AI-Proof: What Machines Can Never Be — Volume I: Awakening, Part II: The Machine at the Gate, Chapter 5: When Intelligence Became Cheap

Part II turns from the person to the machine. After the overview, this first chapter of Part II asks what happens to the value of human thinking when its price collapses.

← Previous: Part II: The Machine at the Gate (overview) · Series hub · Next: Chapter 6: What AI Will Replace →

Nothing is more useful than water, Adam Smith wrote, and nothing buys less. Intelligence is moving from diamond to water. Its price is falling fast. Its usefulness is not. Your wage was tied to the price, and your worth never was.

In Brief: For readers who wonder what their thinking is worth when a machine can do it for almost nothing. Chapter 5 is about an economic event: the price of trained thinking falling faster than any price in living memory. It asks what happens to the value of a thing when its price collapses, with Adam Smith’s water and diamonds beside Nārada, who knew every science of his age and was still “in grief”. It takes the optimists and the sceptics seriously, from Jevons’s coal to Acemoglu’s modest numbers, and keeps four words apart that the panic runs together: cost, price, usefulness and wage. Some of what stays scarce moves to choosing ends and answering, which no better machine removes; some to checking, for as long as checking stays hard; some to the owners of the machines. Whether anyone is paid for the scarce things is still being decided.

When thinking becomes cheap, what stays scarce is choosing what it is for and answering for what it does. Whether anyone is paid for that is still being decided, and a person’s worth never depended on it.

Opening: Three Weeks of Spreadsheets

Three weeks after the dashboard turned green, Priya Chandran was still opening her spreadsheets every morning.

She did not open them casually, the way you glance at a clock you know is right. She opened them with purpose. She compared the machine’s overnight findings with her own catalogue of edge cases, all fourteen spreadsheets and the private wiki that only she fully understood. She went line by line, category by category, for an hour each morning. This was the desk in Koramangala where she had spent eight years building the reputation her reviews called “irreplaceable”.

In three weeks she had not found a single difference.

She kept looking.

What she was doing was not checking. It was a ritual of relevance. Each morning the spreadsheets let her argue, silently, to nobody, that her catalogue still meant something. She had spent two years building it, and the machine had not erased it in eleven minutes. Or so the ritual said. The argument was losing, so she made it louder. She came in earlier. She stayed later. She wrote up edge cases the machine had already found, in her own format, as if her handwriting gave the problem a value the machine’s output did not have.

The worst of it was that she knew. The three seconds of Chapter 1, when she noticed her own fear, had not faded. The noticing sat under her busy mornings like a crack in a wall. She could feel it, and she worked harder so as not to feel it. Noticing without courage is its own kind of suffering.

The event that put her in that chair is the subject of this chapter. For most of history, trained intelligence was scarce. Whole economies were built around how rare and expensive it was. In a few years it has become cheap. So this chapter asks one question. When thinking becomes cheap, where does its value go?

5.1 The Automation Era vs. the Intelligence Era

Every earlier wave of automation had the same shape. A machine arrived that could do one specific thing faster and cheaper than a hand or a mind. The power loom, the calculator, the spreadsheet, the robot on the assembly line. Each one took over tasks that could be fully described in advance, where the inputs, the steps and the right output were all known.

Each wave left something untouched. Economists named it in 2003. Machines, they found, had taken over routine tasks, the ones that follow explicit rules. And they had increased the value of non-routine ones: solving unfamiliar problems, persuading, managing, deciding.¹ The engineer who designed the assembly line. The accountant who told the client what the numbers meant. The fund manager who decided which strategy the trading program should follow. In 1959 Peter Drucker named these people, as Chapter 2 recalled: knowledge workers, people paid to think. For them this was the high ground.

The high ground is flooding. What machines now do is not only the routine task but the thinking act: analysis, synthesis, drafting, design. They write a legal analysis of a new question. They read an unclear scan. They propose the structure of a program nobody has written before. Not perfectly, and not in every field. But well enough, and improving fast enough, to put the old bargain of the knowledge economy under a pressure it has never faced. The bargain was simple: trained thinking is rare, so trained thinking is well paid.

Is this a new kind of change, or only more of the old kind? Some researchers argue that today’s systems are very large pattern-matchers, and that what looks like reasoning is an effect of scale. Others argue that the distinction no longer matters. This book took that question seriously in Chapter 1 and will return to it in Chapter 8. For this chapter one point is enough. A junior lawyer whose research memo is no longer needed does not experience a difference between “real reasoning” and “pattern-matching” that produces the same memo. The economic change is real whatever we decide to call what the machine does.

5.2 The Collapse of Cognitive Scarcity

Consider two lengths of time.

The first is about twenty years. That is roughly how long it takes to produce one professional mind: birth, school, university, apprenticeship, and enough early mistakes to be trusted with hard work. A doctor, a lawyer, an engineer. The cost is carried by the person, their family, their society and the institutions that trained them. The result is one mind, in one body, able to work for perhaps forty or fifty years.

The second is a matter of months. That is roughly how long it takes to train a leading AI model. The cost is very large, hundreds of millions of dollars. But the result is not one mind. It is a system that can serve millions of people at once, in dozens of languages, all day and all night.

And once it exists, using it keeps getting cheaper. Researchers at Epoch AI tracked the price of reaching a fixed level of performance on standard tests. Between 2022 and early 2025 that price fell by between nine times and nine hundred times a year, depending on the task. On one demanding test of graduate-level science, the price of a given score fell about forty times a year.² No other price most of us have known has fallen like that. Three cautions belong with the numbers. They measure the price of reaching a fixed level, which means yesterday’s best, not today’s. And the fall was unequal: fast for some tasks, slow for others. The third is that these are prices. A price can sit below cost for a while, when sellers compete for customers or investors cover the loss. A price held down that way can rise again. But the cost underneath is falling too. Stanford’s AI Index for 2025 found the cost of the hardware falling by about 30 per cent a year. Its energy efficiency was rising by about 40 per cent a year.² Still, the direction is plain. The cost of a thought, of a certain quality, is falling toward the cost of the electricity that runs it.

This is the collapse, and it is an economic event. The scarcity that made trained thinking expensive, the slow biological work of forming a human mind, is being bypassed for a wide and growing range of tasks.

What happens to the value of a thing when its price collapses? Adam Smith asked a version of this question in 1776, and his answer has puzzled students ever since.

“Nothing is more useful than water: but it will purchase scarce any thing; scarce any thing can be had in exchange for it. A diamond, on the contrary, has scarce any value in use; but a very great quantity of other goods may frequently be had in exchange for it.”³

Smith had found what economists call the paradox of value. The most useful things are often the cheapest, and some of the most expensive are nearly useless. A century later economists explained it. Price is set not by how useful a thing is in total, but by how much one more unit is worth to someone who already has some. Water is cheap because there is plenty of it. A diamond is dear because there are few.

Intelligence is moving from diamond to water. For most of history a trained mind was a diamond: rare, and priced accordingly. Now its output is becoming water: more useful than ever, and cheaper every year. Not all of it, and not all at once. The output that can be checked against a known standard turns to water first: the summary, the first draft, the routine analysis, the standard contract. The Prologue used the same image and ended it with a filter. When water is everywhere, someone has to say which water is safe to drink. In the end, the filter is a person. Smith lets us ask the next question. What is the filter worth, when the water is free?

Four words need to be kept apart here, because the rest of this chapter depends on them. Cost is what it takes to produce something. Price is what someone will pay for it. Usefulness is how much good it does. A wage is what a person is paid for producing it. Smith’s paradox is that usefulness and price can move in opposite directions. When the price of thinking falls, its usefulness does not fall with it. The world still needs good analysis, sound diagnosis and honest advice as much as ever, probably more. What falls is the price that any one person can charge for producing it. And the wage followed the price, not the usefulness. The clerk was paid for a scarcity. When the scarcity goes, the usefulness of her memo does not write her next invoice.

There is a fifth word, worth, which belongs to persons and not to their output. It is none of the other four, and we will come to it.

So the question professionals must now ask is not “What can I do?” Most of them can do a great deal. The question is “What that I do is still scarce?” Markets pay for scarcity, not for usefulness. The print edition of this book answered: “consciousness”. That answer was too quick. Everyone is conscious. Consciousness cannot be scarce in a market of people.

When something becomes plentiful, scarcity tends to move to whatever the plenty depends on. When information became cheap, the economist Herbert Simon saw that it consumes “the attention of its recipients”, as Chapter 3 recalled.⁴ When thinking becomes cheap, three things it depends on stay scarce, at least for now.

The first is choosing ends: deciding what is worth pursuing, what must not be traded away, and what should not be optimised at all. When a machine can write a hundred drafts in a minute, the scarce act is to say “this one”, and to say why. That is the taste of Chapter 3, with a price attached. Sometimes the scarce act is to refuse the efficient answer: the fastest route, the cheapest supplier, the message most likely to be clicked. The second is checking: telling good output from output that only looks good. The third is answering: being the person who stands behind a conclusion, who will explain it, undo it where possible and put right what can be put right.

These three are not the same kind of thing, and the difference decides how long each stays scarce. Checking is a task. It needs skill, and skill can be built into machines. Systems already check other systems, and a second model can be set to hunt for the first one’s mistakes. As of 2026, checking has not become cheap in the way producing has. But this book’s rule is that what a machine cannot yet do carries a date, and checking carries one. Choosing ends and answering are different. They are not mainly questions of skill. Who may decide what a hospital, a court or a company is for is a question of right. The Prologue’s third never says that “being intelligent will never give a machine the right to decide for people”. A machine can generate a diagnosis; who may prescribe is not settled by who diagnoses best. And who answers when it goes wrong falls under the Prologue’s second never: a machine’s makers and users “will never be free of what it does”. A better machine changes neither.

Here is the case against this comfort, and it should be heard in full. Stated bluntly, it says that cheap intelligence makes average intelligence worthless. Most people’s working lives were built on average intelligence, applied steadily. The clerk, the junior analyst, the translator, the competent writer of reports.

And notice that the word “scarce” has been doing two jobs. In this book’s sense, choosing, checking and answering are scarce because each needs a particular person’s time. In a market’s sense, scarce means few positions. Both can be true at once. The price of thinking may move toward choosing, checking and answering. The number of people who are paid to do those things may not. A firm that needed nine engineers may now need three people who can choose, check and answer. Both facts are the scarcity. Only one of them is a living. And if machines learn to check, the three scarce things become two, and neither of those was ever a mass occupation. The right to choose ends is not spread evenly either. In most firms it sits with owners and managers. A clerk may be asked to check and to answer while the ends are chosen above her.

The early evidence does not say which way this goes. In one large study of customer-support agents, an AI assistant raised the number of problems solved per hour by 14 per cent on average. For the newest and least skilled workers the rise was 34 per cent. For the most experienced there was little effect.⁵ There, the machine narrowed the gap between the average worker and the best. In a field experiment with small-business owners in Kenya, an AI adviser did the opposite. The strongest performers did just over 15 per cent better, and the weakest did about 8 per cent worse.⁶ There, it widened the gap. Cheap intelligence may lift the average toward the expert, or let the expert pull away. As of 2026 it is doing both, in different places.

I do not have a comforting answer, and I will not invent one. Chapter 4 separated four losses: income, practice, standing and voice. Cheap intelligence threatens the first most directly, for the people whose income rested on average cognitive work. And value can move faster than people can follow it. A twenty-five-year-old analyst cannot become a trusted adviser in a year. Chapter 4 called formation the slowest of its three clocks. What this chapter can do is show where value is moving, so that nobody mistakes the falling price of their output for the falling worth of their life.

That last distinction is old, and its clearest statement is not an economist’s. Smith separated the usefulness of a thing from its price. In 1785 the German philosopher Immanuel Kant drew a sharper line, between things and persons. In the translation of T. K. Abbott:

“In the kingdom of ends everything has either value or dignity. Whatever has a value can be replaced by something else which is equivalent; whatever, on the other hand, is above all value, and therefore admits of no equivalent, has a dignity.”⁷

Kant’s German word for the first is Preis: price. A few lines later he says where work belongs. “Skill and diligence in labour have a market value.”⁷ Here his word is Marktpreis: market price.

Read that sentence slowly, because it was written more than two hundred and forty years before anyone could buy skill by the minute. Kant put skill on the market side from the start. Skill always had a price, which means it could always, in principle, be replaced by an equivalent. What is new is only that the equivalent has arrived, and it is cheap. Kant never placed a person’s dignity in their skill. So the falling price of skill does not touch it. Dignity is the fifth word: not a cost, a price, a usefulness or a wage.

I am applying Kant here, and I should be exact about what the application allows. For Kant, dignity belongs to every person as a being capable of moral choice. It does not depend on being employed, being useful or being needed. Chapter 4 said the same: a person’s worth does not depend on being able to answer. Nor does it depend on judging well. A book that replaced “you matter because you are clever” with “you matter because you judge well” would have changed the test, not removed it. This book removes it. In one place the book goes further than Kant. Kant writes that “morality, and humanity as capable of it, is that which alone has dignity.”⁷ Read strictly, that seems to leave out a person who will never wake from a coma. Critics press this point hard, and his defenders answer it in several ways. This book takes his line between price and dignity, not his ground for drawing it. It offers no new ground of its own. It starts from a commitment: worth belongs to every person, and depends on nothing a person can do or fail to do. Traditions that ground worth in God, in the self or in shared human life can all keep that line. But dignity does not pay rent. Kant tells you what was never for sale. He does not tell you how to live when what you were selling stops selling. That second question is an economic and political one, and it belongs to the rest of this book.

5.3 Knowledge Is Now a Commodity

On a Thursday evening in January 2025, Rajan Subramanian took a call from a new buyer in Copenhagen. The company wanted brushed cotton for a winter range of bedding. It had found Rajan the old way: a name passed between importers at a trade fair in Frankfurt.

The call lasted forty minutes. The buyer had exact requirements: an organic certificate that European shops would accept, a particular weight and feel, delivery to a warehouse in Rotterdam by March. Rajan answered every question from memory. Which certificate, and how long it took. Which three spinning mills could supply certified organic yarn in that quantity. Whether his finishing house in Karur could brush the cloth to the feel they wanted. Why the ship should go through Colombo and not direct from Chennai to meet the date. The name of the freight agent who knew the Dutch customs papers. More than thirty years of cotton, poured into forty minutes.

He put the phone down feeling good. It was the old feeling: being the person in the room who knows.

Karthik phoned on Sunday, as he always did. He had been trying an AI trade assistant, built on shipping records, customs data, certification rules and twenty years of trade journals. Not to test his uncle, he said, but to test the machine. He had asked it the Copenhagen buyer’s questions, and he read the answers aloud.

The certificate: right. The finishing: right, though it named the process and not the house. The yarn: two of Rajan’s three mills, and a mill in Karnataka that Rajan had not considered. The shipping route: right, including Colombo. The freight agent: a different name, but a good one, at similar rates.

Rajan listened. Then he said: “The machine does not know about the mill in Karnataka. I bought from them two monsoons ago. The yarn was uneven. I saw what it did on my looms.”

Karthik was quiet. Then he asked something Rajan had not expected. “Two monsoons is a long time. Have you tried them since?”

He had not.

Rajan had survived the end of the quotas, 2008, the cash withdrawal of 2016 and the pandemic. Survival had taught him that comfort is the enemy of clear sight. So he said the true thing. “Then that is one answer out of six where I am not sure. The machine got the other five.”

He was right about the five. About the one, he did not know, and Karthik had shown him why. His own knowledge had a date on it, like everyone else’s. What he knew about the Karnataka mill was true as of the monsoon of 2023. The machine’s knowledge had a date too. Both needed checking. The difference was that Rajan knew his was dated and could check it: a sample order, a visit, a call to someone he trusted. The machine’s answer carried no date at all, and nobody would check it unless someone thought to ask. Some systems now name their sources, with dates, and that is a real gain. This one did not. And a date on a source says when something was written, not when anyone last went to look.

On Monday he wrote to Copenhagen. He named his three mills and added the fourth, the one the machine had found. Beside it he wrote one line: I had a bad lot from this mill in 2023 and have not tried them since. They may be cheaper. I can order a sample if you wish. He did not know whether he was being fair to the mill. He knew he was being honest about the date. The choice was the buyer’s, because the risk was the buyer’s.

What happened in that phone call is what economists call commoditisation. A commodity is a product that is the same whoever sells it, so that buyers choose on price alone. Rice of a certain grade, cotton of a certain staple, steel of a certain strength. Knowledge of the kind that lives in records, rules and databases is becoming a commodity. Five of Rajan’s six answers are now available to anyone with a subscription. He did not lose them. They simply stopped being his alone.

An old Indian text tells this story better than any economist.

In the seventh chapter of the Chāndogya Upaniṣad, a sage named Nārada comes to the teacher Sanatkumāra and asks to be taught. Sanatkumāra asks him first to say what he already knows. Nārada gives a list that is, for its time, the whole of learning. In F. Max Müller’s translation:

“I know the Rig-veda, Sir, the Yagur-veda, the Sâma-veda, as the fourth the Âtharvana, as the fifth the Itihâsa-purâna; the Veda of the Vedas (grammar); the Pitrya (the rules for the sacrifices for the ancestors); the Râsi (the science of numbers); the Daiva (the science of portents); the Nidhi (the science of time); the Vâkovâkya (logic); the Ekâyana (ethics); the Deva-vidyâ (etymology) …; the Nakshatra-vidyâ (astronomy) … All this I know, Sir.”

Then he adds: “But, Sir, with all this I know the Mantras only, the sacred books, I do not know the Self. I have heard from men like you, that he who knows the Self overcomes grief. I am in grief.”⁸

Sanatkumāra’s answer is the line this chapter needs. “Whatever you have read, is only a name.”⁸ The Sanskrit word is nāma, name: the whole of what can be said, listed, recited and passed on.

Notice what the teacher does next, because it is not what we expect. He does not tell Nārada that his learning is worthless. He tells him to revere it. “He who meditates on the name as Brahman, is, as it were, lord and master as far as the name reaches.”⁸ Name has real power, and a real range. Then Nārada asks the question every professional is now asking. “Sir, is there something better than a name?” Sanatkumāra says yes. He begins to lead him up a long ladder, one step at a time.

I am applying this text, not reporting it, and I should say exactly where the fit ends. The Chāndogya is leading Nārada toward the Self and toward release from grief, not toward a better career. A machine does not hold knowledge as Nārada held it, as a lifetime’s relation to learning. It holds representations of every recorded science, in every language. Nārada’s grief is not Priya’s, either. His is that he does not know the Self. Hers is that her catalogue no longer sets her apart. What fits is narrower, and it is enough. Name has a real range. The teacher reveres it within that range. And his answer is neither “your learning is useless” nor “your learning is enough”. It is: this has its range; now let us see what lies beyond it.

Sanatkumāra’s ladder leads to the Self, and this book does not climb it. What lies past the range of the name, for this book, is smaller and more ordinary. It is not a secret faculty, and it is not the Self. It is what Rajan did with Karthik’s question. Knowing that his knowledge had a date. Saying so, in writing, to the person whose risk it was. Being the person the Copenhagen buyer could phone in March if the cloth did not hold. That is the name used with care, not the name surpassed. Knowing everything that can be named is not the same as knowing what the knowing is for. Machines can hold the names, and they will hold more of them every year. A machine can even answer the phone, and it may soon answer it better than Rajan. But when the cloth fails, the buyer is not looking for an answer. She is looking for someone who will answer for it, and whose name will suffer if he does not. However good the machine becomes, that someone is a person, or a company, which is people under one name: Rajan, or whoever sold the machine’s advice. The law may one day name a machine in a dispute, as it names a company. The explanation, the money and the remedy would still come from the people who made it and used it. This is not a fact about 2026. It is the Prologue’s second “never”.

5.4 Execution Is No Longer a Moat

In business, a moat is whatever protects a company from competitors, as the water around a castle protects its walls. A brand people trust, a patent, a cost nobody else can match. For many firms, the moat was simply doing the work faster and better than anyone else. That moat is drying up.

In Kigali, Diane Uwamahoro built a software firm on speed. Her team of nine could build a payment connection for an East African business in three weeks, from first brief to working system, with clean code and good documents. In a region where software projects often stalled for months, three weeks was a weapon. Her clients came back. She hired her ninth engineer on a Tuesday in October 2024. For the first time in the three years since she had started, she felt she was building something that would last.

Six months later a competitor opened with three people and AI coding tools. It could produce in four days what Diane’s team produced in three weeks.

The code was not better than hers. It was competent. The documents were adequate. The design followed patterns any experienced engineer would recognise as sound, because the tools had learned from millions of programs that had established what sound looked like. The competitor did not know what Diane knew about the patchwork of rules across Rwanda, Uganda and Kenya. Where mobile-money systems failed to connect. Which banks’ systems said one thing and did another. Which compliance officers wanted a human being in the room before they would sign. But for the simple jobs, the ones that followed known patterns and fitted a clear brief, the competitor was faster and cheaper, and the market noticed.

Diane lost four clients in two months. Not the difficult ones. The simple ones: the ones whose fees had paid for the difficult work. By the end of that year she had let three engineers go. She told each of them herself, in her office in Kimihurura, with the hills dark outside. One of them had been her first hire.

She kept the difficult clients. She kept the projects where the brief was unclear, the rules contradicted each other, and the meaning of “finished” had to be agreed. In one of those meetings a compliance officer at a bank asked her a question she had heard many times, and only now understood. “If this breaks at the end of the month,” he said, “who do I call?” He was not asking who could write the code. Anyone could write the code now. He was asking who would answer. Part of what he wanted was liability: a firm that could be held to a contract and made to pay. Insurance can sell that. The other part cannot be bought from an insurer: someone who would explain what broke, fix it, and put right what it cost.

There is a lesson here, and a warning inside it. The lesson is that when doing becomes cheap, what remains valuable is what surrounds the doing: seeing what the problem really is, deciding what to build, and standing behind it. The warning is about speed itself. When every firm has the same tools, everyone becomes fast, and speed stops being an advantage. Then the danger changes. A firm that moves fast without looking, because it can, becomes the firm that ships the error at scale. Fast is no longer rare. Careful is. But rare is not the same as paid, and Diane is smaller.

Diane’s story is not a story of a winner. She is smaller now, and three people she trained work elsewhere. What she kept, she kept because she was the one the compliance officer could call. What she let go was the place where the next person would have learned to deserve that call. Her first hire had learned on exactly the simple jobs the competitor took. This is Chapter 4’s missing rung, seen from the employer’s chair: the expert keeps the work, and the apprentice’s first step disappears. None of it is a promise that everyone like Diane will survive. It is a description of where her value moved, and of what it cost to move it.

5.5 The Rise of AI as Infrastructure

You have probably used AI today without choosing to.

Not the kind you open on purpose and type into. The other kind, which was already running when you arrived. The email program that suggested your reply. The phone camera that adjusted the light, smoothed the shadows and sharpened the face before you ever saw the picture; the raw image you will never see. The search page that answered your question with a paragraph instead of a list of other people’s pages. The system that approved your last payment in a fraction of a second. The hiring platform that scored a CV before any person opened it.

When AI first reached the public, it was a product: a thing you chose to use. You opened it, asked, judged the answer, and decided what to do. You were the operator. That has not ended. But a second stage has begun beside it. AI is becoming infrastructure: part of the roads and pipes of daily life, running inside the tools people already depend on, below the level of notice.

The difference matters for this book, because of what it does to choice. When AI is a product, a person decides when to use it. When AI is infrastructure, a person decides only when to override it, and only if they notice it at all. Most of the time, the choice has already been made, by the default: what happens if nobody does anything.

Defaults are more powerful than they look. In a well-known study published in 2003, two psychologists compared European countries with two kinds of rule for organ donation. In some, people had to opt in to become donors. In others, they were donors unless they opted out. In the opt-in countries, between about 4 and 28 per cent of people had agreed. In the opt-out countries, between about 86 and almost 100 per cent had.⁹ The people were not very different. The default was.

Now imagine every default in a working day set by a system. The hiring manager who never sees the candidates the platform scored low is not making an independent judgement. She is approving one that was made for her. The photographer who never sees the raw image is not making an artistic choice. He is accepting the software’s taste. The hiring manager is no longer making the decision. She is deciding whether to interrupt one, if she notices it at all. Nobody decided to give up these choices. They were given up by the way the tools were built.

This is a new road to the abdication the Prologue warned about. In the Prologue, abdication was a decision: “the machine decided” offered as the end of an answer. With infrastructure, abdication can happen without any decision at all. And this is where the Prologue’s second “never” applies with full force. A machine’s makers and users “will never be free of what it does”. When a choice is built into a default, the people who chose the default answer for it, whether or not anyone ever noticed them choosing. A decision hidden inside a default is still a decision. Someone chose the score, the threshold, the acceptable rate of error, and whether a rejected person could appeal. Hiding the choice does not make it nobody’s. It only makes it harder to see whose it was.

None of this makes a default wrong. Most defaults are delegation done well. Nobody wants to approve every payment by hand. A default becomes abdication when nobody can say who set it, or who answers for what it does.

So the role of the person changes. The operator started the process. The governor watches it, understands roughly what it is doing beneath the surface, and knows when the cost of trusting it is higher than the cost of questioning it. The operator needed skill with the tool. The governor needs to notice that there is a tool at all. And notice who never gets the chance to be a governor: the candidate the platform scored low. She never sees the tool, and unless a law requires it, nobody tells her it was there. Watching a default is a privilege that the default itself hands out.

5.6 The Marginal Cost of Intelligence Approaches Zero

The marginal cost of something is what it costs to produce one more of it. For a lawyer’s memo written by a junior associate, the marginal cost was a day or two of a trained person’s time. In a large city that was a great deal of money. For a memo of similar form produced by a machine, the marginal cost is a few minutes of computing and a small amount of electricity. A human still has to check it. But the cost of the first draft has fallen by a factor that most professions have never seen.

The same thing is happening, at different speeds, wherever the main product is a document, an analysis, a design or a piece of code. The cost is moving from human time to computer time, and computer time keeps getting cheaper.

“Approaches zero” needs two corrections. First, it approaches a floor, not zero. The machines run on electricity. The International Energy Agency estimated that data centres used about 415 terawatt-hours of electricity in 2024. That is about 1.5 per cent of the world’s total, and it could more than double by 2030.¹⁰ Thinking has become cheap partly by moving its cost onto the power grid and the climate. Second, the cost of checking has not fallen with the cost of producing. A memo that takes three minutes to draft may still take an hour to verify, and the hour belongs to a person. And as producing gets cheaper, there is more to check. The cheaper it becomes to produce answers, the more of our time goes into deciding which of them deserve belief. That is true as of 2026. Whether machines will make checking cheap as well, as 5.2 asked, is one of the open questions on which the rest depends.

What happens when the price of something essential collapses? History offers two kinds of answer, and both deserve to be heard.

The first is the optimists’. In 1865 the English economist William Stanley Jevons noticed something odd about coal. Steam engines were becoming far more efficient, and people assumed this would mean Britain burned less coal. Jevons argued the opposite. “It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth.”¹¹ When something becomes cheaper to use, people find many more uses for it, and total use rises. This is now called the Jevons paradox. If it holds for thinking, cheap intelligence will not mean less thinking. It will mean far more of it: more analyses, more drafts, more decisions, and therefore more need for people to choose among them, check them and answer for them.

There is history to support this. When cash machines spread through American banks from the 1970s onward, people expected bank tellers to disappear. They did not. The number of tellers held steady and even rose for years. Each branch needed fewer tellers, so banks could afford to open more branches. And the tellers moved from counting cash to helping customers with problems.¹² The task was automated. The role changed and survived, at least for a long while. Later the numbers fell. Tellers in the United States held about 560,000 jobs in 2010 and about 339,000 in 2025. The US Bureau of Labor Statistics links the fall to online banking and to branches closing.¹²

Economists put the general point in two words. A machine can substitute for a person, doing the task instead of her. Or it can complement her, making her more productive at some other task. The cash machine did both: it replaced the counting and freed the teller for the customer. Most machines do both, in proportions nobody can know in advance, which is why the history is so hard to read.

The second answer is the sceptics’. In 2024 the economist Daron Acemoglu estimated how much AI would raise total factor productivity in the American economy over ten years. That measure is output per unit of labour and capital taken together, so its growth is the part of output growth that more labour and capital do not explain. It is not output per worker. His answer: no more than about 0.66 per cent, and probably less.¹³ On his reading, cheap thinking will change many tasks but add surprisingly little to total wealth. Meanwhile it may shift a good deal of income from workers to the owners of the machines. Other economists think his numbers are far too low. The argument is open, and it should be stated as open.

Under the argument about how much there is a plainer question: who gets it? If cheap thinking makes work more productive, the gain can go to several places. To customers, as lower prices. To firms, as profits. To workers, as wages. To the owners of the machines, as rent. Or to the public, through taxes. Usually it goes to some mix. Scarcity may also move away from people altogether, to whoever owns the computing power, the data and the customers. None of this is decided by the machines. It is decided by competition, contracts, tax law and the bargaining power of the people involved. Chapter 4 said that economics and law decide who gets the gains. This is where that sentence comes due.

Both answers carry a lesson about time. The last great transition of work, from the farm, took generations. In the United States, farm workers were the great majority of the labour force when the country was founded. Today they are a small fraction.¹⁴ The people who left the land moved into factories, then offices. Their descendants moved into the very knowledge work that is now becoming cheap. That transition made the country far richer. It also ruined many families who could not move, and the ruin lasted longer than the hopeful stories admitted. And it is not finished everywhere. In India, nearly half of all workers still work in agriculture.¹⁵ Transitions are neither universal nor quick.

The present change has one feature the farm transition did not. It is fast. Farm work declined over a century. Office work was computerised over decades. Cognitive work is being cheapened in years. The old reassurance, that technology creates more jobs than it destroys, may still hold over a long enough period. Whether it holds for the particular people working now, in the particular years they have, is a question honesty requires this book to ask and not to answer.

India is placed at a sharp point in this change. For thirty years, part of its economic success was built on a simple difference in price. A trained Indian engineer, accountant or support agent cost much less than one in Europe or America, and the work could be sent down a cable. Economists call making money from such a difference arbitrage. In 2019 the economist Richard Baldwin gave the next stage a name: “globotics”. By this he meant globalisation and robotics arriving together in office work, with remote workers abroad and machines at home competing for the same tasks.¹⁶ Cheap intelligence is cheaper still than cheap skilled labour. Priya’s company, her team of twelve, and the IT industry around Koramangala live inside that difference in price. Chapter 4 recorded the first signs: Indian IT firms employing fewer people, and hiring fewer graduates. What the industry does next is not decided by machines. It is decided by what Indian firms choose to sell when trained thinking is no longer the cheap thing they can offer.

So where does value go, when the marginal cost of thinking falls? Not to “consciousness”, which every person has. It splits. Usefulness stays where it was, or grows. Price falls. Some of what stays scarce moves to choosing ends and answering, which no better machine removes. Some moves to checking, for as long as checking stays hard. Some moves to the owners of the machines. The wage follows none of these by itself. And worth, in Kant’s sense, never moved at all.

Choosing ends and answering need a person whose time is limited, whose name is on the result, and whose losses, as Chapter 3 put it, cannot be restored from a copy. That is also why answering cannot be spread thin. A manager who signs for a million decisions she never saw is not answering for them. She would carry blame without control, the place Chapter 1 called a moral crumple zone. Whether a society pays such people, forms enough of them, and lets the gains reach anyone else is not a question machines answer. Cheap intelligence makes one input plentiful. Whether it makes lives more plentiful depends on what else stays scarce: attention, trust, opportunity, ownership.

Under the hood

This chapter has called checking a task, scarce only for as long as it stays hard. How hard it is depends on where the truth lies.

In some work the truth lies inside the text. A sum is right or wrong. A program passes its tests or fails them. There a program can do the checking, millions of times, for almost nothing. One team that built a reasoning model described exactly this in January 2025. It rewarded the model for a final result in maths, written in a set format, that a rule could mark, and for code that passed tests run by a compiler. It chose not to use a learned judge, because a model in training could learn to game one.¹⁶

In other work the truth lies in the world. Whether the cloth held its feel, or the patient recovered, is written in no text. No program can mark it from a desk. Someone has to go and look, as Rajan could with a sample order, and the looking takes a person’s time.

So checking gets cheap at different speeds in different work. Where a rule can mark the work, it is cheap already. Where the truth has to be found in the world, checking costs as much as finding it. That is the date that checking carries.

5.7 The Shift from Information to Judgement

Three weeks and two days after the dashboard turned green, Priya did something new. It was a Thursday morning, and the rain had left the office humid and close.

She did not open her spreadsheets.

She sat at her desk for a full minute, hands on the keyboard, the familiar tabs waiting. She could feel the pull. The spreadsheets were the shape of her professional self: the proof that she was the one who knew where the product broke. Opening them would have felt like putting on a coat on a cold morning. Easy. Warm. Wrong.

She opened a blank document instead, and she wrote a list. Not of edge cases. Of decisions she had made that week.

The bug in the payment module she had escalated on Monday. It was rare, but it touched a tax calculation for users in three states. She knew from years of support tickets that those users complained loudly. She also knew that the product team’s standing with the finance team depended on that module.

The reply to a client in Pune, on Tuesday. Their invoices had failed twice that month. The support system had drafted an answer that was accurate and polite, and promised nothing. She added two things: a date by which the fix would ship, and her own name. If the date slipped, the client would now call her.

The slowdown she had let pass on Wednesday. The module was due to be rebuilt next quarter. Fixing it now would delay a feature the sales team had already promised two large clients.

The cluster of tiny failures she had flagged late on Wednesday. Each was below the threshold that mattered. Together, she thought, they suggested a weakness in the design that would grow under load. She had spent twenty minutes on a call explaining to a senior engineer why many acceptable failures could add up to an unacceptable one. She could not prove it. She said so, and he agreed to watch it for a month.

And one more, which she nearly left off. On Tuesday the machine had marked a test failure as “low risk”. Her instinct said it was serious; she had seen something like it in her second year. She spent an hour on it before admitting that the machine was right. The resemblance was on the surface. She had wasted an hour and learned something about her own memory.

Five decisions. After each one she wrote her reasons: not the data, which the machine had supplied, but what the data did not carry. The history of the product. The politics between teams. Which customers would be hurt, and how badly. Which risks were real and which were noise.

The list filled a page.

She looked at it the way she had once looked at the green dashboard. But the feeling was different. The dashboard had brought fear: the floor shifting, the thing she had dreaded made real. The page brought something quieter. Recognition. These were decisions somebody had to make and stand behind. The machine had supplied every fact on the page. It had not supplied the choosing, and it would not be the one asked about the choices if a customer in one of those three states lost money.

She saved the document. She called it Decisions: week of October 21. She did not know where it would lead. She knew only that the spreadsheets were closed and something else was open.

The question that had lived in her since the dashboard turned green had changed shape. It had been: If the machine does not need what I know, what am I for? Now it was: What does the work need that the machine cannot sign?

It was a better question. It was not yet an answer, and it paid nothing. Nobody at the company had asked for a page of decisions, and nobody would read it. A page that only she would read was still a ritual: a better one, with the catalogue taken out. Only one line on it had left her desk: the date and the name she had sent to Pune. That was the line someone could hold her to.

She had been making judgements for years. She had called it testing. Now the testing was automated, and the judgement stood on its own, exposed, like a rock at low tide. Some of it, as Tuesday had shown, was not as good as she had thought. And nobody else could see the rock yet.

Practice

Put a date on what you know. Choose three things that people at work come to you for: a supplier you trust, a rule you know, a client you understand, a system only you can fix. Beside each, write the date you last checked it yourself. Not the date you last remembered it: the date you last checked it. Then put the same question to a machine. Note what it says, and whether its answer carries any date at all. Where your date is old, check it this month, with a call, a visit or a sample. Where the machine’s answer is better than yours, write that down too. At the end you will not have a score. You will have a list of the places where your knowledge is still yours, because you keep it fresh. And you will see the places where it has quietly become a name. Be honest about what the list can do. It tells you where you stand. It does not promise that standing there will pay. If all three items come back stale, the exercise has told you something no exercise can mend alone.

Ending: Two Decisions on Thursday

The next morning Priya opened one spreadsheet, out of habit, and closed it again without reading it.

Then she opened the page of decisions and added Thursday’s. There were two. She was not sure about the second.

Words to keep

• answering: standing behind a decision: explaining it, undoing it where it can be undone, and putting right what can be put right.

• arbitrage: making money from a difference in price between two places, such as the gap between the cost of skilled work in India and in Europe.

• checking: telling good output from output that only looks good.

• choosing ends: deciding what the thinking is for, and what is worth thinking about at all.

• commodity, commoditisation: a product that is the same whoever sells it, so that buyers choose on price; the process by which something becomes one.

• cost, price, usefulness, wage: what it takes to produce something; what someone pays for it; how much good it does; what a worker is paid for producing it. They can move in different directions.

• default: what happens if nobody does anything.

• dignity and price (Kant): what has a price can be replaced by an equivalent; what has dignity cannot. For Kant, skill has a market price; persons have dignity.

• governor: the person who watches a process run by a machine and knows when to question it. Compare operator, who starts the process.

• infrastructure: the roads and pipes of daily life; things we depend on without noticing them.

• Jevons paradox: when using something becomes cheaper, total use often rises rather than falls.

• liability: being bound by law or contract to pay or put right when something goes wrong. It can be bought, through contracts and insurance. Answering cannot.

• marginal cost: what it costs to produce one more unit of something.

• moat: whatever protects a business from competitors, like the water around a castle.

• paradox of value: the most useful things, like water, are often the cheapest, and some of the most expensive, like diamonds, are nearly useless.

• routine and non-routine tasks: tasks that follow explicit rules, and tasks that do not.

• substitute and complement: a machine substitutes for a person when it does her task instead of her; it complements her when it makes her more productive at another task.

• worth: what a person has as a person, apart from any price. Kant’s word is dignity.

• nāma: name; in the Chāndogya Upaniṣad, all that can be said, listed and learned.

Sources

1. David H. Autor, Frank Levy and Richard J. Murnane, “The Skill Content of Recent Technological Change: An Empirical Exploration,” Quarterly Journal of Economics 118, no. 4 (2003): 1279–1333.

2. Ben Cottier, Ben Snodin, David Owen and Tom Adamczewski, “LLM inference prices have fallen rapidly but unequally across tasks,” Epoch AI data insight, 12 March 2025: declines of “9x to 900x per year” depending on the performance milestone; GPT-4-level performance on PhD-level science questions “fell by 40x per year.” The authors note that the fastest drops were in the most recent year. On the costs beneath the prices: Stanford HAI, The 2025 AI Index Report, Top Takeaways, p. 4: “At the hardware level, costs have declined by 30% annually, while energy efficiency has improved by 40% each year.”

3. Adam Smith, An Inquiry into the Nature and Causes of the Wealth of Nations (1776), Book I, Chapter IV, paragraph 13 (Glasgow edition I.iv.13), original spelling.

4. Herbert A. Simon, “Designing Organizations for an Information-Rich World,” in Martin Greenberger (ed.), Computers, Communications, and the Public Interest (Baltimore: Johns Hopkins Press, 1971), pp. 40–41; see Chapter 3, 3.5.

5. Erik Brynjolfsson, Danielle Li and Lindsey Raymond, “Generative AI at Work,” Quarterly Journal of Economics 140, no. 2 (2025): 889–942 (first as NBER Working Paper 31161, 2023): productivity “by 14% on average,” with “a 34% improvement for novice and low-skilled workers” and “minimal impact on experienced and highly skilled workers.”

6. Nicholas G. Otis, Rowan Clarke, Solène Delecourt, David Holtz and Rembrand Koning, “The Uneven Impact of Generative AI on Entrepreneurial Performance,” Harvard Business School Working Paper 24-042 (December 2023), a field experiment with Kenyan entrepreneurs: “high performers benefited by just over 15%,” while “low performers did about 8% worse.”

7. Immanuel Kant, Groundwork of the Metaphysics of Morals (1785), Second Section, Akademie edition 4:434–435; trans. Thomas Kingsmill Abbott, Fundamental Principles of the Metaphysic of Morals (1873; Project Gutenberg eBook 5682). German: “Im Reiche der Zwecke hat alles entweder einen Preis, oder eine Würde”; “Geschicklichkeit und Fleiß im Arbeiten haben einen Marktpreis.” Later translators (for example Mary Gregor, Cambridge, 1997) render Preis as “price”. Abbott also has “Thus morality, and humanity as capable of it, is that which alone has dignity” (4:435). On whether this ground leaves out human beings with severe cognitive disabilities, “one of the most important criticisms of Kant’s moral theory”, and the replies of his defenders, see Adam Cureton, “Kant’s Moral Philosophy,” Stanford Encyclopedia of Philosophy (substantive revision 2 October 2025), section 11, “Non-rational Beings and Disabled Humans”.

8. Chāndogya Upaniṣad VII.1.1–5, trans. F. Max Müller, The Upanishads, Part I, Sacred Books of the East 1 (Oxford: Clarendon Press, 1879). Some items in Nārada’s list are omitted (marked …); Müller’s glosses are kept.

9. Eric J. Johnson and Daniel Goldstein, “Do Defaults Save Lives?”, Science 302, no. 5649 (2003): 1338–1339: effective consent rates of 4.25–27.5 per cent in opt-in countries and 85.9–99.98 per cent in opt-out countries.

10. International Energy Agency, Energy and AI (10 April 2025), executive summary: data centres used about 415 TWh in 2024, about 1.5 per cent of world electricity, projected to reach about 945 TWh by 2030.

11. William Stanley Jevons, The Coal Question (London: Macmillan, 1865), chapter VII, “Of the Economy of Fuel”. Wording checked against the 1865 text as transcribed by Yale’s Energy History project and against the second edition (1866), chapter VII, at the Library of Economics and Liberty, which read alike; the shorter form often quoted (“It is a confusion of ideas … equivalent to diminished consumption”) is not the text of either.

12. James Bessen, “Toil and Technology,” Finance & Development (IMF), March 2015; summarised in Timothy Taylor, “ATMs and a Rising Number of Bank Tellers?”, Conversable Economist, 3 March 2015. Later figures: US Bureau of Labor Statistics, Occupational Outlook Handbook, “Tellers”: “Tellers held about 560,000 jobs in 2010” (archived 17 October 2012, UNT Libraries Web Archive) and “about 339,200 jobs in 2025”, with the number of branches “in decline due to technological change” and more people using “online banking tools” (page modified 27 August 2026).

13. Daron Acemoglu, “The Simple Macroeconomics of AI,” NBER Working Paper 32487 (May 2024): “no more than a 0.66% increase in total factor productivity (TFP) over 10 years,” and, allowing for harder tasks, “less than 0.53%”; published in Economic Policy (2025).

14. US Department of Agriculture, Growing a Nation: The Story of American Agriculture (National Agriculture in the Classroom), historical timeline: farmers about 90 per cent of the labour force in 1790 and 2.6 per cent in 1990.

15. Ministry of Statistics and Programme Implementation, Government of India, Periodic Labour Force Survey Annual Report, July 2023–June 2024: 46.1 per cent of workers (usual status) in agriculture and allied sectors, as stated in the Ministry of Labour and Employment’s answer to Rajya Sabha Unstarred Question No. 3760, 3 April 2025.

16. Richard Baldwin, The Globotics Upheaval: Globalization, Robotics, and the Future of Work (New York: Oxford University Press, 2019). For “Under the hood” in 5.6: DeepSeek-AI, “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning,” arXiv:2501.12948v1, 22 January 2025, section 2.2.2: a “rule-based reward system”; for maths, “the final answer in a specified format (e.g., within a box)”; for programming problems, “a compiler can be used to generate feedback based on predefined test cases”; no neural reward model, because it “may suffer from reward hacking in the large-scale reinforcement learning process.”

Named people in this chapter are composites, and their conversations and circumstances are constructed; the studies, laws and figures cited are real and sourced.

Revised from the print edition (2026)

• What the print edition claimed: that the answer to cheap intelligence “is not a set of skills. It is a quality of being. Consciousness”; that a junior associate’s London memo cost “roughly four hundred pounds” in 2019; that farm labour fell “from roughly ninety per cent of the population to fewer than three” (unlabelled as American); that the marginal cost of intelligence “approaches zero”; that Diane’s complex work was something “the AI could not do”; “the contradiction the Manifesto names”; “the next volume will map it”; that Priya had spent “six years” at her desk; and, in closing, “The intelligence was the floor. Consciousness is the ceiling.”

• What I now hold: cost, price, usefulness and wage are four things, and a person’s worth is a fifth. When the price of thinking falls, its usefulness does not; the wage followed the price (Smith’s paradox of value). What stays scarce is choosing ends and answering, which are questions of right and of the second never, and checking, which is a task and carries a date. Scarcity in the book’s sense is not scarcity in the market’s sense: few positions, and some of the gain may go to the owners of the machines. It does not move to consciousness, which every person has. The marginal cost of thinking falls toward a floor set by electricity and by the cost of checking, which has not fallen. The farm figures are American, and India still has nearly half its workers on the land. Kant placed skill on the market side and persons beyond price, so the falling price of skill does not touch a person’s dignity, though dignity does not pay rent. Rajan’s knowledge has a date, and he tells the buyer so. Both Jevons’s optimism and Acemoglu’s scepticism are stated, and the question for people working now is left open. Defaults turn infrastructure into a new road to abdication; a decision hidden in a default is still someone’s, and its makers answer for it. Diane’s lost simple work was also her apprentices’ first rung. Priya has eight years at the company, and her first page of decisions includes one where the machine was right. The right to choose ends sits mostly with owners and managers. Answering spread over decisions nobody saw is not answering. The book takes Kant’s line between price and dignity, not his ground for it, so worth reaches those who cannot choose.

• What changed my mind: the audit’s corrections to the figures; Epoch AI’s price data; the customer-support and Kenyan studies, which point in opposite directions; Smith’s water and diamonds; Kant’s price and dignity; the Chāndogya’s Nārada; Jevons, Bessen and Acemoglu read together; Johnson and Goldstein on defaults; the IEA’s figures on electricity; Baldwin and the end of the price difference India’s IT industry was built on.

• What survives: Priya’s ritual of relevance and her first page of decisions on 24 October 2024; Rajan’s “one answer out of six”; Diane’s four lost clients and the complex work she kept; the shift from operator to governor; and the reef, now a rock at low tide.

Next in the series: Chapter 6: What AI Will Replace.