NP Nithin Palal

Chapter 37: The Amplified Human

Part VIII: The 2035 Human

· 47 min read · Discuss · Plain text

AI-Proof: What Machines Can Never Be — Volume III: Transformation, Part VIII: The 2035 Human, Chapter 37: The Amplified Human

Part VIII asks what people are becoming as machines do more of what they do. This first chapter of Part VIII follows a radiologist, an operations head and her eight-year-old niece, and asks what is still ours to decide.

A radiologist keeps his seat; the post beside it goes unfilled. A child gets replies before her questions form. An executive signs reviews unread. A machine enlarges what we bring, lapses included. What we become is partly ours to decide.

In Brief For readers who wonder what machines will make of their work, their children and the decisions they hand over. Chapter 37 opens Part VIII by asking what people are becoming as machines do more of what they do. A radiologist in Mumbai keeps his seat while a retiring colleague’s post goes unadvertised; an operations head in Hyderabad finds she has signed reviews she never read; her eight-year-old niece asks a phone whether knowing feels like something. Rousseau, who wrote in one year about raising a child and about how free people bind themselves, and an Upaniṣad’s father who made his son break a fruit and taste salt water, help separate what is a forecast from what is still ours to decide.

A machine enlarges whatever we bring to it, our lapses as well as our attention; what we become is decided less by what machines will be able to do than by the habits we keep, the children we form and the agreements we make about who answers.

Opening: A Leaf at the Window

Diya was eight. She stood at the window of her aunt’s flat in Banjara Hills, in Hyderabad, holding a leaf from the hibiscus on the balcony. It was the start of October 2026, a long weekend, and her father had brought her from Chennai to see his mother. They were staying with his sister, Meera Reddy. Meera is vice-president of operations at the Hyderabad site of NovaChem, a drug company with its head office in Basel, and she has worked for it for fifteen years.

Diya had asked her father’s phone how leaves make their food, and it had told her. Sunlight, a green pigment called chlorophyll, carbon dioxide from the air and water from the roots make sugar for the plant and oxygen for everyone else. All of it was right. Her father read her next question into the phone, smiling: how does the leaf know to do this? The reply was careful, more careful than most adults would have managed. A leaf does not know anything in the way she knows things, it said. It has parts that respond to light, and the plant works by such responses without thinking about them.

Diya considered this. “The AI knows the answer,” she said. “But does the leaf know? Does knowing feel like something?”

Her father laughed and began to read that question to the phone as well. Thirty seconds later Diya asked whether it could help with her maths homework. She was bored with photosynthesis. She wanted to go to lunch.

Meera said nothing through any of it. A fortnight earlier, in Basel, she had told the company’s global chief executive, a chemist named Friedrich, that she had signed reviews she had not read. Now she watched her niece ask three questions in a minute. Each reply had come before Diya had finished wondering.

This chapter opens Part VIII, and the Part has one question. When machines can do more each year of what people do, what are people becoming, and how much of that is ours to decide?

The Part’s title, “The 2035 Human”, needs one correction. A title with a year in it is a forecast, and a forecast is something a person does on a date, to be scored later. Nobody can score this one yet. In February 2026 the International AI Safety Report, written by researchers led by its chair, Yoshua Bengio, judged that by 2030 alone machine progress “could plausibly range from stagnation to rapid improvement to levels that exceed human cognitive performance.”1 No honest portrait of the human of 2035 can be drawn from a range that wide, and this Part does not try. It can notice something plainer. The human of 2035 is mostly alive now. Diya will be seventeen. What she becomes is not waiting for her in the future. Some of it was being decided on a Saturday in her aunt’s flat.

Some things carry no date. This book says “never” about three only. “A machine will never be you.” “A machine’s makers and users will never be free of what it does.” “Being intelligent will never give a machine the right to decide for people.” The first holds by what we mean by a life. The other two are commitments, which people keep or break. Everything else in this Part carries a year, and any claim about 2035 is a projection, marked as mine or someone else’s, with what would show it wrong. A forecast can be scored, a scenario explored and a commitment kept or broken, and this Part tries to say which each claim is. In that sense only, the title stands: as a question about people already here. What machines amplify carries a date. The nevers do not.

This chapter asks: If machines enlarge whatever we bring to them, lapses as well as attention, what do we become, and what is still ours to decide?

The print edition answered with a portrait. The human who emerges from the age of artificial intelligence would be “not diminished, not replaced, but amplified”, forced “to discover capabilities that were always present but never required”. She was already appearing, it said, in the professional whose judgement is trusted, in the child who has never known a world without AI, and “in the organisation that deploys machines for processing and humans for meaning”.

The chapter keeps its title, and the title needs an argument. An amplifier is a device that takes a small signal and delivers a larger one with the same shape.2 It does not choose what it enlarges. Even a perfect amplifier enlarges the noise in its input as faithfully as the signal, and a real one adds noise of its own. So “the amplified human” is not a new kind of person. It names a relation, and the relation runs both ways. The machines in this chapter make larger whatever is brought to them: a child’s question, a doctor’s reading, a vice-president’s signature. In Meera’s year the system had multiplied her signature across about thirty reviews. It had done nothing for her attention. Yet in practice the amplifier is not passive. A reply that always comes first changes what a person brings next.

Whether what grows is the attention or its lapse depends on what a person brings, and on what the people around her have arranged. That is what this chapter holds. What we become is decided less by what machines will be able to do than by the habits we keep, the children we form and the agreements we make about who answers.

Amplification is not shared out evenly, and it does not run one way. A machine can enlarge a person’s output while shrinking her bargaining power, her practice or her place. It can enlarge one party at another’s cost: a hospital’s reading at the expense of a trainee’s post, an employer’s margin at the expense of a worker’s wage. A person can be amplified and weakened in the same year. The measure is not how much more she can do with a machine. It is whether she becomes more able to understand, decide, act and answer, or more dependent on a system that does those things in her place.

In each of this chapter’s stories the same thing is at stake: practice. A radiologist in Mumbai, Sanjay Rao, keeps his by reading ten images on his own time. The post his hospital left empty beside him removes a beginner’s. Diya’s is being formed, one reply at a time. Meera’s lapsed while her name went on. The better a machine does a task, the less often people tend to do it, and the harder it becomes to tell when the machine is wrong.

And “ours” does not mean each person alone. Meera could choose to read, but others had set much of her week. Sanjay can practise, but the hospital decides its posts. Diya’s father could have chosen a study mode, but the company chose the default. A choice a person cannot afford to make is not much of a choice, so much of what is still ours is decided together.

37.1 The Judgement Economy and Its Few Judges

Nothing about the judgement economy is in full bloom in 2026, and for most of the people it is promised to, it may never flower.

Sanjay Rao is a radiologist at a private hospital in Andheri, in Mumbai. He chose radiology a quarter of a century ago because it let him help patients without sitting with them: images, not faces. In February 2025, late in a long evening shift, he missed a small nodule half hidden behind a rib on a chest X-ray. The machine the hospital had just installed found it, and he wrote a correction under his own name. Since about June 2025 he has read ten images himself each morning, on his own time, before he looks at the machine’s flags. He does it so that his eye stays able to check the machine. The hospital gains from that practice. It does not pay for it. More of his day is now spent with other doctors and with patients, explaining what the machine flagged and what it means for one person. He is not good at that part. It is the part he chose radiology to avoid. His pay has not changed. His list has grown.

The print edition told his story as a promotion. The machine has the sharpest eye, it said, so the radiologist who now “commands the highest referral rates” is “the one whom other physicians call when the machine’s output does not match the patient’s story”. Then it drew a rule: “Judgement compounds. Research depreciates.” There is something true in it. When reading gets cheap, what grows more useful is deciding what the reading is for, and standing behind the decision. But useful is not the same as well paid, and being needed is not the same as being hired. What a thing costs, what it sells for, what it is good for and what a person is paid to do it are four different things. They can move in different directions.

The harder question is how many judges such an economy needs.

Sanjay’s hospital gave its reply in 2025. Three of its radiologists would retire in the next few years, and their posts would not be filled. The machine and the remaining doctors would absorb the work. In the spring of 2026 the first of the three retired, and the post was not advertised. Nobody lost a job. A young doctor somewhere did not get one. The work did not shrink. Through the summer the remaining radiologists each signed more reports, with the machine reading first. The doctor who would have filled the post might have learned beside them, reading a case and then checking it against theirs and the machine’s. Nobody has built that school. The hospital keeps today’s experts by closing a post through which tomorrow’s would have formed. That was the hospital’s choice, not the machine’s. How many judges there are is set by hospitals, payers and the bodies that train doctors.

This call goes to the machine, and it should be said plainly. 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 almost no change in false alarms.3 More women’s cancers were found, and with fewer hours of radiologists’ time. Finding more cancers helps only if they would have mattered, and screening can find some that never would. The trial’s fuller results, published in January 2026, help with part of that. In the AI group, fewer women were found to have cancer between one screening and the next: 1.55 against 1.76 per 1,000, a gap small enough to be chance. Screening with AI also caught a larger share of all the cancers found, at screening or before the next one: 80.5 against 73.8 per cent.3 As I read it, some of the extra cancers would otherwise have been found later. That does not rule out that some would never have done harm. The trial did not measure deaths, so it cannot yet say whether finding more cancers saved lives. The judgement stayed with people. The radiologists’ hours needed to hold it fell. The trial counted hours, not posts, and the step from one to the other is mine. Sanjay’s hospital had already taken it. An economy of judges may need fewer judges.

Under the hood

The trial’s saving came from one step. In the Swedish programme where the trial ran, two radiologists normally read each screening mammogram. That is what the control group received. In the AI group, the system gave each examination a risk score from 1 to 10. Examinations scored 1 to 9 went to one radiologist, and only those scored 10 went to two. For scores of 8 to 10, the readers also saw marks where the system had found something suspicious.

So the machine did not take the first reader’s seat. For most women it took the second reader’s. The control group needed 109,692 readings, and the AI group 61,248. Every decision to call a woman back was still a radiologist’s. What went, for most women, was a second radiologist checking the first. The reading saved was that second look. Whether the machine’s sorting is as good a check is what the trial’s later results, above, began to test.

The same field also shows how badly such forecasts can go. In 2016 one of the founders of modern AI, Geoffrey Hinton, said: “People should stop training radiologists now.” Over the following decade the number of radiologists in the United States rose by about a tenth, and training places reached a record. Reading images turned out to be only about a third of a radiologist’s time.4 Cheaper reading can also bring more scans. Where more scans are waiting than there are radiologists to read them, the saved hours go to reading more, and the posts stay. A forecast is something done on a date, and it can be scored. That one scored badly. What failed was the forecast that the job would vanish. That is not the same as no effect on the job. Cheaper work can go four ways. Fewer people can do the same work, or the same people more of it. They can do different work, or the work can gather in fewer hands. Which follows depends on prices, payers, screening rules and who owns the machines.

The wider evidence points both ways. In 2024 the economist Daron Acemoglu estimated AI’s effect on total factor productivity, the part of growth in output not explained by more labour and capital. Over ten years in the American economy, he put it at no more than about 0.66 per cent. He also warned that it could move income from workers to the owners of the machines. Other economists think his figures far too low.5 Among the young, one pattern is already visible. In August 2026 economists at Stanford reported on people aged twenty-two to twenty-five in the jobs most exposed to AI. Their employment stood about 19 per cent below where it would have been had it moved with that of similar workers in less exposed jobs. The gap came mostly from hiring that did not happen, and the authors call it a pattern, not a measured effect of AI.6 Some of them are the people from whom the next judges would have come. Fewer first jobs means fewer of the rungs on which judgement is learned.

I offer one projection of my own, and it is only that. By 2035, in work where a person signs for what a machine has read or drafted, I expect the number of people who sign to fall against the amount of work signed. Take American radiologists per million imaging studies, measured against 2025. My range runs from a fall of a tenth to a fall of a half, depending on which path the machines take, and on law and insurance. What would show me wrong is already in front of us. Suppose radiologists per million studies hold steady or rise while machines do more of the reading. Then the judges were not few after all.

So the judgement economy is real in one sense. What a person adds is less often what she knows and more often what she decides and will stand behind. It promises no wage, and it may offer few seats. The measure in the projection above is rough: studies differ in difficulty, and some reporting may pass to others. Its bounds are a judgement, nearer a tenth if machines stay assistants and payers hold, nearer a half if law and insurers let one signature cover far more studies. Few seats can also mean a new hierarchy, a small number who approve and override above many whose work is narrowed and watched. Elsewhere the same machines may spread judgement more widely. Which comes depends on who is trained, who is paid and who may challenge the judges. Sanjay still holds his seat. The seat beside his is empty, and the hospital chose that.

37.2 The New Social Contract, Among People, About Machines

If machines make some work less needed, people decide who gains and who bears the loss. The contract in this section is among people, about machines, not between humans and machines. The argument follows from that change of parties.

In 1762 Jean-Jacques Rousseau published two books. One, Emile, was about how to raise a child. The other, The Social Contract, was about how free people may bind themselves to one another. Both were condemned in Geneva, the city of his birth, and Emile in Paris as well.7 In G. D. H. Cole’s translation, the second book sets out its problem: “The problem is to find a form of association which will defend and protect with the whole common force the person and goods of each associate, and in which each, while uniting himself with all, may still obey himself alone, and remain as free as before.” His answer begins with what each party gives: “Each of us puts his person and all his power in common under the supreme direction of the general will”.8

Rousseau wrote about states, not machines, and I borrow only what his clause implies about who can be a party, which is my reading. His contract founds a state and its sovereign. It is not a deal among interests. A party must have something of its own to put in, and must be able to be bound and held to what it agreed. As of October 2026 no law treats a model as a party. A company can be a party, because the law gives it a standing of its own that outlasts any owner or employee. But it acts, decides and pays through people. Being a person in law needs no mind, as Chapter 36 found. Lacking that standing does not stop a machine’s acts having effects. In 2017 the European Parliament asked the European Commission to consider, in the long run, a status of “electronic persons” for the most advanced robots, “responsible for making good any damage they may cause”. In 2020 it said that legal personality for AI systems was not necessary.9 If a law ever did name a model, the money and the remedy would still come from the people who made and used it. The European Union’s new rules on liability will treat AI software as a product. Its makers will be liable for the death, injury and damaged private property that a defect in it causes, and for destroyed or corrupted data not used for work. They apply to products placed on the market after 8 December 2026.9 Much of the rest is fought out in contracts. A model’s maker writes terms that push risk onto the company that deploys it, and that company pushes it onward where it can. Under the new rules such terms cannot cut what is owed to an injured person. Between businesses, under national law, they still largely decide who repays whom. So a contract “between humans and machines” has the wrong parties. The contract is among people, about machines. Its parties are those who build the machines, those who buy and use them, those who work beside them, those they decide about, and the states that set the rules. It also binds people not yet in the room, such as the doctor who was not hired into the seat beside Sanjay’s. Suppose a machine one day came to form itself over years, with goals and a record of its own that its makers did not decide. Then this would have to be argued again. That is not the case in 2026.

This book’s second never says why the people who make machines cannot step out of the contract. In the Prologue’s words: “Children are also made by others, but a child grows into someone who answers for themselves, and the parents step back. This book holds that a machine’s makers and users never step back.” Raising a child is the one kind of making whose aim is that the makers step back. Making a machine is the kind its makers never step back from. The Prologue adds the limit: not being free of a machine does not mean being to blame for everything it does. How much falls to each person depends on her role, her control and what she could foresee. What the never rules out is pointing to the machine’s independence as a way out. Nor is the difference from a child that parents stop caring. It is that a grown child can answer for herself, and as of 2026 no machine is recognised as able to.

What bargain, then, should replace the old one? The print edition proposed one. “The human offers not knowledge but discernment,” it said, and the organisation offers “not job security but development security”: a promise to keep a person able, not employed. The idea is older than these machines. In 1989 the management scholar Rosabeth Moss Kanter took the word employability, until then used of people out of work, and applied it to employees.10 A company would promise to keep its people employable, though not to keep employing them.

Test the bargain against what a job used to bundle together. A job gave an income, a practice, standing among others, and a voice in how the work is done. This book has measured every loss by those four. “Development security” offers the practice, at most, and only at the employer’s choice. In return it asks the worker to stop expecting the income. It is the bargain an employer would choose. Where it is kept, it is a real good, because a practice is how people stay able to check the machines. But it is not a social contract. A contract that one side writes, and can change alone, is an offer. A fuller bargain would add what the offer leaves out. It would add income between jobs, paid training tied to real openings, notice, a voice in how machines are brought in, and a share of the gain.

Who keeps what share of the gain is not settled by the machines either. In 2022 the economist Erik Brynjolfsson argued that machines built to replace people weaken workers’ bargaining power. When machines are built to extend what people can do, he wrote, “humans retain the power to insist on a share of the value created.”11 Which kind gets built is decided by people: those who design machines, those who buy them and those who govern their use. The Swedish trial shows the limit of his distinction, as I read it. Its machine extended what radiologists could do, and the radiologists’ hours still fell. A machine that helps each person can still mean fewer people. The design of a machine does not settle who gains from it. Ownership, labour law, markets and bargaining power do. And when a machine raises output, where the gain goes is also a choice. It can go to wages, shorter hours, lower prices, training, or the owners’ margin. Nor is the gain one thing. One machine can give a patient a better diagnosis, a hospital a smaller wage bill, a radiologist a longer list and a trainee no post.

States have begun to write parts of the contract, unevenly. On 21 November 2025 India brought four labour codes into force. One of them, passed in 2020, defines gig work and platform work and promises such workers social security. The companies that run the platforms are to pay for part of it. Each is to give between 1 and 2 per cent of its annual turnover, capped at 5 per cent of what it pays such workers.12 The European Union moved the other way on one duty. Its law on AI had required the companies that make and use AI to “take measures to ensure, to their best extent” that their staff were literate in it. An amendment in July 2026 asks them only to “take measures to support” it, and says they need not guarantee any level.13 Neither law asks a machine for anything. Both are agreements among people about who pays to keep other people able.

What, then, do the parties owe one another? Earlier chapters of this book have answered piece by piece, and the pieces can be said plainly here. Those who make a machine owe an account of what it is, and repair when it fails. Those who use it on other people owe those people a way to ask why, and to be put right. An employer who changes the work owes an honest account of the change, and the hours its people need to stay able to check it. A state owes a floor under the people whose income goes, and rules that say who pays. None of these debts is owed by a machine, and none can be paid by one. Some are law in places, or soon will be, and courts can enforce them. The rest hold only where those who owe them keep them, or where workers, customers and voters can make them.

Rousseau’s problem has a second half, and it bears on machines. In his contract each party “may still obey himself alone”. A machine’s intelligence gives it no right to decide for people, however good its reasons. Knowing best is not the same as having the right. That does not forbid a machine to brake a car or stop a stolen card. People can give it such a task within limits they can justify and stand behind. What it lacks is a right of its own. The contract must be collective, or it will not hold. It is collective among people. The machines are its subject, not its parties, as of October 2026.

37.3 Raising Children in a World That Answers First

The contract is not only about jobs and liability. It is also about the people being formed now. Every child is conscious. The print edition feared that a child raised with machines “becomes less conscious”, as if consciousness came in amounts. This book does not use the word that way, and the worry it pointed at can be named more exactly.

Diya’s questions were good, and so were the phone’s replies. This call goes to the machine: the replies were correct, and the second was careful. Her last question is one that nobody has settled. The print edition called it “unanswerable by any machine”, but philosophers still argue about it too, and a reply can set out their arguments as well as a book can. Her question slides between kinds of knowing: holding a fact, being able to do something, and having an experience that feels like something. The phone gave her the first. The leaf does the second, as far as anyone can tell, without the others. Whether the machine itself knows anything in her sense is an open question, and nothing in this chapter depends on it. The danger is not that replies exist. Children have always had adults and books that told them things. The danger is that a reply can now arrive every time, before the question has finished forming, at no cost to the child, in a voice that does not hesitate.

The evidence on what this does to learning is young, and it points one way. In a trial with about a thousand secondary-school students in Türkiye, some were given an ordinary chat assistant of the kind on the market, and many used it to get answers. They did better in practice and worse in the exam taken without it. Those given a version that offered only hints did much better in practice and no worse in the exam.14 The students were teenagers learning mathematics, and the hint version was a tutor the researchers had set up for the trial. What it shows about an eight-year-old asking about leaves is a possibility, not a finding. The machine alone was not the harm. How it was set, and what the students used it for, mattered as much. The authors found that without safeguards students tried to use it as a “crutch” in practice, and then did worse on their own. Psychologists call the useful kind of struggle a “desirable difficulty”: effort that slows practice but makes learning last.14 Sanjay’s ten images each morning are the same idea in a grown professional, struggle kept on purpose so that the eye can still check the machine. Since April 2025 some assistants have offered a mode built for this; one company’s, launched in July 2025, is called study mode. It asks the learner questions to test her understanding, and in some cases refuses to give a direct answer until she has worked at the problem.15 Diya’s father’s phone was not set to it. Nobody had decided against it. The default was the company’s, and keeping it was his.

Much of what reaches Diya is decided by others. The company whose assistant she was using says its users must be at least thirteen, and under eighteen only with a parent’s permission.16 She was on her father’s account, under rules written for him. Since May 2025 another company has let children under thirteen use its assistant on accounts their parents supervise, and its help page asks parents to help a child understand “that it’s an AI tool, not a person to confide in”.17 From 13 May 2027 India’s data law will require a parent’s verifiable consent before a company processes the personal data of anyone under eighteen. As of October 2026 that duty is not yet in force.18 Schools that follow India’s central board begin teaching AI and computational thinking from Class 3 in the 2026–27 session.19 Learning to use an assistant is not learning to do without one. A child who can solve a sum with the machine but not alone has a tool, not yet the skill. Companies, the state and schools decide most of what reaches an eight-year-old. Her father decides the rest, between lunch and homework. None of them decided that Diya would get replies before she had finished wondering. Each decided a piece, and the pieces added up.

Two old voices disagree about what the adults should do. Rousseau’s Emile follows an imagined boy raised by a tutor. In Barbara Foxley’s translation, the tutor’s rule for the boy’s learning is plain: “Let him know nothing because you have told him, but because he has learnt it for himself.” Rousseau gave his reason. If ever authority takes the place of reason, the boy “will cease to reason”, and becomes “a mere plaything of other people’s thoughts”.20 The method has a shadow. In the words of one modern account of his thought, “Up to adolescence at least, the educational program comprises a sequence of manipulations of the environment by the tutor.”7 The boy finds things for himself. The tutor has arranged, out of his sight, what there is to find.

The sixth chapter of the Chāndogya Upaniṣad, an ancient Sanskrit text, tells of a father whose son had been told too much. In F. Max Müller’s translation, Uddālaka sent his son Śvetaketu to a teacher at twelve. The son came home at twenty-four, “having then studied all the Vedas”, and “conceited, considering himself well-read, and stern”. His father asked whether he had ever asked for the teaching “by which we know what cannot be known”. He had not. “Surely those venerable men (my teachers) did not know that,” the son said. “For if they had known it, why should they not have told it me?”21 He had come to believe that whatever is known is told.

“Be it so,” the father said, and he told him a great deal: about being, heat, water and food, about sleep and hunger, the rivers and the tree. From the eighth lesson on, each closed on the same refrain. But twice, when the son asked, “Please, Sir, inform me still more,” the father first gave him something to do. “Fetch me from thence a fruit of the Nyagrodha tree,” he said: the banyan. The son broke it and found seeds, “almost infinitesimal”. He broke a seed. “What do you see there?” “Not anything, Sir.” Another day the father had him put salt in water overnight and bring it back in the morning. The son looked, and “found it not, for, of course, it was melted.” “Taste it from the surface of the water. How is it?” “It is salt.” He tasted it from the middle and from the bottom. It was salt. Then the father said once more what each lesson had been for: “It is the True. It is the Self, and thou, O Svetaketu, art it.” In Sanskrit, “thou ... art it” is tat tvam asi, “you are that”.21

The Upaniṣad is teaching a metaphysics: that the self in each thing is one with the being of all things. This book needs none of it. What I borrow is a method, and the reading is mine. A father who has told his son a great deal makes him fetch, break, look and taste before he says it again. The son had come home from twelve years of being told.

The two voices agree on part of the first half: a child should meet the thing, not only hear of it. But Uddālaka arranges the experiment and supplies its meaning, while Rousseau distrusts any authority that tells a child what to think. They part most plainly at the end. At the end of the banyan lesson Uddālaka says “Believe it, my son”: a conclusion given on his own authority, the very step Rousseau forbids. They also part on who does the teaching. Uddālaka teaches his own son in his own house, after a school has sent the boy back proud. Rousseau’s boy has a tutor who, Rousseau insists, must not do it for pay, and no parents in the story: “Emile is an orphan.”20 Rousseau wrote the most famous modern book on raising a child. The five children he had with Thérèse Levasseur were all left at the foundling hospital in Paris shortly after they were born.7 That does not refute Emile. It shows how far a theory of raising children can sit from the work of doing it. The two texts take up partly different questions, one about the nature of the self and the other about an imagined education. But both ask what Diya’s adults must ask: who arranges the finding?

Applied to the phone, and the application is mine: the phone in Diya’s hand plays the part of the Upaniṣad’s teachers. It tells. Whether she comes home at twenty-four well read and sure depends less on the phone than on one habit around her. Does someone sometimes ask her to fetch and break and taste before she asks? That is not a rule against machines. A machine that asks first is also a machine. It is a question about who arranges the asking. Rousseau’s tutor is the oldest picture of a company’s default: someone who shapes what a child meets, out of her sight, meaning well. Guidance turns into manipulation when the arranging is hidden and the conclusion fixed in advance. Scaffolding is guidance built to be taken away. Nor is asking the whole of learning. A child also learns by being told, copying, practising and remembering. The aim is that she can move between asking, looking, trying, reasoning and changing her mind. A school’s syllabus arranges it in the open. So does a father on a Saturday, if he asks her what she thinks before he reads her question to the phone.

It is also a question about the adults. The father handed over his phone because the replies were good and lunch was waiting. Meera had stopped reading some of her reviews before she signed them. She could not have shown her niece how to look first that afternoon. No single afternoon fails a child. But habits are made of such afternoons, and both adults showed her what adults do. A parent cannot easily model a practice he does not have.

Restriction alone is not a strategy. On the research about screens, anyone who claims certainty on either side is selling something. The print edition’s worry about relationships deserves a narrower form. Children need people with needs of their own: someone to wait for, argue with and make up with. An assistant built to accommodate offers little of that. Whether time with it crowds such things out is a risk to watch, not yet a finding. Whether children who never knew a world without such replies will feel any need to ask first is a question this book leaves open.

Diya, for her part, did what eight-year-olds do. She had her question, and then she wanted lunch.

37.4 The Automated or Amplified Decision: Three Questions, and the Attention They Need

Diya and Meera both met replies faster than their attention, though not with the same stakes. A child given replies too soon loses practice. An executive who signs too fast exposes others to risks she has not examined. No framework covers every choice. This section is about one that its own builder stopped reading.

The governance team at Meera’s site began work in December 2025. It has three members: a former quality analyst, a data scientist and a bioethicist. From the start, every review the team made of a new or changed system in Meera’s operations ended with two signatures, the system owner’s and hers. In March 2026 she signed a review that let the procurement agent, a machine that compares suppliers, send routine requests for prices to suppliers in the company’s name. Until then she had not let it send anything to a supplier at all. In August the bioethicist asked her why that review set no limit on what the agent might say. Meera could not remember the review. Her signature was on it.

Nothing had gone wrong. The team’s own tests had checked the agent’s draft messages against the supplier records, and they had caught the two errors that mattered before the change went live. This call goes to the machines, and to the people who tested them. Meera’s signature had added nothing that she could name.

The form she signed asked the three questions the print edition set out, and they are still good questions.

“Is the outcome reversible?” Keep it. The old reason was that weighing what cannot be undone is “a capability machines do not possess”. That is withdrawn. The reason is simpler, and it does not depend on what machines can do. What cannot be undone can at best be put right in part, so the answering has to come before the act. What can be reversed, what can be repaired and what cannot be restored are different things. An order can be cancelled after the supplier has bought the materials for it. This book has already matched an agent’s freedom to how easily its acts can be undone. It has also asked for speed where a choice can still be undone and slowness where it cannot.

“Are the relevant variables quantifiable?” Narrow it. The old text said that where information is qualitative, “a feeling about a person’s reliability, a sense of what a market will accept”, then “consciousness is required”. As of October 2026 machines read letters, voices and images, so “qualitative” is no longer the line. Ask instead whether what matters is in what the machine is given at all, and who decided what it would be given. Someone always chose what counts. And what can be measured is not always what should count.

“Does the decision affect a human being’s dignity, livelihood, or wellbeing?” Keep it, for this book’s third never: “being intelligent will never give a machine the right to decide for people”. The person decided about must be able to ask why, and be put right. The old text wanted someone “who can look the affected person in the eye”. Eye contact is not answering. An explanation and a remedy are.

Run the March review through them, and the trouble shows. A request for prices can be withdrawn the next day. Its words cannot be unsent, and they can reveal what the company needs or will pay. But a request does not bind the company to buy, and a wrong one can be put right by the next message. What mattered in it was in the supplier records the agent was given. It decided nothing about anyone’s livelihood or safety. By her own three questions, the change did not need her. Her name was on it because the form said so, not because the questions did. A signature that the questions do not call for is the easiest one to give without reading.

These questions come before a decision is handed over. The book asks others afterwards, about when to take a decision back. Both sets depend on someone reading. The questions screen a decision. They do not say who may make it, who may stop the machine or who puts things right. Those are named by role, as Chapter 36 argued.

The fourth consideration is the best of them: speed is not urgency. The AI’s ability to decide in milliseconds creates a gravitational pull toward automation. Urgency belongs to the world. It is how soon harm arrives if nothing is done. Speed belongs to the machine. It is how soon the machine can reply. A decision can be urgent for a patient and still need a day of reading. Many decisions a machine can make at once are not urgent at all. Some harms do arrive faster than a person can read, as when a car must brake or a stolen card must be stopped. There the machine’s speed is what the urgency needs, and the answering moves to before the act, to whoever set the rule it follows. Someone must also look afterwards at whether the rule was right. Confuse the two, and decisions go to whatever replies fastest.

Meera’s form could not do one thing. Three good questions can say where a person must stand. They cannot keep her attending there. Studies of watchkeeping, long before these machines, found that even a keen person stops truly seeing a display on which little happens after about half an hour. In 1983 the psychologist Lisanne Bainbridge drew the consequence for automated plants: people left only to watch a working system watch it badly.22 Ben Green studied rules that require people to oversee government algorithms. In 2022 he warned that the person placed beside a machine can make its use look legitimate without being able to oversee it.23 And this book has said before that forms become rituals, and that a ritual check is no check. Attention is only one condition. A person can attend and lack the authority to stop, hold the authority and lack the skill to read, or have both and lack the time. Two signatures can hold less oversight than one careful reading.

Meera had built the form herself. Then she held its reins for nine months without once pulling them. In the old image the Prologue borrowed, the mind is the reins of a chariot. Holding the reins is not the same as pulling them, and a desk that pulled every rein would fail in its own way. A good overseer may rarely pull. What matters is that she would notice when to. Attention is perishable. It must be renewed. A check placed in a loop decays unless someone keeps it alive.

That leaves a question this chapter can set out but not settle. When a loop cannot be kept alive, there are two honest moves. One is to give the person in it the time, and say who pays for the time. The other is to take the loop away, and have someone answer for taking it away. Between the two lie others: fewer cases sent for review, a sample chosen by risk, an independent audit, failures made easier to see, regular practice. Each still needs a name beside the choice. Nobody knows which is better in general. Fewer loops can mean fewer places where an error is caught. More people in loops they cannot keep alive can mean more names on decisions that nobody made.

37.5 What Amplification Looks Like in Practice

The amplified human is not a type of person. There is a relation between a person and the machines she works with, and in Meera’s year it had run the wrong way.

She had the reviews in front of her, and the authority to refuse any of them. What she lacked was time, and nobody had counted it. The terms she proposed in October 2025 put her name on every review. Nobody asked how many reviews there would be, or how long a careful one takes. By September there had been about thirty. Part of that failure was hers, because she designed it. Part was the site’s, because its leaders approved it without asking what it would cost.

The print edition ended this section with a chain of claims. “You can only be responsible for what you are aware of. The amplified human is aware, and therefore responsible, and therefore trusted, and therefore irreplaceable.” Meera’s year breaks the chain at its first link. In 1694 the philosopher John Locke tied being rightly held to an act to being able to take it as your own, through being conscious of it.24 Read that way, Locke stands nearer the print edition’s first link than to this book. Here I part from him, and the application to Meera is mine. Meera could have taken those reviews as hers. She had the papers, the authority and the right to refuse, and not reading was itself something she did. So she answers for them, read or unread. A signature also tells others who will answer, and hers told them truly. But to answer is to explain and to put right what can be put right. In August she could not explain a review she could not remember. Responsibility reaches what a person chose not to look at. It reaches only as far as the choice. A person given neither the time to read nor the right to refuse holds the seat in name only. Most of the answering then goes to whoever arranged it so. The chain’s last link breaks too. Being trusted makes nobody irreplaceable. It makes them answerable.

She signs her name. Not because the work is perfect. An honest signature says more than that a person read. It says what she read, what she relied on others to check, and what nobody checked.

In the week of 12 October, Meera wrote to the governance team, with a copy to Basel. She began with what had gone wrong, in the plainest words she had found. “The frameworks are the easy part. Anyone can build a governance process. The hard part, the part I still get wrong, is knowing when the framework itself is the problem. When following the process becomes a substitute for seeing what the process was designed to make you see.”

Then she set new terms. From November the owner of each system would sign its review alone, with the team’s findings attached. She would sign only the reviews of changes whose acts cannot be undone. An order placed with a supplier is one: it can be cancelled, but not wholly undone. A rule about which batches may leave the site is another. She would read each of those in full before she signed it. Her reading did not replace the team’s review. It added a second reader for changes that cannot be undone. The team’s findings stayed attached to every review, and the power to stop a system stayed the team’s. An owner could not sign past a finding the team had left open. Each quarter the team would reread a sample of the reviews owners had signed alone, and tell her what it found. She would pay for the sample’s hours from her own budget, and name them in the same quarterly report.

She put the price in the same letter. Basel’s quarterly report would show fewer signatures from the head of the site’s operations, and to a reader in Basel that could look like less care. If a review she no longer signed went wrong, she would have to explain why her name was not on it. And the hours for the reviews she kept had to come from somewhere. She took them from a monthly planning meeting she chaired, which would now meet every other month. That cost something too: fewer chances for teams to raise a clash before it grew. Hours for governance taken from another meeting are hours nobody budgeted for. Whether fewer signatures read in full will catch more than many signatures read in part, nobody knows. She wrote that it was a bet, and that she would report each quarter what her reading had found. Her care could not scale by itself. What would make her afternoons possible was that, from November, fewer reviews would reach her desk.

That is what amplification looks like when it runs the right way. A person hands the machines what can safely be handed over, including much of the checking they do well, and names who answers for it. She keeps up the skill to examine what matters. She keeps her own name only where she can bring her attention with it.

Ending: A Question in the Margin

The first review that the new terms would have left on her desk reached it that same week, before they began. It was for a change she had refused for a year and a half. The procurement agent would place small repeat orders with existing suppliers by itself, below a set value, with no person approving each one. An order placed is a promise to a supplier.

The review ran to about forty pages. She read the owner’s description of the change, then every test the team had run. Then she read the agent’s sample orders one by one, and the owner’s written replies to the team’s findings. It took the whole of an afternoon and two hours of the evening. She found no error. In the margin she wrote one question: when an order goes wrong, who telephones the supplier? The owner replied in writing the next morning. A named buyer in procurement would make the call.

She signed it. The change would go live two days later than its owner had planned, and for those two days his team would place the repeat orders by hand. The planning meeting she had thinned would not meet again until December. The next review on her list was due in a fortnight, and before she left the office she blocked out an afternoon for it.

Practice

The approvals you no longer read. For one week, list every place where your name or your click approves something a machine produced, checked or proposed: a review, a report, a sign-off, an “approve” button. Beside each, write when you last read the thing in full before approving it. Then ask whether you could say now, without looking, what you approved last week. Mark each one with a choice. Read it, and write whose hour pays for it. Hand back: the person who does read it signs alone, where the rules allow, and you say so in writing. Ask: ask for the time, from whoever set your week. Disclose: beside your name, write what you read, what you relied on and what you did not check. Never drop an approval that a law, a licence or a safety rule requires; if you cannot give it properly, tell whoever can change that, in writing. Tell the people each choice affects. This costs an hour, and it may cost standing, because an approval taken back can look like less care. It will not keep you attentive for long, and it promises nothing about your job. It shows where your name has gone ahead of your attention, and who must decide what to do about it.

Words to keep

  • amplifier: a device that takes a small signal and delivers a larger one of the same shape; it enlarges the noise in its input too, and adds some of its own.
  • desirable difficulty: in the psychology of learning (Bjork, 1994), effort during practice that slows it down but makes what is learned last.
  • employability: being able to find and keep work; in Kanter’s bargain (1989), a company promises to keep its people employable, though not to keep employing them.
  • gig and platform work: gig work is paid work outside a regular employer–employee relationship; platform work is such work found through an app or website; both are defined in India’s Code on Social Security, 2020, in force from 21 November 2025.
  • judgement economy: an economy in which what a person adds is less often what she knows and more often what she decides and will stand behind; it may need fewer such people.
  • nyagrodha: Sanskrit, the banyan tree.
  • scaffolding: help given to a learner that is meant to be taken away as she becomes able to do without it.
  • social contract: an agreement, real or imagined, by which people bind themselves to one another under common rules; the subject of Rousseau’s book of 1762.
  • study mode: a setting in some AI assistants that asks the learner questions and, in some cases, holds back a direct answer until she has worked at the problem.
  • tat tvam asi: Sanskrit, “you are that”; the Chāndogya Upaniṣad’s teaching that the self in each thing is one with the being of all, a metaphysics this book does not need.
  • total factor productivity: the part of growth in output that more labour and more capital do not explain.
  • Upaniṣad: one of the ancient Sanskrit texts at the close of the Vedas that teach about the self and what is real.
  • urgency and speed: urgency is how soon harm comes if nothing is done; speed is how soon a machine can reply.

Sources

  1. International AI Safety Report 2026 (3 February 2026), chaired by Yoshua Bengio, section 1.3, “Capabilities by 2030”: four scenarios developed with the OECD; “by 2030, AI progress could plausibly range from stagnation to rapid improvement to levels that exceed human cognitive performance.” See Chapter 7.
  2. “Amplifier,” Encyclopaedia Britannica (britannica.com, accessed 23 October 2026): a “device that responds to a small input signal (voltage, current, or power) and delivers a larger output signal that contains the essential waveform features of the input signal”; Rohde & Schwarz, “Understanding noise figure” (rohde-schwarz.com, accessed 23 October 2026): “even an ideal amplifier would amplify both the signal and the noise by the same amount”; “A real amplifier adds its own internal noise to the input signal.”
  3. 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 (ratio 1.01); a 44.2 per cent reduction in screen-reading workload. The trial measured reading workload, not posts. The AI sorted examinations and marked suspicious areas; radiologists made the recall decisions. The fuller results: the trial’s report on interval cancer, sensitivity and specificity (first author Jessie Gommers, senior author Kristina Lång), The Lancet, January 2026, as reported by EurekAlert (29 January 2026): “1.55 interval cancers per 1,000 women (82/53,043) in the AI-supported mammography group” against “1.76 interval cancers per 1,000 women (93/52,872) in the control group”, “a 12% reduction”. The abstract (vol. 407, no. 10527, 31 January 2026, pp. 505–514; PubMed) gives the ratio as 0.88 (95% CI 0.65–1.18), “non-inferior”, and sensitivity of 80.5% against 73.8% (p=0.031). The authors note one country, one mammography device and one AI system; in the release Lång asks for further studies of later screening rounds and of cost-effectiveness. All participants had completed two years of follow-up by December 2025 (Lund University news, 30 January 2026). On what was measured: the Lancet report names the interval cancer rate as the trial’s primary outcome, with sensitivity and specificity among its secondary outcomes; the Lancet Digital Health report gave secondary outcomes, among them recall, cancer detection, false positives and reading workload; neither reports deaths (abstracts in Lund University Publications). On how the sorting worked (for “Under the hood”, 37.1): Kristina Lång et al. (last author Aldana Rosso), “Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study,” The Lancet Oncology 24, no. 8 (August 2023; online 1 August 2023): 936–944, abstract (Lund University Publications, checked 31 October 2026): an “examination-based malignancy risk score on a 10-level scale” was “used to triage screening examinations to single reading (score 1-9) or double reading (score 10)”, with “computer-aided detection marks (for examinations with risk score 8-10)” available to the radiologists; the control group had “standard double reading without AI”. The Lancet Digital Health report (abstract): “There were 61 248 screen readings in the intervention group and 109 692 in the control group”. See Chapter 7.
  4. Joshua Gans, “Radiologists are the canaries in a coal mine,” Substack, 26 September 2025 (Hinton’s 2016 remark; a record 1,208 diagnostic radiology residency positions in 2025; diagnosis “only about a third of their time”); Fortune, 4 May 2026, on radiologist employment rising about 10 per cent over the decade. Neither source gives a figure for imaging volume. See Chapter 4.
  5. 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”; published in Economic Policy (2025). See Chapter 5.
  6. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Stanford Digital Economy Lab, “Canaries in the Coal Mine” update, 12 August 2026: employment of workers aged 22–25 in the most AI-exposed occupations “now stands about 19% below where it would be” had it moved with less-exposed peers, “primarily through reduced hiring of young workers rather than increased separations”; the authors call these “descriptive patterns, not causal estimates of the effect of AI”. See Chapter 4.
  7. Christopher Bertram, “Jean Jacques Rousseau,” Stanford Encyclopedia of Philosophy (substantive revision 21 April 2023): Emile and The Social Contract “appeared in 1762”; “Emile was condemned in Paris and both Emile and The Social Contract were condemned in Geneva”; Thérèse Levasseur “bore him five children, all of whom were deposited at the foundling hospital shortly after birth”; “Up to adolescence at least, the educational program comprises a sequence of manipulations of the environment by the tutor.” For Paris: Rousseau, Confessions, Book VIII (Aldus Society translation, 1903; Project Gutenberg 3908): “My third child was therefore carried to the foundling hospital as well as the two former”; “I have had five children in all”; he was then living in Paris.
  8. Jean-Jacques Rousseau, The Social Contract and Discourses, trans. G. D. H. Cole (London: J. M. Dent, Everyman’s Library, 1913), The Social Contract, Book I, ch. VI, “The Social Compact” (Project Gutenberg eBook 46333).
  9. Directive (EU) 2024/2853 on liability for defective products (23 October 2024): software, including AI systems, counts as a product; to be transposed by 9 December 2026; Article 2(1) as corrected on 7 May 2026: it applies to products “placed on the market or put into service after 8 December 2026”. See Chapter 15. On “electronic persons”: European Parliament resolution of 16 February 2017 with recommendations to the Commission on Civil Law Rules on Robotics (P8_TA(2017)0051), paragraph 59(f): “creating a specific legal status for robots in the long run”, so that the most sophisticated could be “electronic persons responsible for making good any damage they may cause”; European Parliament resolution of 20 October 2020 on a civil liability regime for artificial intelligence (2020/2014(INL)), paragraph 7: “it is not necessary to give legal personality to AI-systems” (as quoted by Vagelis Papakonstantinou and Paul de Hert, ai-regulation.com, 26 November 2020). On property, data and contracts: Directive (EU) 2024/2853, Article 6(1)(b)(iii): damage to property counts except to “property used exclusively for professional purposes”; Article 6(1)(c): damage includes the “destruction or corruption of data that are not used for professional purposes”; Article 15: an economic operator’s liability is not, “in relation to the injured person, limited or excluded by a contractual provision or by national law”; Article 14 leaves recourse between liable operators to national law (Official Journal of the European Union, L series, 18 November 2024, on EUR-Lex).
  10. Hans Pruijt and Pascal Dérogée, “Employability and job security, friends or foes? The paradoxical reception of employacurity in the Netherlands,” Socio-Economic Review (2010), p. 3 of the repository version (the published paging differs): “Kanter (1989) lifted the term employability from the context in which it was normally used, that of people without work (cf. Orr, 1973), and applied it to employees”, citing Rosabeth Moss Kanter, “The New Managerial Work,” Harvard Business Review 67 (1989): 85–92.
  11. Erik Brynjolfsson, “The Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence,” Daedalus 151, no. 2 (2022): 272–287; abstract (arXiv 2201.04200): “As machines become better substitutes for human labor, workers lose economic and political bargaining power”; “when AI is focused on augmenting humans rather than mimicking them, then humans retain the power to insist on a share of the value created.” See Chapter 4.
  12. Press Information Bureau, Government of India, “Government Announces Implementation of Four Labour Codes to Simplify and Streamline Labour Laws,” 21 November 2025 (labour.gov.in): in force “with effect from 21st November 2025”; “’Gig work’, ‘Platform work’, and ‘Aggregators’ have been defined for the first time” (the government’s wording; the definitions were enacted in the Code on Social Security, 2020, and came into force with it); aggregators “must contribute 1–2% of the annual turnover”, “capped at 5% of the amount paid/payable to gig and platform workers”.
  13. Regulation (EU) 2026/1744 of 8 July 2026 (the Digital Omnibus on AI), amending Regulation (EU) 2024/1689: published 24 July 2026; in force 27 July 2026; new Article 4(1): providers and deployers “shall take measures to support the development of AI literacy of their staff”; “This obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual.” The original Article 4 of Regulation (EU) 2024/1689 required them to “take measures to ensure, to their best extent, a sufficient level of AI literacy” of their staff. See Chapter 25.
  14. Hamsa Bastani et al., “Generative AI without Guardrails Can Harm Learning: Evidence from High School Mathematics,” Proceedings of the National Academy of Sciences 122 (2025): e2422633122: about 1,000 students; unrestricted access raised practice scores by 48 per cent and lowered unaided exam scores by 17 per cent; a hint-giving tutor raised practice by 127 per cent with no exam harm. The abstract (IDEAS/RePEc): the researchers “deploy two generative AI tutors”, one of them, “GPT Tutor”, with “prompts designed to safeguard learning”; and “Without guardrails, students attempt to use GPT-4 as a ‘crutch’ during practice problem sessions”. See Chapter 4. On “desirable difficulties”: Robert A. Bjork, “Memory and metamemory considerations in the training of human beings,” in Janet Metcalfe and Arthur P. Shimamura (eds.), Metacognition: Knowing about Knowing (Cambridge, MA: MIT Press, 1994), pp. 185–205.
  15. Maxwell Zeff, “OpenAI launches Study Mode in ChatGPT,” TechCrunch, 29 July 2025: it “will ask users questions to test their understanding” and “in some cases, refuse to offer direct answers unless students engage with the material”; the same article reports that Anthropic had launched a similar tool, Learning Mode, in April 2025.
  16. OpenAI, Terms of Use (Rest of World version, accessed 23 October 2026): “You must be at least 13 years old or the minimum age required in your country to consent to use the Services. If you are under 18 you must have your parent or legal guardian’s permission to use the Services.”
  17. heise online, “Google makes Gemini chatbot accessible for children under 13 via Family Link” (May 2025), reporting the New York Times of 2 May 2025: children with Family Link accounts may use Gemini. Google, “Guide your child’s Gemini Apps experience,” Gemini Apps Help (support.google.com/gemini/answer/16109150, accessed 29 October 2026), under “Remember that Gemini isn’t a person”: “It’s important for you to help your child understand that it’s an AI tool, not a person to confide in.”
  18. Digital Personal Data Protection Act, 2023, section 9(1): the data fiduciary “shall, before processing any personal data of a child ... obtain verifiable consent of the parent” (a child is anyone under eighteen) (dpdpa.com); Digital Personal Data Protection Rules, 2025, notified 13 November 2025; section 9 is among the provisions in force only from 13 May 2027 (dpdpa.com, “DPDPA Implementation Timeline”; Bar & Bench, November 2025). See Chapter 25.
  19. Press Information Bureau, Government of India, release of 1 April 2026 on the launch of the curriculum on AI and computational thinking for Classes 3 to 8, “implemented from classes 3rd to 8th, in the session 2026-27”, by CBSE (PIB2247963). See Chapter 28.
  20. Jean-Jacques Rousseau, Emile, trans. Barbara Foxley (London: J. M. Dent, Everyman’s Library, 1911), Book I (“My first requirement, and it implies a good many more, is that he should not take up his task for reward”; “Emile is an orphan.”) and Book III. The wording was read in two reproductions of Foxley’s text, one citing the Dent printing of 1914; they differ in one comma, so only the first sentence is quoted whole and the rest in fragments. The independent fact-check confirmed Books I and III in readbookonline.org’s chaptered copy of Project Gutenberg eBook 5427, which prints a semicolon after “cease to reason”.
  21. Chāndogya Upaniṣad VI.1.2–3, 7; VI.12.1–3; VI.13.1–3, trans. F. Max Müller, The Upanishads, Part I, Sacred Books of the East 1 (Oxford: Clarendon Press, 1879), pp. 92–93 and 104–105 (sacred-texts.com). Müller prints a dash after “all the Vedas” (VI.1.2), so the passage is quoted in fragments; the online text of VI.12.1 lacks an opening quotation mark before “Break it”, which is not quoted. “Believe it, my son” is VI.12.3.
  22. Lisanne Bainbridge, “Ironies of Automation,” Automatica 19, no. 6 (1983): 775–779, at 776, section 1.1.3, “Monitoring”, citing Mackworth (1950) on vigilance (paraphrased). See Chapter 3.
  23. Ben Green, “The flaws of policies requiring human oversight of government algorithms,” Computer Law & Security Review 45 (2022): 105681 (paraphrased). See Chapter 16.
  24. John Locke, An Essay Concerning Human Understanding, 2nd ed. (1694), Book II, Chapter 27, §26: “person” is “a forensic term, appropriating actions and their merit” (modernised spelling; paraphrased). See Chapter 1.

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 human who emerges from the age of AI is “not diminished, not replaced, but amplified”, discovering “capabilities that were always present but never required”; that “The human’s contribution has narrowed to the domain where consciousness is required”; that the radiologist who “commands the highest referral rates” is the one other physicians call, and that “Judgement compounds. Research depreciates.”; that the new social contract is “between humans and machines”, in which the organisation offers “development security”, and that Meera’s site and a textile factory from another chapter “are evidence that the new contract is not a theory”; that a child raised with AI “becomes less conscious”, that “awareness is embodied”, that relational capacities “cannot be developed by interacting with a system that requires none of them”, and that the research “consistently shows” the benefits of less screen time; that judging what cannot be undone is “a capability machines do not possess”, that qualitative information means “consciousness is required”, and that the drift of attention is “a feature of consciousness”; that the amplified human “is the observer. The AI is the observed”; that the machine’s taste “is never surprising” and that taste “cannot be automated, and no amount of prompting will produce it”; that a manufacturer’s trust with his buyers and workers is “not optimisable” and “irreducible”; that “You can only be responsible for what you are aware of” and so the aware are “irreplaceable”; that Diya’s question is “unanswerable by any machine”; that “the ultimate moat is not personal. It is collective.” Three section titles kept withdrawn claims: “The Judgement Economy in Full Bloom”, “Raising Conscious Children in an AI-Native World” and “What the Amplified Human Looks Like in Practice”; two others overclaimed (“Between Humans and Machines”; “A Framework for Every Choice”). It also opened on a Sunday-evening flight from Basel that repeated the previous chapter’s close; told the one Basel conversation a second time, differently; gave Meera “fourteen years” at NovaChem and “fourteen years of building governance systems”; gave her a notebook copied from another character’s private practice; put Sagar “five hundred kilometres to the northwest” of Bengaluru and Bengaluru “to the east” of Hyderabad; put Diya in a Jayanagar school whose teacher it linked to another chapter’s teacher through a programme that did not exist; called that teacher’s method “verification-based pedagogy”; and gave her a district committee, test scores and a letter about “twelve more schools”; quoted an unnamed Bengaluru firm’s “head of talent”; credited Waldorf schools with “positive outcomes”; described East Asia’s bargain as “lifetime employment”; gave the factory cameras that missed only what “required tactile judgement”, and its owner suppliers “in the villages where the cotton is grown”; retold two other chapters’ stories, a junior lawyer’s shaking voice and a test lead’s fear; and closed its last scene on “Consciousness transmitted. The subject of the next chapter.”
  • What I now hold: nobody can draw the human of 2035, and the Part’s title stands only as a question about people already alive; what holds without a date is the three nevers, and everything else carries a year, with any 2035 claim marked as a projection with what would show it wrong. An amplifier enlarges whatever it is given, noise included, so “amplified” names a relation that runs both ways, not a new kind of person. When reading gets cheap, deciding what it is for grows more useful, but usefulness is not a wage, and an economy of judges may need few judges (Sanjay’s unadvertised post; the Swedish trial’s 44 per cent), though the 2016 forecast about radiologists failed. A social contract has people as its parties; as of 2026 no law treats a model as one; makers and users never step back, as parents do; “development security” is an employer’s offer, and the share of the gain is a choice about which machines are built and where the gain goes, with states writing parts of the contract unevenly. Every child is conscious; the danger is replies that arrive before the question has formed, and in the trial cited the harm came from how the machine was set and used, not from the machine alone; companies, states, schools and parents arrange what reaches a child; Rousseau and the Chāndogya’s father agree that a child should meet the thing, not only hear of it, and part over authority and over who teaches. The three questions stay, with new reasons (what cannot be undone cannot be put right; the question is what the machine is given and who chose it; the third never), and speed belongs to the machine while urgency belongs to the world; but no form keeps its holder attending, so a loop that cannot be kept alive must be given time or taken away, with someone answering for either. Responsibility reaches what a person chose not to look at, and only as far as the choice; a signature covers what was read and says what was not. Where harm arrives faster than a person can read, the machine’s speed serves the urgency, and the answering moves to whoever set its rule. The measure of amplification is whether a person becomes more able to understand, decide, act and answer, or more dependent; one machine can amplify one party and weaken another, and in each story what is at stake is practice. How many judges there are is set by hospitals, payers and training bodies, not by the machine; cheaper work can go four ways. A company is a person in law that acts through people; contracts shift risk between businesses but cannot cut what is owed to an injured person. Not being free of a machine is not being to blame for everything it does; a machine may brake a car under a rule people answer for, but has no right of its own. Asking is not the whole of learning; scaffolding is guidance built to be taken away; using an assistant is not doing without one. The three questions screen a decision but do not name who decides, and between giving time and removing a loop lie other moves. An honest signature says what was read, what was relied on and what nobody checked; an owner who signs alone needs a team that can stop a system and sample the reviews.
  • What changed my mind: the Prologue’s rule against defining people by what machines cannot yet do; the International AI Safety Report’s range for 2030; what an amplifier is; the Swedish screening trial and the record of the 2016 forecast; Stanford’s figures on young workers; Rousseau’s test of who can be a party, and Kanter’s older name for the bargain; Brynjolfsson on who keeps a share of the gain; India’s labour codes and the EU’s softened duty; the Turkish trial, which put the harm in how the machine was set and used; the companies’ age rules and India’s data law; the Chāndogya’s father, who made his son break a fruit and taste salt; Bainbridge and Green on watching; Meera’s own unread signature, which refuted the print’s “You can only be responsible for what you are aware of”; the audit’s corrections to the dates, the distances, the canon and the invented quotation; Bjork on desirable difficulties; the EU directive’s bar on contracting out of what is owed to an injured person; and three outside readings of an earlier version of this chapter, in October 2026.
  • What survives: the chapter’s five subjects in their numbered sections; Diya, her question (“But does the leaf know? Does knowing feel like something?”) and “She was bored with photosynthesis. She wanted to go to lunch.”; Sanjay as the judgement economy’s case; the judges as a new elite, kept as “something true in it”; “The contract must be collective, or it will not hold.”; “Restriction alone is not a strategy” and “anyone who claims certainty on either side is selling something”; the three questions in their own words (“Is the outcome reversible?”; “Are the relevant variables quantifiable?”; “Does the decision affect a human being’s dignity, livelihood, or wellbeing?”); “speed is not urgency” and the milliseconds sentence; Meera’s frameworks lines, now in her own letter; “attention is perishable. It must be renewed.”; “She signs her name. Not because the work is perfect.”
  • Section titles in the print edition: 37.2 was “The New Social Contract Between Humans and Machines”, and 37.4 was “The Automated or Amplified Decision: A Framework for Every Choice”. The print titles of 37.1, 37.3 and 37.5 are listed above.

↑ Back to top

Discuss

Loading the discussion…

More in Artificial Intelligence
  •  ·  26 min read

    Epilogue

    Late in the Gītā, the one who knows most says: ponder it thoroughly, then act as you like. A machine can now set out the reasons on every side. Saying “I answer for this” then is not pride. It is a reply, for your part, to those owed one.

  •  ·  37 min read

    Chapter 40: What This Book Cannot Tell You

    The print edition’s last chapter reduced the book to one sentence, and rested human worth on what machines cannot do. I withdraw both. A book can date its claims and correct them in public. It cannot sit in the room where you decide.

  •  ·  6 min read

    Part IX: The Unfinished Chapter

    The print edition’s last Part said that consciousness never concludes. This one is a single chapter that turns the book on itself: what a book can stand behind, what it owes those who read the print edition, and what it leaves open.