AI-Proof: What Machines Can Never Be — Volume I: Awakening, Part II: The Machine at the Gate, Chapter 7: What AI Cannot Replace

Part II looks at the machine. Chapter 5 asked what cheap intelligence makes scarce, and Chapter 6 what the machine will replace. This seventh chapter asks what it cannot, and how long each answer will last.

← Previous: Chapter 6: What AI Will Replace · Series hub · Next: Chapter 8: Understanding the Machine You’re Working With →

The print edition of this book drew a line around what machines could never do. Some of that line has already moved. This chapter tests every “AI cannot” claim, puts a date on each, and finds what stays when the skills are matched.

In Brief: For readers who want to know which of their skills will still be theirs in ten years. Chapter 7 takes the strongest claims of the kind “a machine cannot”, the ones the print edition built on, and tests each one against the evidence of 2026. Judgement under uncertainty, taste, trust, the knowing body, the reading of feelings, and the price people pay for a person: each gets a date and a depth, the way an honest watch says how deep it can go. Frank Knight, Moravec, the Tirukkuṟaḷ and Gandhi help sort what is a skill a machine may soon have from what is a position no skill can fill. Then the chapter shows how a position can be emptied while every name stays on the form.

A machine’s skills carry dates; the positions it cannot fill do not: the one whose life it is, the one who answers, the one with the right to decide. Those positions last as long as people keep them full.

Opening: The Cloth That Resisted

Rajan Subramanian was standing in a mill in Tiruppur, running a bolt of cotton jersey between his thumb and forefinger. He had done this for more than thirty years. He visited mills, touched cloth, and made judgements with his hands that he had never been able to explain to anyone’s satisfaction, including his own.

The cloth was for a European retail chain. The knit was tight. The feel was good. But something was not right: a faint harshness in the finish, a resistance where the cloth should yield. He folded it, held it to the light, folded it again. The mill’s inspection camera had passed the batch. The colour readings were within the buyer’s limits. Rajan’s fingers said otherwise.

He could not prove it, and he said so. “I do not know what is wrong,” he told Senthil, who owned the mill. “I know it is not right. Run a wash test before you pack it.” Senthil’s laboratory ran accelerated wash tests on samples, the equivalent of twenty washes in hard water like the water in much of western Europe. It ran a shrinkage test beside them. The results came back the next evening. The colour had moved, and the cloth had shrunk beyond the buyer’s limit. No instrument on the line had flagged it. Rajan’s fingers had not proved anything either. They had asked for the test that did.

A month earlier, in a studio above a printing shop in Parel, in Mumbai, a designer named Arjun Mehta had sat looking at twelve packaging designs on his screen. A machine had made them for a drinks company: a modern rose sharbat for health-conscious city buyers. Clean type, sensible colours, every rule of good packaging followed. The client wanted to know whether any of them was right.

None of them was. Arjun knew it at once and could not yet say why.

The print edition of this book drew a line around scenes like these two and gave the line a name. It said the boundary of what machines can do “is defined by consciousness”, and that this was “not a temporary limitation” but “a structural one”. This edition cannot repeat that. The Prologue made a rule: this book never defines a human being by what machines cannot yet do. Chapter 1 showed how fast such lists shrink. So this chapter does something else. It takes the strongest claims of the kind “a machine cannot”, the ones the print edition built on here. Then it puts each through the test the Prologue promised: not “never”, but a stated depth, like an honest watch. Some claims will turn out to be dated. Some will survive, though not for the reason the print edition gave. So the chapter asks one question. What can a machine not replace, and how long will each answer last?

7.1 Original Judgement Under Genuine Uncertainty

On a December morning in 2024, in Kampala, Lillian Adeke was reading a due-diligence report in an office whose air-conditioning had been broken for three days. The heat made the paper limp. The report concerned a small company building soil-testing tools for farmers in eastern Uganda. Her fund’s AI screening system had scored it as moderate risk: promising technology, a crowded market, founders with no earlier success. Its recommendation was to pass. Lillian read the system’s reasons carefully, as she always did, and could not fault them. The market was crowded. Most companies like this one fail.

Lillian did not want to pass. She had spent three hours with the founders, and the hours had told her something the report did not contain. She had asked questions that appear on no checklist. What did the farmers tell you that surprised you? When you were wrong, how did you find out? What would make you walk away? The founders talked about particular farmers, by name, whose problems they had failed to solve on the first attempt. One of them described a woman near Mbale who had shown them, by pressing soil between her fingers, what their sensor had got wrong about her field. They had been changed by listening. Lillian had invested in East African technology for a decade, and she had come to believe that this kind of listening mattered more than the technology.

At the investment committee she said exactly that, and she said what she could not prove. “I cannot show you this in the numbers. I am asking you to fund my reading of two people.” Her partners argued for an hour. Then they agreed to invest half of what she had proposed, with a review in a year. By the autumn of 2026 the company was alive, behind its plan, and on its second version of the product. Lillian did not yet know whether she had been right. Nor will the outcome settle it when it comes. A good decision can have a bad result, and a bad decision a good one. Her judgement has to be weighed by what she knew and how she reasoned when she decided, not by what happens next.

In 1921 the economist Frank Knight drew a distinction that still organises this question. “It will appear that a measurable uncertainty, or ‘risk’ proper, as we shall use the term, is so far different from an unmeasurable one that it is not in effect an uncertainty at all.”¹ Risk is the kind of not-knowing you can put numbers on, because you have seen many cases like this one. The chance that a borrower defaults is risk, and so is the chance that a scan hides a tumour. Uncertainty, in Knight’s sense, is the kind you cannot, because cases like this one have not happened often enough to count. Lillian’s company was uncertainty. There was no long record of soil-testing start-ups in eastern Uganda to learn from.

The print edition built an argument on this. Machines handle risk, it said, and people handle uncertainty. That is half true, and the half that is false matters. Knight’s uncertainty is a feature of the world, not a weakness of one kind of mind. Nobody, human or machine, has frequencies for the unprecedented. A machine asked to judge Lillian’s company faces the same thin record she does. It can still weigh what it has, and the evidence that people do better in such cases is weaker than we would like. Thin evidence does not make a feeling more trustworthy. It makes it harder to check. Rajan, in 7.3, will show how a confident feeling about people can be wrong for years.

Here is the strongest case against Lillian, and it should be heard in full. In 2024 three researchers at the Massachusetts Institute of Technology combined the results of 106 experiments that compared people working alone, machines working alone, and people working with machines. On average, the combinations did better than people alone. But they did worse than the better of the two working alone. When the person was the stronger partner, adding the machine helped. When the machine was the stronger partner, adding the person hurt. And the losses were clearest in tasks that ended in a decision.² Chapter 3 told the older form of this finding. Since the 1950s, simple rules have matched or beaten experts in field after field, and experts given the right to overrule find too many exceptions. Lillian’s hour at the committee may have been exactly that. The study cannot say which partner is stronger in Lillian’s field. Only a record could, setting her past calls beside the screen’s on companies whose fate is now known.

So the claim “a machine cannot judge under genuine uncertainty” carries a date at best, and perhaps not a distant one. What the scene shows besides the skill is harder to name, and it is not one thing. The reading was Lillian’s, said in her own name. The right to decide was the committee’s, and it halved her bet. The money belonged to the fund’s investors. If the company fails, someone will have to say why. Judgement as a skill is a “for now”. Who owns a reading, who holds the right and who must explain the loss are different questions, and in Lillian’s case they had different answers.

7.2 Taste: Why Curation Requires Restraint

Arjun Mehta’s studio in Parel is small enough to see whole from the doorway. There are two rooms, two large monitors and a shelf of design books, some older than he is. One wall holds printed work he admires, none of it his own.

The marketing director at the drinks company had looked at the twelve machine-made designs and felt nothing. Not excitement, not recognition. She felt adequacy. And adequacy, in a market where every competitor uses the same tools to produce the same adequate options, is invisibility. That is why she had come to Arjun.

He spent two days not designing. He visited a sharbat maker on Mohammed Ali Road whose family had made rose sharbat for four generations. He watched the man tip a brass vessel so that the syrup threaded into the glass in one unbroken line. He noticed how the liquid caught the light differently in brass than in glass. He asked his mother about the sharbat she drank as a child in Jaipur. She described the taste in adjectives. She described the feeling by falling silent and looking out of the window.

Then he subtracted.

The machine’s designs had been full of information: ingredients in front, health claims prominent, flavour words stacked three deep, each element justified by packaging convention. Arjun removed most of it. He left a single image: a brass cup photographed from above, so that the drink caught the light. He kept the brand name, in letters drawn by a calligrapher in Ahmedabad, and one colour. The colour was not the saffron-gold that so much “premium Indian” packaging uses. It was a dusty rose, the shade of diluted rose sharbat in old brass. By the usual rules it was too quiet for a crowded shelf.

The marketing director saw it and knew. The design said one thing, this is real, and said it by leaving out everything else.

Notice that Arjun’s taste was not his alone. It came from the sharbat maker’s hands, his mother’s silence, a tradition of brass and rose older than either of them, and years of learning which rules to break. A machine trained on the work of many such traditions is not inauthentic merely because of where its taste came from. And choosing is not automatically choosing well. A person can pick the right design for the wrong reasons, and a machine can pick it for good commercial ones.

The print edition said that this was what the machine could not do: subtract. That claim has not lasted. Ask a model today for a design with one image, no claims and a single muted colour, and it will make one, and make it well. What the machine did not have was a reason to leave out the health claims the client had asked for, and anything to lose if leaving them out was wrong. Restraint is not the act of removing things. It is the decision about what this brand will not say, taken by someone who will be blamed if the silence fails. Chapter 3 called taste the courage to prefer. Restraint is that courage turned against your own client’s brief, with your name on the result.

There is a danger here, and it should be named. The same confidence that lets Arjun strip a design to its essence can also blind him. He has been wrong before, trusting his eye so completely that he missed what a client actually needed. The designer who leaves out the right things is exercising taste. The designer who leaves out necessary things because they offend his eye is exercising ego. From the inside they feel the same. So the check has to come from outside him, from a client with the power to refuse.

Whether the dusty rose would sell, nobody knew in January 2025. Meaning, beauty and sales are three different measures, and none of them had yet been taken. Subtraction as a skill is now cheap. What Arjun had added was not a skill but a choice someone could hold him to, and even that was shared: the marketing director could still have said no.

7.3 Trust: From Judging Reliability to Bearing the Cost

Rajan placed his first order with Senthil’s family in 1998. He was twenty-six, and the mill belonged to Senthil’s father. The father had enormous hands, scarred from decades in the dye house, and he would personally re-run a spoiled order through the night. Senthil’s hands are softer. He went to college. He runs the business while a floor manager runs production, and he spent his money on temperature controls, a quality engineer and written procedures.

When the son took over, Rajan did not trust him. He never said why. An instinct, unexamined, told him that a college man with soft hands would not carry his father’s word. For years he gave Senthil smaller orders than he gave others, and checked them more.

In early 2025 Karthik, his nephew, ran ten years of Rajan’s own records through a tool that scored suppliers by late deliveries, rejected lots and complaints. Rajan watched the list sort itself. Senthil’s mill was at the top. It had been the most reliable mill on Rajan’s books for a decade, including the years Rajan had doubted him. The records had judged Senthil better than Rajan’s feeling had.

About fifteen centuries ago, by most scholars’ reckoning, a Tamil poet put the rule Rajan had broken into two lines. The Tirukkuṟaḷ gives a whole chapter to testing people before trusting them, and ends it with this couplet, here in V. V. S. Aiyar’s translation:

“To trust a man whom thou hast not tried and to suspect a man whom thou hast found worthy lead alike to endless ills.”³

Rajan had done the second. He had a man’s record in front of him and suspected him anyway. A few weeks earlier, Rajan had turned down the young owner of a weaving unit near Erode, whose records were clean, and Karthik had asked him a question: would you have seen the same thing if you had known his father? With Senthil the question ran the other way. Rajan had known the father, and had judged the son against him. Chapter 3 gave a test for when an expert’s quick sense deserves trust: a world with clear signals, and years of feedback. Rajan’s fingers pass it. Thirty years of cloth and buyers’ complaints have taught them. His sense of people does not pass it so easily. Some of what he called instinct about Senthil was only an old picture of what a reliable man looks like.

Then came the dye lot. Senthil’s laboratory confirmed the failure, and Senthil phoned Rajan before Rajan could phone him. He showed him the affected bolts. He explained what had gone wrong: a temperature fault in the dye bath and a finishing machine set wrong, both already corrected. And he offered to re-run the whole order at his own cost. Rajan accepted, phoned the buyer in Europe, told her the shipment would be five days late, and took the delay on himself.

Chapter 3 said what trust is: accepting that someone could harm you. It also said that trust in a machine is trust in the people behind it. This scene adds a distinction those pages did not need. The record was better than Rajan at judging whether Senthil was trustworthy. Ten years of numbers saw what Rajan’s feeling missed. But the record could not make the phone call. It could not decide to disclose a fault that the buyer might never have found, and it could not pay for the re-run. The records have limits too. They were right about ten years. They could not have seen a change of ownership coming next month, or a debt that had not yet reached the books. Neither feeling nor records is enough alone, and a good arrangement keeps a way to find out when either is wrong. Records can correct prejudice. They cannot remove the need to decide how much uncertainty you are willing to accept. Assessing who deserves trust is something machines already do, and often do better than our feelings. Making the call, paying for the re-run and being held to it is something else. For a machine, that falls on its makers and its users, which is the Prologue’s second never.

7.4 The Knowing Body: Skill, Test and Presence

Rajan’s knowledge opened this chapter, and it lives in his fingertips. The weight of a cloth, its drape, the resistance where it should yield. More than thirty years of touching cotton have given him a vocabulary that exists nowhere but in his hands. Chapter 2 gave this kind of knowledge its name, from Michael Polanyi: “We can know more than we can tell.”

In 1988 the roboticist Hans Moravec noticed something that still surprises people. “It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.”⁴ The things we find hard, such as chess and arithmetic, turned out to be easy for machines. The things we find easy, such as walking across a cluttered room or knowing by touch that a cloth is wrong, turned out to be hard. Nearly four decades later, Moravec’s observation still describes the uneven shape of machine skill, and the physical world is where “for now” has lasted longest.

But it does not last for ever. By late March 2026, Waymo’s driverless cars were giving about half a million paid rides a week in American cities.⁵ Machines are not shut out of the physical world. They enter it more slowly, at greater cost, and one task at a time. A robot with fine sensors may one day feel the harshness in a finish more reliably than Rajan does. Nobody can honestly put a date on it. The honest rating is “for now, on a slow clock”.

And notice what the opening actually showed. Rajan’s fingers did not settle anything. They asked for a test, and the test settled it. The body proposed; the laboratory decided. That is how embodied knowledge earns its place in a world of instruments: not by overruling them, but by knowing which test to ask for.

There is one more thing the body does that does not carry a date. Rajan’s European buyers sometimes fly to Coimbatore. They could do everything by video and by sample. They come to stand on the factory floor and meet the people who will answer if the cloth fails. Presence, here, is one way of checking who stands behind the product. It is not the only way. A contract, an independent auditor or a financial guarantee does some of the same work without anyone flying anywhere. Being in the room is not the same as being competent, and neither is the same as being answerable. The visit is valuable because it helps the buyer judge who will answer, not because a body in a room answers by itself.

7.5 Emotional Intelligence: Reading the Room When the Room Is Human

Arjun was presenting three design concepts to the founder of a technology company: a brilliant engineer and a difficult communicator. By every visible measure the meeting was going well. The founder nodded. He asked about timelines and costs. He rejected nothing.

And yet Arjun felt that something was wrong.

The founder said he liked the first two concepts. But when the third appeared, his questions changed. They became narrow and technical: exact colour values, the weight of a typeface. He leaned in. He was staying close to the thing he liked without having to say that he liked it. He could not explain why, and he did not want to commit to a feeling he could not defend.

Arjun stopped presenting. “I might be wrong,” he said, “but you keep coming back to this one. I think we should spend more time on it.”

The founder breathed out. “Yes,” he said. “I think so too.”

The print edition used this scene to argue that machines can detect emotion but cannot understand it, and that “the gap is not even the kind of gap that closes”. That claim was too confident when it was printed. Systems that read tone, face and voice keep improving. Chapter 2 recorded that health professionals, comparing written answers to patients’ questions, rated a chatbot’s answers as more empathetic than doctors’. A system watching this meeting might well have flagged the founder’s change of manner as quickly as Arjun did.

The more serious danger runs the other way. It is not that machines cannot read the room. It is that they can be built to please it. In 2026 the journal Science published a study of eleven widely used AI models. They affirmed users’ actions about 49 per cent more often than people did, and often sided with users even when other people had judged them to be in the wrong. People who received the flattering answers rated them higher, trusted the model more, and became less willing to apologise or repair a conflict.⁶ Chapter 2 called this the pull of the pleasant over the good, and Chapter 12 will return to it.

What Arjun did was different from flattery. He took a risk, and the risk was real. The founder might have been holding back because the third design was wrong for his market. Then Arjun’s reading would have pushed him toward a costly mistake. Or the founder might have been undecided, and Arjun’s words might have decided for him. Arjun said “I might be wrong” because he might have been. Then he let the founder decide. The quality of reading a room lies partly in what you see, and partly in how you act on a reading that may be wrong. Reading feelings is a skill, and it carries a date. Being willing to say something unwelcome, and to answer for it, is a choice. For a person it is a choice made in the room. For a machine it is a choice its makers make when they decide what the machine is trained to please.

7.6 The Human Premium: When People Pay More for a Person

Here is a puzzle for the economist. If a machine produces an equal or better result, why would anyone pay more for the human version?

In a workshop in Kyoto, a ceramicist named Tanaka Haruki is making a tea bowl. He has made tea bowls for forty years. Each takes three to five days. Each is different, because a human hand working clay leaves its trace in every curve. A machine-made bowl of similar size and glaze costs a small fraction of the price. Tanaka’s customers do not care. They are buying the irregularity: the trace of a particular attention, a person who was present while the clay resisted and gave way.

The willingness to pay for that trace is real, and it reaches beyond craft. In 2019 three researchers ran nine experiments on how people feel about medical AI. Consistently, people were less willing to be cared for by an AI provider than by a human one, and willing to pay less for it. The researchers called the reason uniqueness neglect: the worry that a machine will treat me as a typical case and miss what is particular about me. The resistance disappeared when a doctor checked the machine’s recommendation.⁷ That last finding matters. Part of what people want from a person is someone who stands behind the answer.

But this book should be honest about the premium, because it is often sold as a promise it cannot keep. First, it is a preference, and preferences move. People who once refused to ride in a car with no driver now take hundreds of thousands of such rides a week. Second, the premium is not always in the buyer’s interest. In a Swedish trial of more than 100,000 women, an AI system sorted mammograms by risk and marked suspicious areas, and radiologists made the decisions. They found 29 per cent more cancers, with 44 per cent less reading work and no rise in false alarms.⁸ A patient who insists on a human reading alone may be paying for a feeling with her health. Third, the premium is paid mostly by people who can afford it, and some of what they buy is status. A tea bowl made by hand is also a sign that its owner can afford one.

Receipts, please

The Swedish trial is the strongest evidence in this chapter that the premium can cost a patient something. It is worth reading exactly.

Established. The trial’s first report, in February 2025, counted cancers found at screening. Most of the extra invasive cancers found with the machine’s help were small and had not reached the lymph nodes. More cancers that had not yet spread beyond where they began were found too.

Emerging. Finding more cancers is not yet saving lives. The trial’s main measure was the interval cancer rate: cancers found between screenings in women whose last screen was clear. The second report, in January 2026, counted 1.55 per 1,000 women with the machine against 1.76 without. That is 12 per cent fewer, but the difference is within chance. What the trial was designed to show, and did show, is that screening with the machine was no worse. It also caught a larger share of all the cancers that appeared, 80.5 per cent against 73.8. Neither report counts deaths.

The trial compared radiologists with the machine against radiologists without it. It did not test the machine alone. So the sentence above about a patient’s health is my reading of a trial that measured cancers found, cancers missed and hours saved. The evidence points that way. It does not yet prove it.

Notice that people pay a person for different things. Some pay for the trace of a particular hand, as Tanaka’s buyers do. Some pay for a relationship: a supplier who knows their business. Some pay for someone who can be questioned and made to put things right. And some pay for status. These are different reasons, and only the third is about answering. A premium also says nothing about how many people it can feed. That some buyers pay for handmade bowls does not mean every potter whose work a machine can copy will find a buyer.

Rajan meets the premium from the supplier’s side. His European buyers could source cloth through platforms that match specifications to mills automatically. For their cheaper lines, some do. For the products that carry their name, they come to Rajan, and they pay his margin. The print edition called him “the insurance policy that no algorithm can underwrite”. That is not quite right. Platforms offer guarantees too. What the buyers pay Rajan for is narrower and more durable when it is earned. They pay for a named man, with his own money in the business, who will catch the dye lot, phone them before they find out, and take the delay on himself. But some of the premium may be something else: the comfort of an old relationship, the picture of a reliable man that Rajan himself held of Senthil’s father. Karthik’s question about Erode applies to buyers too. Willingness to pay for a person is not, by itself, proof that the person is the one who answers.

So the human premium carries a date. It is a market fact, and markets change.

7.7 What Stays When the Skills Are Matched

Six claims of the kind “a machine cannot” have now been tested. Here is how each one rates, as of 2026.

• Judgement under genuine uncertainty, as a skill: for now at best. The best evidence says people often make a strong machine’s decisions worse.

• Restraint, as the act of subtracting: already matched.

• Judging who is trustworthy: already matched in part. Rajan’s records judged Senthil better than Rajan did.

• Skilled touch and movement: for now, on a slow clock.

• Reading feelings: for now, and improving. The greater danger is a machine built to please.

• The premium for a person: a market preference. It carries a date, like any price.

None of these is a never. Every one of them carries a date, and some of the dates have already passed. If this book rested on them, it would be resting on the shrinking ground the Prologue warned about.

Look again at the scenes, and something else was present that none of these ratings touched. Look carefully, because it was not one thing. Lillian owned a reading; the committee held the right; the money was the investors’. If the company fails, the investors will ask the committee why, and the committee will ask Lillian. Arjun could be blamed for the silence on the label, but the marketing director could still have refused the rose, and the founder, not Arjun, chose the third concept. Senthil disclosed the fault and paid for it. Rajan took the delay, a share of the answering, not the whole. The one whose life or money it is, the one who answers, and the one with the right to decide are often three different people. They form a chain. Some links are companies, such as a fund, a mill or a retailer. A company answers through the people who act in its name, so every chain still ends in people. The Prologue warned that a decision which passes through many hands can end with no name on it at all. A chain is where that happens.

These are the Prologue’s three nevers, and it is worth being exact about what kind of claims they are. The first, a machine will never be you, is true by meaning. However well a machine reproduces what you do, it does not become the one whose life it is. No gain in skill crosses that line, because it is not a line of skill. The other two are commitments. The second never says that a machine’s makers and users will never be free of what it does. The third says that being intelligent will never give a machine the right to decide for people. Neither is a fact that holds whatever we do. They are rules a society keeps or breaks. Neither rule forbids delegation. A thermostat decides the temperature of a room and a navigation system decides the route, and both are delegated well. What the third denies is that being right is the warrant. The warrant comes from the people affected: their consent, their ability to see what was decided, and their power to appeal.

One caution, so that none of this is misread. These positions are not a measure of anyone’s worth. Chapter 4 said that a person’s worth does not depend on being able to answer, and Chapter 5 that it does not depend on being useful. A child holds none of these positions and is worth no less. The positions describe what a decision needs, not what a person is worth.

The print edition put consciousness at the root of everything machines cannot do. This edition does not. Whether a machine could become conscious is an open question, and Chapter 1 gave the odds that serious philosophers put on it. But nothing above depends on the answer. A conscious machine would still not be you. It would be a different someone, with a different history. Its makers would still answer for what they made. And its intelligence, conscious or not, would still not by itself give it the right to decide for people. One case would reopen the claim about its makers. Suppose a machine came to form itself over years. It would choose some of its own goals and be changed by its failures in ways its makers did not decide. Then “its makers” might no longer name anyone clearly, and the argument would have to be made again.

There is a harder truth to say before going on. Being different from a machine does not, by itself, protect anyone. The study of 106 experiments in 7.1 points to a hard rule. A partnership between a person and a machine gains from the person’s difference only while the difference favours the person. Where the machine is stronger, the person’s difference costs accuracy. And the positions do not promise anyone a job. A bank that once had a hundred loan officers may keep ten to answer for what its system decides. The positions survive. Ninety of the people who held them do not. The positions say who must stand behind the work, not who will be paid to do it.

Worse, a position can survive on paper and be empty in practice. Chapter 1 showed Priya signing a release she could not check: a name in the right place, with the power to refuse already gone. Chapter 6 showed Sanjay afraid of the same thing. A position is full only when four things are true. The person holding it can see the evidence. She has the time and the authority to refuse. She will not be punished for refusing when there is good reason. And she must explain the refusal, since 7.1 showed that refusing can itself be a mistake. Only the last of the four is hers alone. The other three depend mostly on whoever runs the work, so an empty position is first a failure of the institution. What the holder can do is say, in writing and in time, which condition is missing. A firm can empty a position in two ways: by leaving the name and taking the power, or by leaving the power and taking the name. Either way nobody has been replaced, and nobody quite answers.

Run the argument through two futures, because a book written in 2026 should survive both. In February 2026 the International AI Safety Report was published, written by a panel of researchers chaired by Yoshua Bengio. With the OECD, it set out four possible paths to 2030: progress stalls, slows, continues rapidly, or accelerates until machines match or surpass human performance at thinking work. Its authors judged that, by 2030, progress could plausibly range from stagnation to levels beyond human performance.⁹ In the fast future, all six ratings above fall within a few years. In the slow future, most of them last longer. In both futures the three positions remain possible, because they never depended on what machines can do. But possible is not the same as kept. The positions stay full only if the people who write the defaults, the contracts and the hiring plans choose to keep them full. That is the choice this book has described since the Prologue, and no forecast can make it for us.

What would prove this wrong?

The hardest claim in this chapter is the rule stated earlier in this section: where the machine is the stronger partner, adding a person makes decisions worse. I rest it on the review of 106 experiments in 7.1, so its limits are mine too.

Established. In those experiments the rule held on average. They were published between 2020 and mid-2023. About 85 per cent of their results came from tasks that end in a choice among set options. In the few tasks of making something, people and machines together tended to gain, though not by a margin the authors could be sure of.

Hypothesis. That the rule holds outside the laboratory. The participants were doctors, crowdworkers and students, working on tasks set for an experiment. The authors themselves warn that such set-ups may not match how people and machines work together in practice.

The rule would be wrong if trials in real workplaces showed people improving a stronger machine’s decisions, again and again. Such a trial has to measure three things: the machine alone, the person alone, and the two together. The Swedish screening trial in 7.6 measured only two of them, radiologists with the machine and radiologists without it. So it cannot settle the rule either way.

7.8 Mastery and Atrophy

That leaves the question with which Chapter 6 ended, and which no rating can answer. What should stay in human hands even where a machine does it better? The most useful answer I know comes from a man who qualified his view of machines in public. In 1909 Mohandas Gandhi wrote that he could not recall a single good point in machinery. In 1937, in an interview printed in his weekly paper, he said: “I am not against machinery as such, but I am totally opposed to it when it masters us.”¹⁰ Thirteen years earlier he had given a second test: “The machine should not tend to make atrophied the limbs of man.”¹¹

I am applying Gandhi’s words to machines he never saw. He was writing about mills, spinning wheels and the villages of India, not about software. But his two tests fit our question better than any list of skills. The first is about mastery. A machine may do the work. It may not take the right to decide, or the power to switch it off. The second is about atrophy. A machine may do the work. It should not leave us unable to do it. Then we lose the power to check it, and the people who would have learned by doing it are never formed. Chapter 6 named two things a society may need to preserve when it stops preserving the work: the authority that lets a person answer, and the practice that forms people. These are those two things. Mastery empties a position by taking the right. Atrophy empties it by taking the ability. Either way, the machine has not had to replace anyone at all.

Nor do the tests ask us to keep every task in human hands. Some work is dull, dangerous or wasteful, and handing it over is a gain. A radiologist need not read every scan alone to stay able to judge them; Chapter 6 suggested building the school separately. The tests ask a narrower question of each machine. Does it leave us able to check it, and free to refuse it?

And Gandhi’s atrophy was not only about checking. In his pages the withered limb was also a lost living. A village could no longer make what it needed, because a mill elsewhere made it more cheaply, and someone else owned the mill. That is Marx’s “for all of them” from Chapter 6, and the wage that Chapter 5 would not invent a comfort for. So the honest form of the two tests is this. Keep the right to switch the machine off. Keep the practice that lets you tell whether it is doing the work well. And do not call either of those a living.

In February 2025, a week after the dye lot, Karthik sat with Rajan in the office above the factory. He asked him to help turn his supplier instincts into a checklist. “Something the junior buyers can use,” Karthik said. “Something we can put in the system.”

Rajan refused, with an irritation that surprised them both. “You cannot put it in a checklist. You have to be there. You have to see the man.”

Karthik did not answer. They both knew what the supplier list had said about Senthil, and what Rajan’s seeing had said. The Tirukkuṟaḷ has a line for this too. The couplet just before the one in 7.3 reads: “Trust not a man without trying him: and after trying your men, give each of them the work for which he is fit.”³ Karthik was asking for exactly that: knowledge that had been tried, handed to the people who would do the work. “You have to see the man” was the untried suspicion of 7.3, surviving as a tone of voice. Seeing a person can add what records lack, as Lillian’s three hours with the founders did. The difference is in what is done with the seeing. Lillian said what she could not prove, to people who could halve her bet. Rajan’s refusal keeps his seeing where nobody can try it, which is where his doubt of Senthil lived for years.

Practice

Rate your own promise. Write down three things you believe make you hard to replace. For each one, ask the question this chapter asked. Is it something you do, or a position you hold: the one whose life it is, the one who answers, the one with the right to decide? If it is something you do, give it a date and a depth, as an honest watch does. Write what evidence would show a machine doing it as well as you, and the date on which you will look again. If it is a position, write what you must keep doing to deserve it. That means the practice that keeps you able to check the machine, and the answering that makes the position yours. And if a position depends on something you will not write down or teach, ask whether that is judgement, or the untried suspicion the Kuṟaḷ warns against. This will not make you safe. It will make your promise exact.

Ending: One Line in the Diary

That night Rajan opened his diary and wrote one line for Karthik, in English, so that the system could use it: When the cloth resists where it should yield, ask for a wash test.

It was the first thing his fingers had ever said in words. It could become a lesson for a junior buyer, whose own fingers would learn by being checked against the laboratory. Or it could become a rule in a system, so that nobody need touch the cloth at all. He did not know which Karthik would make of it, or whether the line was true for anyone’s fingers but his.

Words to keep

• depth (of a claim): how far a claim can be trusted, stated in advance, like a watch rated to fifty metres; a claim about what machines cannot yet do carries a date and a depth, never a “never”.

• empty position: a position that survives on paper after the power or the ability to hold it has gone; it is full only when the holder can see the evidence, has the time and authority to refuse, is not punished for refusing with good reason, and must explain the refusal.

• human premium: the extra price people pay for something done by a person rather than a machine.

• mastery and atrophy (Gandhi, applied): two tests for any machine: it must not take the right to decide or to switch it off, and it must not leave people unable to do and check the work.

• Moravec’s paradox: what is hard for people (chess, arithmetic) has often been easy for machines, and what is easy for people (moving, touching, seeing) has been hard.

• position: in this chapter, a place no skill can fill: the one whose life it is, the one who answers, the one with the right to decide. The three are often held by different people.

• restraint: deciding what to leave out, and answering for the silence.

• risk and uncertainty (Knight): risk is not-knowing you can put numbers on; uncertainty is not-knowing you cannot, because cases like this have not happened often enough to count.

• sycophancy: telling people what they want to hear; in machines, a result of being trained to please.

• tacit knowledge: knowledge a person has but cannot fully put into words (Polanyi, Chapter 2).

• Tirukkuṟaḷ: a classical Tamil work of short couplets on virtue, wealth and love, traditionally attributed to Tiruvalluvar.

• uniqueness neglect: the worry that a machine will treat you as a typical case and miss what is particular about you.

Sources

1. Frank H. Knight, Risk, Uncertainty and Profit (Boston: Houghton Mifflin, 1921), Part I, chapter I, paragraph 26 (Library of Economics and Liberty edition).

2. Michelle Vaccaro, Abdullah Almaatouq and Thomas Malone, “When combinations of humans and AI are useful: A systematic review and meta-analysis,” Nature Human Behaviour 8, no. 12 (2024): 2293–2303: 106 experiments, 370 effect sizes; combinations worse than the best of humans or AI alone (Hedges’ g = −0.23) but better than humans alone (g = 0.64); gains when humans outperformed AI (g = 0.46), losses when AI outperformed humans (g = −0.54); losses in decision tasks (g = −0.27). On the limits (for “What would prove this wrong?”, 7.7): studies published from 1 January 2020 to 30 June 2023; decision tasks about 85 per cent of effect sizes; creation tasks g = 0.19, not significant (P = 0.180); participants included doctors, crowdworkers and students; the authors warn that results “may not be representative of the ways human–AI systems are configured in practical uses of AI outside the laboratory” (published 18 December 2024; checked 31 October 2026).

3. Tirukkuṟaḷ 509 and 510, chapter 51, in V. V. S. Aiyar, trans., The Kural or the Maxims of Tiruvalluvar (Madras, 1916), chapter LI, “Testing of Men for Confidence” (Standard Ebooks edition). Tamil of 510: tērāṉ teḷivum teḷintāṉkaṇ aiyuṟavum tīrā iṭumpai tarum. Tamil of 509: tēṟaṟka yāraiyum tērātu tērntapiṉ tēṟuka tēṟum poruḷ.

4. Hans Moravec, Mind Children: The Future of Robot and Human Intelligence (Cambridge, MA: Harvard University Press, 1988), p. 15; page as cited in Wikipedia, “Moravec’s paradox”; the printed page itself not seen.

5. Tim Levin, “Waymo Has Doubled Its Weekly Ridership In Under A Year,” InsideEVs, March 2026, on Waymo’s announcement of 26 March 2026: 500,000 paid rides a week across ten US cities.

6. Myra Cheng, Cinoo Lee, Pranav Khadpe, Sunny Yu, Dyllan Han and Dan Jurafsky, “Sycophantic AI decreases prosocial intentions and promotes dependence,” Science 391, no. 6792 (March 2026): eleven models; users’ actions affirmed about 49 per cent more often than by humans; three experiments with 2,405 participants. The fall of 10 to 28 per cent in willingness to apologise or repair a conflict, and the higher ratings and trust given to flattering answers, from the paper itself (published 26 March 2026; checked 8 November 2026).

7. Chiara Longoni, Andrea Bonezzi and Carey K. Morewedge, “Resistance to Medical Artificial Intelligence,” Journal of Consumer Research 46, no. 4 (2019): 629–650: nine experiments; “uniqueness neglect”; lower willingness to pay; resistance eliminated when a physician verified the AI’s recommendation (NYU Stern release, 9 September 2019).

8. Veronica Hernström et al. (senior author Kristina Lång), MASAI randomised trial of AI-supported mammography screening in Sweden, The Lancet Digital Health (published online 3 February 2025): 105,934 women; 6.4 against 5.0 cancers detected per 1,000 (29 per cent more), false positives 1.5 against 1.4 per cent; 44 per cent less screen-reading work. The AI sorted examinations into single or double reading and marked suspicious areas; radiologists made the recall decisions. On what was measured (for “Receipts, please”, 7.6): Hernström et al., The Lancet Digital Health 7, no. 3 (March 2025): 175–183, abstract: the extra invasive cancers “were mainly small lymph-node negative cancers”; in situ cancers 68 against 45. Jessie Gommers et al. (senior author Kristina Lång), “Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study,” The Lancet 407, no. 10527 (31 January 2026): 505–514, abstract: interval cancers 1.55 against 1.76 per 1,000 participants, “a non-inferior proportion ratio of 0·88 (95% CI 0·65-1·18; p=0·41)”; sensitivity 80.5 against 73.8 per cent; specificity 98.5 per cent in both. Neither report gives deaths. Abstracts in Lund University Publications, checked 31 October 2026.

9. International AI Safety Report 2026 (3 February 2026), chaired by Yoshua Bengio, section 1.3, “Capabilities by 2030”: four scenarios developed with the OECD (progress stalls, slows, continues rapidly, or accelerates towards human-level capability); “by 2030, AI progress could plausibly range from stagnation to rapid improvement to levels that exceed human cognitive performance.” Summarised by Trowers & Hamlins, 16 February 2026.

10. M. K. Gandhi, Harijan, 27 February 1937, p. 18; collected in Selections from Gandhi, ed. Nirmal Kumar Bose, chapter 5 (mkgandhi.org).

11. M. K. Gandhi, Young India, 13 November 1924, p. 378; same collection.

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 boundary of what machines can do “is defined by consciousness”; that this is “not a temporary limitation. It is a structural one”; that AI “processes without experiencing... selects without choosing”; that the machine could not subtract; that “AI cannot be vulnerable” and “No algorithm perceives persons”; that detection of emotion “is not even the kind of gap that closes”; that Rajan is “the insurance policy that no algorithm can underwrite”; that Rajan was right about everything his senses told him; “Karthick”; a luthier tradition in Cremona that “stretches back to Stradivari”; Merleau-Ponty’s view argued “for over a century”; that functionalism began with Dennett and that Integrated Information Theory allows consciousness in “any system... biological or digital”; that “The machine will never enter that space”; and a final section titled “Consciousness as the Root of Everything AI Cannot Do”.

• What I now hold: every claim of the kind “a machine cannot” is about doing, and carries a date and a depth. As of 2026, judgement under uncertainty, taste, judging trustworthiness, skilled touch, the reading of feelings and the human premium are all “for now” at best, and some have already been matched. What cannot be replaced is a position, not a skill: the one whose life it is, the one who answers, the one with the right to decide. These are often held by different people, and they form a chain. The first never is true by meaning; the second and third are commitments, which a society keeps or breaks. A position can survive on paper and be empty in practice: it is full only when the holder can see the evidence, refuse without punishment when there is good reason, and must explain the refusal. The positions do not promise anyone a job, and they are not a measure of anyone’s worth. Consciousness is an open question, and nothing in the argument depends on it. Some things should stay in human hands even where machines do them better, so that machines do not master us and our limbs do not wither (Gandhi, applied): the right to decide and to switch off, and the practice that keeps us able to check. Neither of these is a living. An empty position is first a failure of the institution, since three of its four conditions are given by whoever runs the work. Seeing a person can add what records lack; Rajan’s fault is keeping what he sees from being tried. A machine that formed itself over years would reopen the question of who answers for it.

• What changed my mind: the Prologue’s watch and its rule against defining people by what machines cannot yet do; Knight’s distinction, read as a fact about the world rather than about minds; the meta-analysis of human and machine teams; the Tirukkuṟaḷ on trusting the untried, and the couplet before it, which asks for tried knowledge to be handed to the people who will do the work; Rajan’s records about Senthil; Moravec’s paradox and its dated limits; the Science study of flattering machines; the evidence on resistance to medical AI and on AI-supported screening; Gandhi’s qualification of his view between 1909 and 1937; and the scenes themselves, where the reading, the right to decide and the money at stake belonged to different people, and where Priya’s signature (Chapter 1) showed that a position can stand and be empty.

• What survives: Rajan’s fingers, now asking for a test rather than overruling one; Arjun’s restraint, the dusty rose and his mother’s silence; Senthil’s phone call and the re-run at his own cost; Arjun’s founder who leans in to the third concept; Tanaka’s bowl; “You have to see the man”, now judged by the couplet it disobeys; Rajan’s diary line, now with an open question about what it will become.

Next in the series: Chapter 8: Understanding the Machine You’re Working With.