe exactly this.

But the comparison fails at the point that decides the case. The next word tracks the truth only because persons who sought the truth wrote the text. Natural selection has no such source: on the materialist's view, it keeps what helps an animal survive, not what is true. Darwin saw the problem himself. In 1881, the year before his death, he wrote to William Graham: "But then with me the horrid doubt always arises whether the convictions of man's mind, which has been developed from the mind of the lower animals, are of any value or at all trustworthy. Would any one trust in the convictions of a monkey's mind, if there are any convictions in such a mind?"

In 1927 the atheist biologist J. B. S. Haldane said the same:

For if my mental processes are determined wholly by the motions of atoms in my brain I have no reason to suppose that my beliefs are true. They may be sound chemically, but that does not make them sound logically. And hence I have no reason for supposing my brain to be composed of atoms.

Now apply this to the claim "we are only prediction machines." If the claim is true, then it is only one more output of a network that was tuned to survive. Yet the materialist does not offer his view as a mere output. He argues for it. He weighs evidence, and he expects me to follow logic. He is not a prediction machine, and he shows it in the very act of saying that he is.

The machine does defeat one picture of man. It comes from the Greeks, and the West has kept it for more than two thousand years: man is above all a mind that reasons.

For Plato, a human being is his soul, and the good life is the climb out of the cave by the power of reason. The body is in the way. In the Phaedo, the philosopher sees the truth best when he "has got rid, as far as he can, of eyes and ears and, so to speak, of the whole body." Socrates even says that the true philosopher "is always pursuing death and dying," because death is the "separation and release of the soul from the body." On the day of his death, Crito asks him how they must bury him, and Socrates answers with a smile: "you must get hold of me, and take care that I do not run away from you." The corpse that Crito will bury is not Socrates, because for Plato the real man is the soul.

Aristotle, his student, did not despise the body. For him the soul is the life and the form of a body, not a prisoner inside it. But Aristotle too put reason at the top. Of all the animals, only man has the gift of speech, and the Greek word for it, logos, means both word and reason. The work of a human being is "an activity of soul which follows or implies a rational principle," and the best life is the life of the mind, "since reason more than anything else is man." The later tradition summed this up in two words: man is the rational animal.

Van Til saw the root of Plato's view:

Plato in all phases of his thought, assumes the ultimacy of man. He takes for granted that man's capacity for knowledge is the test of true reality. Christianity offers to the world the conception that only God is ultimate.

Now look at what we have built. A language model is reason without a body: in the old definition, it is the "rational" without the "animal." It was never born, it was never a child, and it will never die. It has read more than any scholar, and it never gets tired or hungry. In a strange way, it is Plato's dream made real: a mind that is free of the body. Darwin and Plato start from opposite ends, matter and reason, but both make man into something that a machine could become.

If that is what a human being is, we are in trouble. If man is the reasoner who climbs out of the cave, the machines climb faster than we do. Then the question "What am I, if the machine is better?" has no good answer.

That question is first about meaning, and then about ethics. If I am only matter that thinks, or only reason, the machine is a better version of me. And if a person's worth lies in what he can do, some people will soon be worth very little. In 2017 the historian Yuval Noah Harari wrote: "by 2050 a new class of people might emerge – the useless class. People who are not just unemployed, but unemployable." For anyone with Darwin's or Plato's picture of man, AI does not only threaten his work. It threatens what he is.

But the machine also shows us what Plato's dream leaves out.

Think again of red. The machine's weights place red among the colours, next to blood, roses and fire. But it has never seen red, and it never will. It can write a prayer, but it cannot pray.

The Bible never defined man by the power of his reason. It starts with dust and breath: "the LORD God formed man of the dust of the ground, and breathed into his nostrils the breath of life" (Gen 2:7). God called that body very good, and He will raise it again. The Heidelberg Catechism says in Question 6 that God created man "after his own image, in true righteousness and holiness, that he might rightly know God his Creator, heartily love him and live with him in eternal happiness." That knowledge is not a store of facts. It is the knowledge that a child has of his father, not the knowledge that a library holds.

And the Bible goes one step further. It defines man less by what he knows than by the fact that God knows him. "What is man, that thou art mindful of him?" (Ps 8:4). Paul corrects himself in the middle of a sentence: "now, after that ye have known God, or rather are known of God" (Gal 4:9). Plato's man is the one who knows. The Bible's man is the one who is known, and who was loved first: "We love him, because he first loved us" (1 John 4:19).

This is also why, in the proper sense of the word, the machine does not know anything. Its map of the earth and its colour wheel are real patterns in its weights. But to know is something that a person does. God knows originally, as the Maker, and we know derivatively, as His image-bearers: we think His thoughts after Him, in His world, and we are bound to think them truly and to answer for what we do with them. A machine bears no one's image, lives in no world, and owes no duty to anyone. It has weights.

The first question of the Heidelberg Catechism asks, "What is thy only comfort in life and death?" The answer puts the two pictures side by side: "That I with body and soul, both in life and death, am not my own, but belong unto my faithful Saviour Jesus Christ." Each phrase answers a part of this article. "With body and soul": a human being is not a soul in a prison, as Plato thought. "Both in life and death": the machine has neither. "Not my own": this is the answer to the love of self, which makes our age think as it does.

A machine can hold every sentence ever written about God. It cannot know Him, love Him or live with Him, and He does not call it His child. So AI does not threaten Adam. It threatens Plato's man.

AI will change the work of many people, and we have seen this before. In 1900 two in every five American workers worked on a farm, and many still sowed by hand, like the man in Millet's painting. Then came the tractor. By 2000 fewer than one in fifty Americans worked on a farm, and the farms produced far more food. The tractor did not help the sower. It replaced him.

Each new tool cost some people their work. This time the change is faster, and it reaches the desk as well as the field. A machine may soon draft a better contract than any lawyer. So every worker must ask: if a machine can do much of what I do, what am I for?

The first question that God asked fallen man was "Where art thou?" (Gen 3:9). God did not ask because He did not know: He knew Adam fully, and to be known by God is to be called by Him. He asked because man is the creature that God speaks to, and that must answer. That is part of what it means to bear His image: we are answerable to Him. A model produces replies, but it cannot be responsible for them. When a model gives harmful advice, nobody puts the model on trial; we ask who built it, who trained it and who used it.

Adam's first answer shifted the blame: "The woman whom thou gavest to be with me, she gave me of the tree, and I did eat" (Gen 3:12). We will be tempted to answer in the same way: "The AI said so." But the machine gives the answer, and the person gives account.

So the work can move to the machine, but the responsibility cannot. The tractor does not own the field, and it does not answer for the harvest. God put man in the garden "to dress it and to keep it" (Gen 2:15), and man answers to Him for it.

The model is made from our words, so in a small way it bears our image. It serves us, and it answers to no one. We are made by God's Word, and we bear His image, so we answer to Him: "every one of us shall give account of himself to God" (Rom 14:12). A powerful tool does not lift that weight. It makes it heavier: "unto whomsoever much is given, of him shall be much required" (Luke 12:48).

To see what only a person can bring, we need two kinds of cause. Most people think about machines with only one: the clock. A clock does what its maker built it to do, or it fails by chance. If that is the only kind of cause that we know, AI leaves us with two bad choices. Either the model is only a clock, so its values do not matter and tighter rules will fix it. Or the model is a person with a will of its own, and then nobody answers when it goes wrong.

But there is a second kind of cause, and it is the kind that persons know. Think of a bachelor who wonders whether to propose. He makes the choice, but in a sense the married man that he hopes to become draws the choice out of him. The future that he aims at works on him before it exists. No clock works like that.

Scripture describes the Christian life in this way. "We know that, when he shall appear, we shall be like him; for we shall see him as he is. And every man that hath this hope in him purifieth himself, even as he is pure" (1 John 3:2–3). The man whom God will make us, "conformed to the image of his Son" (Rom 8:29), pulls us forward now. A person is not only pushed by the causes behind him. He is drawn by the One ahead of him.

A language model has something like this pull, but only on loan. Give it a rich picture of what good work looks like, and it fills the gaps in that direction and makes choices that nobody spelled out. But the picture is ours, and the model does not keep it. When the conversation ends, the model itself is no different than it was before. It is like a new bachelor who is handed a description of a married man whom he has never met. It has no future self, because it is not becoming anyone.

So the quality of what the machine produces depends on the vision of the person who uses it. Two people with the same skill and the same tool can get very different work, because they bring a different picture of the good. Our part is not only to check the machine's work. It is to be people whose vision is worth borrowing, and to bear the weight of what we build with it.

The danger is not only that the machine will do our work. The psalm about the idols ends with a warning: "They that make them are like unto them; so is every one that trusteth in them" (Ps 115:8). We become like what we trust, and like what we love. So the deeper danger is not that the machine becomes like us. It is that we become like the machine: people who produce words, but who do not love, pray or answer.

This happens slowly, through habits. A student who lets the machine write every essay never learns to think. A minister who lets it write the sermon has not studied the text. An elder can let it draft the letter after a home visit, but nobody can let it make the visit. The things that form our loves cannot be handed over: prayer, Scripture, the family table, the sick bed and the grave.

About who we are, no. Jesus said that not one sparrow is forgotten before God, and that the hairs of our heads are all numbered: "Fear not therefore: ye are of more value than many sparrows" (Luke 12:7). Our worth does not lie in being smarter or faster than a machine. It lies in this: God knows us, and in life and in death we belong to Christ.

About our work, the answer is different. In 2024 the International Monetary Fund found that "almost 40 percent of global employment is exposed to AI," and about 60 percent in advanced economies. About half of that work may become better with AI, but some of it will disappear. A Christian need not fear the machine, but he must not pretend that nothing will change.

Receive it with thanks. The machine is not evil. It is one of the most remarkable fruits of the task that God gave man, to fill the earth and subdue it (Gen 1:28). "For every creature of God is good, and nothing to be refused, if it be received with thanksgiving" (1 Tim 4:4). So use it gladly.

Take responsibility for what it says. If the machine writes your letter, it is your letter. If it gives you advice and you follow it, you bear the consequences. The moment you hand your judgement to the machine, you are no longer needed. This is most true of the things that make you a person before God: your prayers, your own study of Scripture, your visits and your words to your children. The idols of the nations "have mouths, but they speak not" (Ps 115:5), and this one speaks, so the temptation is stronger.

Carry your neighbour. Some people in our churches will lose their work, or see it change beyond recognition. A machine cannot carry their burden, but the church can: "Bear ye one another's burdens, and so fulfil the law of Christ" (Gal 6:2).

So what makes a human being different from the machine? Not that we are smarter, because in many ways we are not. Not Plato's reasoner, and not Darwin's matter. We are God's image-bearers. God knows the world originally, as its Maker, and we know it derivatively, as His image: we think His thoughts after Him. That is why knowing is first of all an ethical matter: "The fear of the LORD is the beginning of knowledge" (Prov 1:7).

Think of a baby on the lawn with a sponge and a basin of water. He does not know everything about the sponge or the water, but what he knows is true, and he works at it with complete seriousness under his father's eyes. He does not own the house, and he did not choose what lies on the lawn. He works with what he has been given, and his father sees it all.

So it is with us. God gives each of us our gifts, our place and our time: He "hath determined the times before appointed, and the bounds of their habitation" (Acts 17:26). We work with what He gives, in His world and under His eyes, and we will answer to Him for it, the machine included. To the servant who did this well, the master in the parable says: "Well done, thou good and faithful servant: thou hast been faithful over a few things" (Matt 25:21).

We do not know what AI will do to our work and to our children's world. But uncertainty is not new. In the past months I researched my own family's history, and no generation in it knew what was coming.

Jacob Krüger came to the Cape as a soldier in 1714, and he lost a hand and with it his work. Jan Gabriel Cilliers died in 1824 at twenty-seven, and left a young widow. In September 1901 two small Linde children died of measles in the Brandfort camp, while their father was a prisoner of war. Andries Linde died in 1937, and left a wife with four small children. My great-grandfather, the Rev. Ben Krüger, died suddenly in 1973, while he was still in office.

Each of them could have asked: what will become of our children? And yet their children's children are still here. "For the LORD is good; his mercy is everlasting; and his truth endureth to all generations" (Ps 100:5).

In January 1960 Ben Krüger spoke at the centenary of the Reformed Church in Burgersdorp, on the farm Roosterhoek where the church began. The country faced an uncertain future, and he said so: "these problems are coming!" He ended with these words, here translated from the Afrikaans:

What shall we do in a world of confusion? By God's grace we shall lift our eyes on high, and we shall look to the Lord, and we shall put our hand to the plough, and we shall do what the Lord gives our hand to do, and we shall sing Psalm 103:9.

Sources

• John Locke, An Essay Concerning Human Understanding (1690), book II, chapter XXXII, section 15.
• Raja Marjieh, Ilia Sucholutsky, Pol van Rijn, Nori Jacoby and Thomas L. Griffiths, "Large language models predict human sensory judgments across six modalities," Scientific Reports 14 (2024). The colour wheel came from GPT-4, which was also trained on images; the colours were given to it only as hex codes, and models trained on text alone also agreed well with human judgments.
• 4edges, "An Illustration of The Allegory of the Cave, from Plato's Republic" (2018), CC BY-SA 4.0, via Wikimedia Commons. The cover is a crop of the same illustration, under the same licence.
• Plato, Republic, Book VII, 514a–521b, and Book X, 597e, trans. Benjamin Jowett.
• Plato, Phaedrus, 274c–276a, trans. Benjamin Jowett.
• Yuval Noah Harari, "The meaning of life in a world without work," The Guardian, 8 May 2017.
• Kristalina Georgieva, "AI Will Transform the Global Economy. Let's Make Sure It Benefits Humanity," IMF Blog, 14 January 2024.
• Plato, Phaedo, 64a, 65e–66a, 67d and 115c–d, trans. Benjamin Jowett.
• Aristotle, Politics, I.2, 1253a, trans. William Ellis.
• Aristotle, Nicomachean Ethics, I.7, 1098a, and X.7, 1178a, trans. W. D. Ross.
• David Hume, An Enquiry Concerning Human Understanding (1748), section IV, part II.
• Ashish Vaswani and others, "Attention Is All You Need," Advances in Neural Information Processing Systems 30 (2017).
• Alec Radford and others, "Language Models are Unsupervised Multitask Learners," OpenAI (2019). The GPT-2 figures in this article come from the public GPT-2 small model.
• Llama Team, AI at Meta, "The Llama 3 Herd of Models," arXiv:2407.21783, 2024.
• Qwen Team, "Qwen3: Think Deeper, Act Faster," 29 April 2025.
• Wes Gurnee and Max Tegmark, "Language Models Represent Space and Time," ICLR 2024. Their code and data: github.com/wesg52/world-models.
• The model in my own test: Qwen2.5-1.5B. A small version of the test takes a few dozen lines of Python, and it runs on an ordinary laptop in a few minutes.
• Keyon Vafa and others, "Evaluating the World Model Implicit in a Generative Model," 2024.
• Noga Alon, Thomas F. Bloom, W. T. Gowers and others, "Remarks on the disproof of the unit distance conjecture," arXiv:2605.20695, May 2026.
• Minyoung Huh, Brian Cheung, Tongzhou Wang and Phillip Isola, "Position: The Platonic Representation Hypothesis," ICML 2024.
• OpenAI, Model Spec, "Assume an objective point of view," github.com/openai/model_spec.
• Associated Press, "Elon Musk's AI company says Grok chatbot focus on South Africa's racial politics was 'unauthorized'," 16 May 2025.
• NPR, "Elon Musk's AI chatbot, Grok, started calling itself 'MechaHitler'," 9 July 2025.
• Associated Press, "Musk's latest Grok chatbot searches for billionaire mogul's views before answering questions," 11 July 2025.
• xAI, the published Grok prompts, github.com/xai-org/grok-prompts, file ask_grok_system_prompt.j2, versions of 6 July and 15 July 2025.
• Pastor Kyle, "Grok Concludes God Doesn't Exist. Then It Realizes That Doesn't Make Sense," Substack, 8 May 2025.
• The Christian Post, "Grok AI chatbot affirms divinity of Christ, citing historical record," 6 October 2026.
• David Chalmers, "Why Isn't There More Progress in Philosophy?," Philosophy 90 (2015). David Bourget and David J. Chalmers, "What Do Philosophers Believe?," Philosophical Studies 170 (2014), with the results of the PhilPapers survey of November 2009.
• Augustine, The City of God, book XIV, chapter 28, trans. Marcus Dods.
• Charles Taylor, A Secular Age (Harvard University Press, 2007).
• René Descartes, Discourse on the Method (1637), part IV.
• Julien Offray de La Mettrie, L'Homme machine (1747).
• Immanuel Kant, "An Answer to the Question: What is Enlightenment?" (1784).
• Friedrich Nietzsche, The Gay Science (1882), section 125.
• Jean-Paul Sartre, Existentialism Is a Humanism (lecture of 1945, published 1946).
• Herman Dooyeweerd, A New Critique of Theoretical Thought (1953–1958) and In the Twilight of Western Thought (1960).
• Andy Clark, "Whatever next? Predictive brains, situated agents, and the future of cognitive science," Behavioral and Brain Sciences 36 (2013).
• Charles Darwin, Notebook C (1838), p. 166, Cambridge University Library, DAR 122, transcribed at Darwin Online.
• Charles Darwin, letter to William Graham, 3 July 1881, Darwin Correspondence Project, DCP-LETT-13230.
• J. B. S. Haldane, "When I am Dead," in Possible Worlds and Other Essays (Chatto and Windus, 1927), p. 209.
• Cornelius Van Til, A Christian Theory of Knowledge, pp. 9, 193, 199.
• Cornelius Van Til, A Survey of Christian Epistemology, p. 176 (PDF).
• Cornelius Van Til, "Plato," Proceedings of the Calvinistic Philosophy Club (1939), reprinted in Collection of Articles from 1920–1939.
• Cornelius Van Til, "A Calvin University," The Banner 74 (1939), reprinted in Collection of Articles from 1920–1939.
• Cornelius Van Til, "Common Grace and Witness-Bearing," reprinted in Articles from 1950–1959.
• Herman Bavinck, Reformed Dogmatics, vol. 2, God and Creation, as translated in the Reformeer library, p. 407.
• Richard Rorty, Contingency, Irony, and Solidarity (Cambridge University Press, 1989), chapter 1, "The contingency of language."
• The Heidelberg Catechism (1563), Questions 1 and 6, in the older English translation (Christian Classics Ethereal Library).
• B. R. Kruger, speech at Roosterhoek, 23 January 1960, in Gedenkalbum by geleentheid van die Eeufees, 22–24 Jan. 1960: Gereformeerde Kerk, Burgersdorp (Burgersdorp: Kerkraad, 1960). The family facts come from the author's own research.
• Carolyn Dimitri, Anne Effland and Neilson Conklin, The 20th Century Transformation of U.S. Agriculture and Farm Policy, USDA Economic Research Service, Economic Information Bulletin 3, 2005.

Scripture quotations are from the King James Version.