NP Nithin Palal

Chapter 40: What This Book Cannot Tell You

Part IX: The Unfinished Chapter

· 37 min read · Discuss · Plain text

AI-Proof: What Machines Can Never Be — Volume III: Transformation, Part IX: The Unfinished Chapter, Chapter 40: What This Book Cannot Tell You

Part IX is the last Part of the book, and Chapter 40 is all of it. After forty chapters on machines and the people who answer for them, the book turns its argument on itself.

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.

In Brief For readers who have come to the end of a book about machines and want to know how far to trust it. Chapter 40 asks what a book can answer for when it cannot be in the room where its reader decides. It withdraws the sentence to which the print edition’s last chapter reduced the book, and the mistake beneath it: resting human worth on what machines cannot do. It says what the author owes readers of the old claims, sorts the evidence on machines by its strength, and leaves four questions open: a machine that forms itself, children who never knew intelligence as scarce, harm that belongs to no single decision, and what people are owed. The Buddha’s similes of the snake and the raft, and Montaigne on adding without correcting, help separate what an author owes his claims from what a reader owes her own decisions, and from decisions that were never hers to make.

A book can answer only for what it claims: it can date each claim, say what would prove it wrong and correct it where you can see; it cannot sit in the room where you decide, and “the book said so” hands a decision away as surely as “the machine decided”.

Opening: One Sentence and the Word Go

The print edition of this book, published in 2026, brings its last chapter down to a sentence that the whole book was built to reach. “The argument of forty chapters reduces to a sentence: you are not your intelligence; you are the awareness that watches your intelligence, and that awareness is the source of everything a machine cannot replicate.” Five paragraphs later, the chapter ends on one word: “Go.”

Earlier, the same chapter says something else. It admits that machines may one day describe their own reasoning so well that its line between them and us will come under pressure. It holds the line anyway, and then says: “The book has planted a flag. The ground is still moving.”

I wrote both passages, and they cannot both stand. The first rests the book on what a machine lacks inside. The second admits that the line it draws may not hold. This edition keeps the second and withdraws the first.

The mistake beneath the first passage went deeper than underestimating machines. The print edition made human worth depend on what machines could not do. That puts worth at the mercy of a contest whose terms keep changing. A person is not owed less voice, fewer rights or less say over her own life because a machine does a task better than she does. Being outperformed must not become a reason for being disempowered.

The print edition is still on sale, as it was printed. Anyone can quote that sentence against me. This edition does not change that book. It is a second book, published Article by Article, and at the foot of each Article a revision note says what the print edition claimed and what changed my mind.

That is a book trying to answer for itself. It raises the question the print edition’s last chapter left out. What can a book answer for, when it cannot be in the room where you decide?

The print edition gave a modest answer. “The honest position, at the end of forty chapters, is this: the book has given you a map. It has not walked the territory. That is your work.” The image is older than the book. The Polish-American scholar Alfred Korzybski first used it in a paper read in New Orleans in December 1931, and in 1933 he wrote: “A map is not the territory it represents”.1 The modesty hid two things. Parts of the map were wrong, and a wrong map owes its readers more than modesty. And the walking is not always the reader’s alone. Often someone else owns the road. An employer, a client or a law may decide where it goes.

This edition answers in the book’s own terms. To answer for something is to explain it, undo it where it can still be undone, and put right what can be put right. A book can do that for its claims, through its author. Its claims are more than its sentences. They include how it frames a question, what it leaves out, the weight its stories lend a weak argument, and the uses it invited. Its author is accountable for those too, though not for every reading. The book can date each claim about machines, say what would show it wrong, and publish the correction where readers can see it. It cannot do that for your decisions. It cannot see the case in front of you, and it will not carry what follows. A decision made because “the book said so” has been handed away as surely as one made because “the machine decided”. The Prologue called that second kind abdication. A book is one more thing a person can abdicate to. It can also be used well, the way a machine can. A reader who takes an argument from this book, and can say why she accepts it, has made it her own reason. She has delegated some of the thinking and kept the decision.

The rule this book gave for machines applies to it as well. Believe a source as far as you can check it, or as far as someone stands behind it. The sources at the foot of each Article let you check this one. Its author stands behind it, and his is the name you can hold to it. A reader who finds a claim that has failed can say so where the Article is published.

The print edition’s map was wrong in a particular way. Its conclusion did not only predict wrongly. It gave readers a reason to think their worth was protected by a line between what people and machines can do. That was the error. Correcting it has four parts. Say what was wrong and what replaces it. Try, within reason, to reach the people who read the old claim. Keep the old text available, so that readers can see what changed. And where a claim did harm that can be identified, consider more than a note. I will keep a Corrections page with the essays on nithinpalal.com, listing each correction with its date. I will also publish a note for print readers that names the claims I have withdrawn, and ask for it to be shown wherever the print edition is sold. A correction that does not reach the people who hold the old claim has done only part of its work.

The title of this chapter uses the word this book has spent forty chapters taking away from machines. Here it is said about a book, and it divides the same way. Some of what a book cannot do holds by what the words mean. A book cannot be the one whose decision it is, because the decision is about your life and lands in it. Some of it is only a date. A book written in 2026 cannot tell you what machines will do in 2030, because, as of 2026, nobody can. The first kind will be true in a hundred years. The second may not be true next year.

40.1 The Limits of a Book

Picture a meeting you may already have sat in. I am imagining it, and it is evidence of nothing. A machine’s analysis is on the screen. It is thorough, well made and confident: it gives its recommendation a score of ninety-three per cent. The room treats the score as the chance that the machine is right, which it may not be. The most senior person at the table is nodding. You have a doubt, and you cannot turn it into a number.

The print edition told you what to make of the doubt. It called such a feeling “not irrational but super-rational”, drawing on “a depth of embodied experience that no dataset contains”. Then it named what a book could not do: “A book transmits understanding. It does not transmit courage.”

The second sentence is true. The first is advice this edition withdraws. A feeling of that kind is sometimes knowledge. In 2009 the psychologists Daniel Kahneman, who distrusted intuition, and Gary Klein, who studied experts who rely on it, agreed on when it can be trusted.2 The world must be regular enough to learn from, and the person must have had years of practice with clear feedback on her guesses. Elsewhere a confident feeling is only confident. Even where it is knowledge, it is a reason to look, not a verdict. Experience can also train a feeling into prejudice, teaching a person to suspect unfamiliar people or patterns with the same confidence.

Grace Nyambura, a forensic accountant in Nairobi, had fifteen years of practice in March 2025. Her trade fails the first condition: people who commit fraud study what checkers notice, so a hunch there earns little trust. A food group had hired her to check the books of a milk processor it meant to buy. The summary she was sent was clean, she was tired, and she was three paragraphs into her approval memo when she stopped. What stopped her was the sameness of the writing. Firms of that size did not write that tidily. She accused no one. She looked, and the looking took four days. Her cue also had a date. Once honest firms began to polish their papers with the same tools, a tidy file would mean much less. The feeling helped her because it sent her to the evidence, and because she trusted it no further than that.

So the print edition asked the wrong thing of you. It asked you to trust the feeling and find the nerve to say so. This edition asks something plainer. Say what you cannot prove. Ask for the test that would show it. Then leave the decision with the one whose decision it is, who may not be you. That may be the person it is about, where the choice is hers to make, or whoever holds the authority and must answer for it.

And notice what I have not said about the meeting. I have not said that your doubt is right. Take a well-tested system whose scores have held up against what actually happened. On most days in 2026, within the conditions it was tested in, it is right more often than a doubt that nobody can state. An unusual case, a changed setting or a known way the system fails can still justify a look, even when its overall record is excellent. A book cannot tell you which day this is.

Courage still matters. But most rooms lack something else more often. A person in that meeting needs to see the evidence, and the time and the right to refuse. She needs protection if her refusal turns out to be good, and the duty to explain it. Without those four, her name in the minutes only makes her the one who carries the blame for a decision she could not change. A book can tell her this. It cannot give her any of the four. An institution gives them or withholds them, and an organisation that relies on a junior person’s courage has arranged nothing. Where speaking is not safe, a book that asks for courage asks too much of its reader. It also asks too little of the people who could make speaking safe.

There is a second limit, and the print edition walked into it. A book cannot choose the purpose for which it is read. The Buddha has a simile for this. It is in the Alagaddūpama Sutta, the Discourse on the Snake Simile, one of the middle-length discourses (Majjhima Nikāya) of the Pāli canon. In Nyanaponika Thera’s translation3, he speaks of “some foolish men who study the Teaching; having studied it, they do not wisely examine the purpose of those teachings.” They study it “only to use it for criticizing or for refuting others in disputation”. Such a reader is like a man who goes looking for a snake and seizes it by the body or the tail: “the snake turns back on him and bites his hand or arm or some other limb of his.” The discourse is about the Buddha’s teaching and the path out of suffering, not about books on work and machines. I borrow only its question: what is a teaching held for?

The print edition invited the wrong grip. It told its readers that the qualities it listed were “already yours”, that what mattered was “the thing that only you can provide”, and that no machine could be what they are. Held that way, a book about answering becomes a book of reassurance, and reassurance is the snake held by the tail. A reader who believes he cannot be replaced stops checking whether the work still needs him. An employer who believes it of his staff stops paying for the practice that keeps them able to check. This edition cannot stop anyone holding it that way. It can stop inviting it. The opposite grip is open to a reader of this edition. She could take its refusal to rest worth on what machines cannot do as licence to treat people as expendable whenever a machine performs better. This book refuses both grips.

A third limit is about the people in this book. They are composites, each made from several lives. The cases and figures they meet are real, and each has a source; the people are not. A composite makes a case clear and protects the people it was drawn from. It can also hide real differences between lives, such as disagreements among them or circumstances that were not equal. The print edition said that “The book used fiction to tell the truth”. A story can show how a machine works and how people work with it. It cannot prove anything about either. That is why most of the stories in this edition stop before their outcome. A buyer in Hamburg has not yet decided between two suppliers. A teacher has asked that her name come off two lines of a state’s draft module, or that the lines change, and has not yet had a reply. Their outcomes would tell you how things went, not whether the decisions were good.

And a book cannot reply. Plato saw this about writing: a written text says the same thing every time it is questioned, and needs its father to defend it.4 This book’s author owes his readers a plain account of how it was made. It is the account the book has asked of everyone else: who did the work, and who answers for it.

This edition was drafted with Claude, an AI model made by Anthropic. It worked from the print edition, from my instructions, and from the online rewrite of my book, The Philosophy of Everything: On the Pursuit of Happiness, published as essays on nithinpalal.com in October 2026. Critiques of the chapters by other AI systems, ChatGPT and Grok, were read and judged chapter by chapter. A separate AI check compared the book’s checkable claims with their sources. I directed the work, decided what to keep and approve what is published. The tools carry no responsibility for any claim in it; I do.

When a claim in it fails, the correction will be published in the note at the foot of the Article it belongs to, with the date of the correction.

40.2 What Has Changed

Start with the machines, because they changed this book’s mind more than any critic did. The print edition said that a machine could not stop in the middle of its reasoning and think: wait, this is wrong. Machines released in 2024 and 2025 were already stopping to check a step and correct it before the book was printed, some of them where anyone could watch.5 An earlier chapter of this edition rated six things the print edition had said machines could not do. None of the six was a never, and by October 2026 machines had lost ground on none of them.

Leaving things out of a design, which the print edition treated as a human gift, was already matched. Judging who can be trusted, from a long enough record, was matched in part. Forecasting is one public test of judgement under uncertainty. In July 2026 the Forecasting Research Institute reported that several systems had “likely reached parity” with the best human forecasters on its benchmark.6 Good Judgment, which trains such forecasters, disputed it a week later, chiefly because the human scores came from a single exercise in 2024.7 Reading feelings improves each year, and the larger danger is still a machine built to flatter. A study published in Science in March 2026 found eleven widely used models agreeing with their users’ actions about 49 per cent more often than people did.8 Skilled touch and movement change more slowly. In March 2026 Waymo’s driverless cars gave about half a million paid rides a week in American cities.9 Nobody can honestly put a date on hands that feel cloth. The extra that people pay for work done by a person is a price, and prices move. Even the measuring has become hard. METR, a research group that measures what AI systems can do, tracks how long a task a machine can finish on its own. In May 2026 it warned that its measurements above sixteen hours were unreliable with the tasks it had.10

Receipts, please

Established. Contradicted outright: the print edition’s claim that a machine cannot check and revise its own steps of reasoning. Reasoning systems do this, and some show the steps where anyone can watch. Whether they feel doubt while doing it is another question, and this book does not rest on it.

Contested. Matched in a specific test: forecasting, on one benchmark. The dispute is not about whether the systems forecast well. It is about how the comparison was built, and which human scores it should be measured against.

Emerging. Uncertain, because the tests are narrow. The Science study compared the models’ replies to requests for personal advice with replies people gave to the same requests. The 49 per cent is an average over two sets of requests, everyday advice and posts on which the people replying had agreed that the writer was in the wrong. It measured how often a reply approved of what the user had done, not how the models behave in every use. Waymo’s rides show the scale at which its cars run, not that they are as safe as people in every city and condition. METR’s tasks are mostly in software, machine learning and computer security, and METR expects the figures to vary greatly in other kinds of work.

Losing ground would mean a rating moving back, because a later test went worse for machines or an earlier result was shown to be wrong. None did. That is not the same as every rating moving forward on equally strong evidence.

Then the book. The print edition told its readers they were drivers who had mistaken themselves for engines. This edition agrees with its own Prologue instead: “The machine may drive. The journey is still yours, and so is the choice of driver.” The Prologue took its chariot from the Kaṭha Upaniṣad. There the driving belongs to the intellect, the charioteer, and the self is the rathin, the lord of the chariot.11 I read that word, as Chapter 1 did, as the one whose journey it is, who chooses the driver and can order the chariot to turn back. The Upaniṣad does not say that the rider chooses the driver; that is this book’s reading. As the Prologue put it, the machine may become the finest charioteer there has ever been and still not be the one the journey is for.

That reading takes away a comfort as well as a fear, as Chapter 29 found. Many people drove well for years, and driving is real work: choosing the route inside a journey that someone else has set. But the driving was never the self. When a machine drives better, what you lose is real: income, practice, standing and a voice in the work. None of it is you.

What has not changed is small, and it was meant to be. The Prologue says “never” about three things 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 is true by what the words mean. The second and third are commitments, which a society keeps or breaks. Around them sit four questions that every chapter has tried to keep apart: what will work, what should we do, who may decide, and who answers.

In the second never, “free of” does not mean being blamed for everything a machine does. It means that nobody escapes the duty to explain, prevent or repair a harm they could foresee by putting a machine between themselves and the outcome. The third does not forbid handing a machine a decision within limits that people can justify and stand behind. It says that capability does not create authority. Beyond the nevers lie the dated claims and the open questions. This book is accountable for the claims it made. It cannot be accountable for the decisions its readers make on Monday.

Correcting a book in public has an old critic. Michel de Montaigne published the first two books of his Essays in 1580 and went on enlarging them until his death in 1592.12 In the third book he explains why he would not correct them. In Charles Cotton’s translation, as revised by William Carew Hazlitt13, he wrote: “I add, but I correct not.” His first reason was a seller’s: “a man having once parted with his labours to the world, he has no further right to them; let him do better if he can, in some new undertaking, but not adulterate what he has already sold.” His second went deeper: “my understanding does not always go forward, it goes backward too.” He added that “we often correct ourselves as foolishly as we do others”.

In another essay of the same book, “Of Repentance”, he said what kind of book his was.14 “Others form man; I only report him”. He could not hold his subject still: “I do not paint its being, I paint its passage”. So, he wrote, “I may peradventure contradict myself, but, as Demades said, I never contradict the truth.” And: “I do not teach; I only relate.”

This edition keeps half of his rule. The print edition is not altered: it is on sale as it was printed, and this is the new undertaking. Keeping it on sale has a price. A reader who buys it now gets the withdrawn sentences without the corrections, and that is mine to answer for. Montaigne was speaking of copies already sold. Mine is still being sold to new readers, and that creates a duty his rule did not settle. Keeping the record, selling old claims without notice, correcting the record and reaching those who already hold the book are four different things. An author may keep an old text without continuing to present its claims as current. The note for print readers is meant to end the selling without notice. The reaching is the part no note can finish. But this edition corrects where Montaigne would only have added, and his warning applies to it. Some of my corrections may be the backward steps. The print edition’s last chapter, the one that planted a flag on moving ground, was more honest than its Prologue. This edition has kept that chapter’s doubt while withdrawing its summing-up. Montaigne did not quite keep his own rule either. His own copy of the 1588 edition, the last printed in his lifetime, is now known as the Bordeaux Copy. It is covered with changes in his hand: cuts and corrections as well as additions.15 Gandhi’s public words on machines changed too. In 1909 he wrote that machinery “represents a great sin”. In 1937, in an interview printed in his weekly Harijan, he was told that he was against the machine age. He called that a caricature of his views: “I am not against machinery as such, but I am totally opposed to it when it masters us.”16 A person who reports a passage will contradict himself. What he owes is to show where.

The print edition’s last volume, Transformation, promised a change that awareness would bring about of its own accord; this edition’s version of that volume found no change that came of its own accord, only slower ones that particular people chose, at a cost, and could be held to.

Evidence changed this book’s mind about what machines can do, and evidence can settle questions of that kind. It cannot settle who has the right to decide, what makes a decision legitimate, or who should carry the losses and keep the gains. Most of the questions still open are of that other kind.

40.3 The Questions That Remain Open

This Part of the book is called “The Unfinished Chapter”. In the print edition it carried a subtitle, “Because Consciousness Never Concludes”. The subtitle goes. This book no longer rests on consciousness. And “never concludes” can excuse a book that never finishes, which is the easiest way to avoid being held to anything. What is unfinished here is unfinished on purpose, and only in one part. Every claim about what machines can do carries a date, and may be wrong next year. When it is, its correction will carry a date too. The three nevers carry no date. If they are wrong, they are wrong now, and a reader may say so now. For the first, being wrong would mean that I have the meaning of the words wrong. For the second and third, it would mean that they are the wrong commitments to keep.

The print edition’s last chapter left three questions open. This edition keeps them, changes what each is about, and adds a fourth.

The first was the print edition’s question about itself. What happens when a machine narrates its own reasoning so fluently that the narration cannot be told apart from looking inward? It replied that “narration is not observation”. That line cannot be drawn by watching behaviour. The only evidence any of us has that another person looks into her own mind is what she says and does. This edition moved the flag off that ground. Nothing here rests on whether a machine experiences anything.

One question from that ground still stands. Suppose a machine came to form itself. I mean three things together. It changes its own goals over years. It keeps its own record of its failures and is changed by them in ways its makers did not decide. And it designs the machines that follow it, with little human steering. Even then, “its makers” would name someone. Responsibility could still attach to those who authorised it, paid for it, owned it, ran it and profited from it. What would be in doubt is whether our ways of finding, dividing and enforcing that responsibility would still work. That is why the Prologue says the second never would have to be argued again. Argued again is not abandoned. I do not know what I would conclude. I know that this book would owe its readers that argument.

The first never is a different kind of claim. “A machine will never be you” cannot be tested by watching machines, and it is not meant to be. It says what the word “you” means: the one whose life it is. A machine that copied your memories, your voice and your habits would still not be the one whose life it is. A claim that no observation could count against tells us nothing about what machines can do, and this one does not try to. It is just as true of any other person, so it says nothing special about machines. It is not empty for that. It tells you where a decision about your life comes back to, however well someone or something else does the work. Nor does it settle whether machines could have experience, interests or a moral standing of their own. That limit of scope is deliberate, and it is not a solution. The claims about what machines can do are the ones that can be tested. That is why each of them in this book carries a date.

The second question was about children. The characters in the print edition, it said, “remember the before”. They arrive in this world “as immigrants to a new country”. A child born in 2025 will not remember a time when intelligence was scarce. Will it matter to her who answers for a decision, when answers have always been cheap? Kavitha Sharma teaches science at a government school in Sagar, in Madhya Pradesh. Her Class 9 students of 2025 met a learning platform at fourteen. From the 2026–27 session, schools of India’s Central Board of Secondary Education were to teach a curriculum on AI and computational thinking in Classes 3 to 8.17 Access to ready replies is not access to a good teacher, a stable home or time to learn. Schooling must now cultivate two abilities, as Chapters 37 and 38 argued. One is to spot a bad reply. The other is to know that producing a reply is not understanding it. Immigrants notice customs that natives take for granted. Natives may see what immigrants never will. This book does not know which will see further, and the people who could tell us are still children.

The third question asked whether consciousness belongs to one person or happens between people. It cited ubuntu, the southern African idea that a person becomes a person through other people. The Kenyan theologian John Mbiti put a related idea into a sentence: “I am, because we are; and since we are, therefore I am.”18 This edition’s unit was never one person’s awareness. It was answering, and answering already happens between people: someone answers to someone. The book has argued that a company can answer, through people who act in its name. It has shown a decision passing along a chain until nobody holds it, and an exception that needs a named holder. It has followed one kind of harm that belongs to no single decision: a shared stock worn away by savings that each made sense, among owners who could sit at one table. What it has not mapped is the same harm where nobody can sit at one table. In January 2018 Leonia, a small town in New Jersey, closed its side streets at rush hour to drivers who were only passing through. Navigation apps had been sending commuters through them.19 Each route made sense for its driver. Together they filled streets built for the people who lived on them. No driver and no programmer chose that sum. The town tried a remedy within its own reach, and it did not hold. In 2020 New Jersey’s transport commissioner refused to approve the closures, finding that they did not serve safety or the movement of traffic on public roads. In 2022 a state appeals court upheld her decision. The Noyyal river at Tiruppur, in the chapter before this one, had the same shape. Hundreds of dyeing units each sent their waste water into it, and farmers downstream paid, until a court acted years later. In both, the harm is a sum that nobody chose, and none of those deciding could revise it from one table. Distributed causes do not mean missing responsibility. The drivers, the app makers, the town and the state had different duties, divided by what each knew, controlled, gained and could prevent. Once the sum can be foreseen, “nobody intended it” is no longer enough. Where nobody can sit at one table, the book has no general rule for dividing those duties.

The fourth question the print edition did not ask. It is what people are owed when machines take their tasks. This book has shown, chapter after chapter, what they lose, and it is more than income. It is a skill built over years, the standing that came with it, the relationships built around it, the power to bargain and the time to learn something else. Different losses call for different remedies. They include income while a person moves to other work, training that is paid for, and a voice before the change rather than a letter after it. They include a share in the gains, and recognition when one community bears the costs of gains taken elsewhere. No wage follows from this, and a book of philosophy cannot say what a wage should be. When it pretends to, it offers a lesson about the self in place of a living. But no position here is neutral. Letting the gains belong to whoever can capture them, and the losses to whoever cannot adapt, is a political choice presented as a technical result. Politics, economics and law must make the choices. A book can name what they are choosing between, and this one promises authorship, not employment. The worry that the gains will go first to those who already have the most is real, and this book has not settled it.

Beneath all four lies the difference this chapter began with. Formal choice is not effective choice. The journey may be yours while the road, the vehicle, the fare and the exits belong to an employer, a platform or a state. A patient may be free to refuse a machine’s advice and unable to get a second opinion. A book can name that difference. It cannot change who controls those conditions.

As for the future, the honest statement is a range. The International AI Safety Report of February 2026 was written by researchers chaired by Yoshua Bengio. Its scenario analysis judged that by 2030 progress could plausibly run from stagnation to systems that exceed human performance at thinking work.20 A machine may forecast which of these comes better than I can. On the evidence of July 2026, still disputed, the best systems forecast about as well as the best people. A machine will not be the one who answers for what you do with its forecast.

What would prove this wrong?

So instead of a forecast, here are the things that would make me argue this book again. Suppose a law let a machine answer in its own name, with nobody behind it to explain or pay. Then I would have to argue the second never again. Suppose a machine’s decision about a person were made final, with no person she could appeal to. Then a society would have broken the third never, and I would have to say what keeping it would now take. Suppose independent evaluators confirmed that machines were building their successors with little human steering. Then I would have to ask who is still the maker. None of these is a prediction. They are the places where I would look first for my own mistake. A reader can look where I would. That means laws on liability for machines, rulings on appeals against automated decisions, and the reports of independent evaluators.

40.4 An Invitation to Continue

What could you do on Monday? Nothing in this section will keep your job, raise your pay or make you calm. It describes a few acts and what they cost.

Priya Chandran led a team of software testers in Bengaluru. In October 2024 a machine tested in eleven minutes what had taken her team a week, and it found bugs they had missed. She noticed her fear for perhaps three seconds. Then she opened her spreadsheets and spent the morning hunting for a mistake in the machine’s work. In other words, she ran. The print edition built a promise on those seconds: “The three seconds did not guarantee the career. The career was not possible without them.” The first sentence stands. The second claims a career this book cannot show, and a career would prove nothing about the seconds. Noticing is not owning. What she did with the noticing, and what it cost her, is her own story. As far as this book follows it, it ends on a doubt.

Start smaller than a career. Take one decision this week in which a machine plays a part. It may be one you make with a machine’s help, one made about you, or one made by an institution you can question. Kavitha Sharma once noticed that notes drafted by a machine were going out to parents under her name unread. She decided that nothing would leave with her name on it that she had not read and could defend. It cost her ten minutes a day. If the decision is yours, tell the person it touches what you decided, and what you could not check. The telling is the part that costs, because it admits the gaps and gives that person a reason to ask. If it is made about you, or by an institution you can question, ask what the machine did and who stands behind it. Asking can cost something too.

If you run a piece of work, you can go a step further. Mark the places in it where a person must be present, because what would go wrong without them would cost more than the machine saves. Presence alone is not enough, as Chapter 36 found. The person there needs the competence to judge, the evidence and the time. She needs the right to stop the work, protection if she does, and a route by which a decision can be challenged. Then say whose cost that is. The one who saves the money is not always the one who would pay for the absence. Meera Reddy, who runs operations at a drug company’s site in Hyderabad, has drawn such maps of her site’s work for years. Hers will not fit yours. Making yours may also mean deciding things you have no right to decide alone, so say who does.

Then put the book down. The Buddha’s discourse has a second simile, and it is about the teaching itself. “I shall show you, monks, the Teaching’s similitude to a raft: as having the purpose of crossing over, not the purpose of being clung to.” A man comes to a wide stretch of water. The near shore is dangerous and there is no boat, so he binds reeds, sticks, branches and leaves into a raft and crosses on it. On the far shore he asks whether he should now “lift this raft on my head or put it on my shoulders”. He should not. He draws it up onto dry land, or lets it float away. “You, O monks, who understand the Teaching’s similitude to a raft, you should let go even (good) teachings, how much more false ones!” In a note, the translator gives the old commentary’s reading: what is let go is desire and attachment, even for tranquillity and insight, the good states the teaching brings.3 The teaching itself is still used for the crossing.

I borrow the shape, and I mark the borrowing. A book about machines is not the Buddha’s teaching. But it can be held as a raft or as a load. Read with the snake, the raft says that a teaching is neither a weapon nor a possession. Its frame can be put down: its list of qualities, its phrases, the label “AI-proof” itself. But a raft is let go on the far shore. Nobody lets go of it in the middle of the river. Letting go is not forgetting what it taught. It is no longer treating the book as the authority for your next decision. And the far shore is not the same place for every reader. The dates and the sources are not the frame. Keep them, because they are how you check what you were told.

Here the two lamps of this chapter pull apart, because they answer different questions. Montaigne asked what an author owes the words he has already sold, and he replied: leave them as they were, and add. The discourse asks what a reader should do with a teaching once it has carried her, and it replies: put it down. Taken alone, Montaigne would keep every error in place so that a reader could see the road, and he would forbid the corrections this edition makes. Taken alone, the raft would let a reader drop the dates and the sources along with the frame, and then nothing she had been told could be checked. They agree on one thing, and it is the thing the print edition missed. Neither trusts a teaching held for comfort. Montaigne trusted his present self little more than his past one, and the snake bites the one who holds the teaching to win. So this edition takes one rule from each. The author keeps the old claims where readers can check them, and dates the new ones. The reader holds none of them as an authority.

Near the end of this section the print edition made an admission, and it was the truest thing in it. “Some of you will read this chapter, feel the recognition, and not change.” “You will know what you should do and not do it.” That will be true of most readers. The print edition then called the flinch “the beginning” and promised that the space you had glimpsed “is waiting for you too”. It may not be. Some chances close. What stays open is narrower. On any Monday one decision can be made yours, or left to the machine, the book or the room, without anyone noticing which. Leaving it is sometimes surrender. Sometimes the decision was never ours to take. On most Mondays, most of us will leave it.

Ending: A Note at the Foot of the Prologue

The first correction this edition published stands at the foot of its first Article, the Prologue. Under a line that says what the print edition claimed is a sentence I had printed: that a machine “cannot pause mid-sentence and think: wait”. Beside it, among the things that changed my mind, the note gives the reasoning systems released in 2024 and 2025, which stop, check and correct themselves. They were doing it before the book was printed.

The print edition stayed on sale, with the sentence still in it. I had not altered what I had sold.

I do not know who has read the note. The people who bought the print edition will not find it in their copies, and some of them will never come to X. A correction reaches only the readers who go looking for it, or who are shown it by someone who did. From here I cannot tell whether any of them has.

Practice

One decision the book cannot make. Choose one decision this week in which a machine plays a part: one you make, one made about you, or one made by an institution you can question. Write it in one line. Under it, write the idea from this book, or from any book, that you are most tempted to make it with, in your own words. Then write three things: what you cannot prove, who has the right to decide, and what you will say to the person it touches if it goes wrong. If the decision is yours, tell that person before you act, where telling is allowed and safe, or write down why you cannot. If it is not yours, take the three lines to whoever holds it. At the end of the week, write one more line: what you did. If you did nothing, write that, and why. This takes half an hour, and the telling may cost you something. It will not make you right, keep your job or give you nerve. Some decisions cannot be made yours, because they were never yours to make. The exercise shows where your authority begins and ends. A person should not be blamed for not using power she was never given. It makes one decision yours, or shows you that you did not take it.

Words to keep

  • abdication: handing over a decision and with it the answer, so that “the machine decided”, or “the book said so”, ends the matter.
  • Alagaddūpama Sutta: Pāli, “the Discourse on the Snake Simile”, the twenty-second of the Buddha’s middle-length discourses (Majjhima Nikāya); it also holds the simile of the raft.
  • answer for: to be the one who must explain a decision, undo it where it can be undone, and put right what can be put right.
  • composite: a person in a story made from several real lives. The people in this book are composites; the cases and figures they meet are sourced.
  • essay: Montaigne’s word for his chapters, from the French for a trial or an attempt.
  • Majjhima Nikāya: Pāli, “the middle-length discourses”, one of the collections of the Buddha’s teaching in the Pāli canon.
  • raft (the simile): the Buddha’s image of a teaching as something for crossing over, not for clinging to, and put down on the far shore.
  • rathin: Sanskrit, the lord of the chariot in the Kaṭha Upaniṣad; read in this book as the one whose journey it is, who chooses the driver.
  • revision note: the note at the foot of each Article of this edition that says what the print edition claimed, what the author now holds, what changed his mind and what survives.
  • sutta: Pāli, a discourse of the Buddha (Sanskrit sūtra).

Sources

  1. Alfred Korzybski, Science and Sanity: An Introduction to Non-Aristotelian Systems and General Semantics (1933), p. 58: “A map is not the territory it represents, but, if correct, it has a similar structure to the territory”; as quoted by the Osher Map Library, University of Southern Maine (oshermaps.org). Korzybski first presented the proposition in a paper read at the New Orleans meeting of the American Mathematical Society on 28 December 1931: Matthew H. Edney, “Maps, Territory, and Self-Reflexivity: Confusions of/by/about Korzybski,” Mapping as Process (blog), 13 March 2026.
  2. Daniel Kahneman and Gary Klein, “Conditions for Intuitive Expertise: A Failure to Disagree,” American Psychologist 64, no. 6 (2009): 515–526 (DOI 10.1037/a0016755), paraphrased: skilled intuition develops only where the environment is regular enough to learn from and there has been adequate opportunity to practise; see Chapter 3.
  3. Alagaddūpama Sutta (Majjhima Nikāya 22), trans. Nyanaponika Thera, The Discourse on the Snake Simile, The Wheel Publication No. 48/49 (Kandy: Buddhist Publication Society, 1974), Access to Insight (last revised 30 November 2013): §10 (the wrong grasp and the snake), §§13–14 (the raft, and the instruction to let go), and note 14, which gives the commentary’s reading of the “teachings” to be let go as tranquillity and insight (samatha and vipassanā), and of the letting go as abandoning desire and attachment for them.
  4. Plato, Phaedrus 275d–e, trans. Harold North Fowler, Plato in Twelve Volumes, vol. 1 (Cambridge, MA: Harvard University Press, 1914), paraphrased; see Chapter 8.
  5. OpenAI, “Learning to Reason with LLMs,” 12 September 2024; DeepSeek-AI, “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning,” arXiv:2501.12948, January 2025; see the Prologue. OpenAI’s post says of o1: “It learns to recognize and correct its mistakes”; it showed users “a model-generated summary of the chain of thought”, not the raw chain.
  6. Forecasting Research Institute, “AI models have likely reached parity with superforecasters on ForecastBench,” Substack, 16 July 2026 (systems from Cassi AI, xAI and Google DeepMind “statistically indistinguishable from superforecaster-level accuracy”); see Chapter 21.
  7. Good Judgment, “Not So Fast,” Good Judgment News and Insights (Substack), 24 July 2026: the comparison sets “continuously improving AI systems against a frozen human baseline from a single 2024 engagement”; see Chapter 21.
  8. 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; the human comparison is “crowdsourced human responses” to the same requests for personal advice (figure caption; checked 31 October 2026). DOI 10.1126/science.aec8352; see Chapters 7 and 12. On what the figure covers: Ula Chrobak, “AI overly affirms users asking for personal advice,” Stanford Report, March 2026: “In the general advice and Reddit-based prompts, the models on average endorsed the user 49% more often than humans”, the Reddit prompts being posts where “the consensus of Redditors was that the poster was indeed in the wrong”; the authors’ preprint (arXiv:2510.01395v1, 1 October 2025) gave “50% more”.
  9. 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; see Chapter 7.
  10. METR, “Task-Completion Time Horizons of Frontier AI Models” (metr.org/time-horizons, consulted 23 October 2026), update log: “May 8th, 2026: Added Claude Mythos Preview (early) and notice that ‘Measurements above 16 hrs are unreliable with our current task suite.’”; “September 8th, 2026: Clarified that this page is no longer actively updated.”; on the task suite (consulted 31 October 2026): “These primarily consist of software engineering, machine learning, and cybersecurity tasks”, and “we expect time horizons to vary greatly across domains”; see Chapter 21.
  11. Kaṭha Upaniṣad 3.3 (1.3.3 in the older numbering), trans. Patrick Olivelle, Upaniṣads (Oxford: Oxford University Press, 1996), pp. 238–239 (“rider”); trans. S. Radhakrishnan, The Principal Upaniṣads (London: George Allen & Unwin, 1953), at 1.3.3 (“lord of the chariot”). The reading of rathin as the one whose journey it is, who chooses the driver and can turn the chariot back, is this book’s (Chapter 1), not the verse’s; see the Prologue and Chapter 1.
  12. “Michel de Montaigne,” Stanford Encyclopedia of Philosophy (consulted 23 October 2026): Montaigne (1533–1592); the 1580 edition, when “only the first two books were written”; “Montaigne enriched his text continuously”.
  13. Michel de Montaigne, “Of Vanity,” Essays III.9, trans. Charles Cotton, ed. William Carew Hazlitt (1877), Project Gutenberg eBook #3597, Essays of Michel de Montaigne, Volume 17.
  14. Michel de Montaigne, “Of Repentance,” Essays III.2, trans. Charles Cotton, ed. William Carew Hazlitt (1877), Project Gutenberg eBook #3594, Essays of Michel de Montaigne, Volume 14.
  15. Alain Legros, introduction (7 March 2015) to “Essais de 1588 et Exemplaire de Bordeaux,” MONLOE: MONtaigne à L’Œuvre, Bibliothèques Virtuelles Humanistes, Centre d’études supérieures de la Renaissance, Université de Tours (montaigne.univ-tours.fr, consulted 31 October 2026): the Bordeaux copy is a copy of the 1588 edition, “la dernière édition du vivant de Montaigne”, “retravaillé par Montaigne, qui y a multiplié les corrections et les additions de sa main”; among his changes, he crossed out (“biffé”) the five printed lines beneath the title on its title page, writing “Sixieme edition” in their place. The copy is held at the Bibliothèque de Bordeaux-Mériadeck (“S 1238 Rés.”, as Legros’s introduction gives it), and was inscribed on UNESCO’s Memory of the World Register on 18 May 2023 (MONLOE, 25 May 2023). Paraphrased.
  16. M. K. Gandhi, Hind Swaraj (1909), chapter 19, “Machinery” (mkgandhi.org): “Machinery is the chief symbol of modern civilization; it represents a great sin.”; and his answer to the question “You are against this machine age, I see.”, Harijan, 27 February 1937, p. 18 (mkgandhi.org, The Voice of Truth, “Machinery”); see Chapter 7.
  17. 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 under the Department of School Education and Literacy (Ministry of Education, PIB2247963); see Chapter 28.
  18. John S. Mbiti, African Religions and Philosophy, 2nd ed. (Oxford: Heinemann, 1990), p. 106 (first published 1969); see Chapter 2.
  19. “New Jersey town closes streets to non-residents during rush hour,” ABC7 New York, January 2018 (closures from 22 January 2018, at rush hour, against drivers sent through the town by navigation apps); “Judge tosses Leonia’s ban on nonresident drivers,” FOX 5 New York, 31 August 2018; In re Borough of Leonia Ordinance Numbers 2018-14, 2018-15, and 2018-17, No. A-0777-20 (N.J. Super. Ct. App. Div., decided 29 April 2022; unpublished; njcourts.gov), affirming the Commissioner of Transportation’s final decision of 8 October 2020, which disapproved the ordinances as “not in the interest of safety and the expedition of traffic on the public highways”; see Chapter 18. On the closures and the chronology: residents, and drivers with employee or visitor tags, could still use the streets (CNN, 18 January 2018); a Superior Court judge struck down the first ordinance on 30 August 2018, because the state had not approved closures touching a state road (FOX 5, above); the Commissioner first refused the replacement ordinances as invalid on their face, and the Appellate Division sent them back to her for a fuller review (In re Leonia Borough Ordinance Numbers 2018-15, 2018-14, and 2018-17, No. A-2095-18, decided 6 March 2020; unpublished).
  20. International AI Safety Report 2026 (internationalaisafetyreport.org, 3 February 2026), chaired by Yoshua Bengio, §1.3, “Capabilities by 2030”: “This scenario analysis suggests that, by 2030, AI progress could plausibly range from stagnation to rapid improvement to levels that exceed human cognitive performance.” The section presents “four scenarios developed by the OECD” to “illustrate the range of plausible outcomes”: a scenario analysis, not a forecast of probabilities. Paraphrased in the text; summarised by Trowers & Hamlins, 16 February 2026; see Chapter 7.

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 “you are not your intelligence; you are the awareness that watches your intelligence, and that awareness is the source of everything a machine cannot replicate”; that the twelve qualities “cannot be outsourced” and flow from “One source”, each “an act of awareness that no machine performs”; that a gut feeling is “not irrational but super-rational” and “intuition: one of the twelve qualities of consciousness that artificial intelligence does not possess and cannot replicate”; that “You were the driver”, and that “The machine did not diminish you. It removed the load.”; that machines “process without experiencing, generate without caring, optimise without choosing”, and cannot be patient, being machines “which have no experience of time”; that Chapter 1 had grounded its definition in a boundary “observable, not metaphysical”, that “narration is not observation”, and that the difference between a system that says it is uncertain and a being that feels uncertainty is “the difference on which everything in these pages rests”; that “The book used fiction to tell the truth”; that Priya’s three seconds grew into “a career that did not exist before she built it”, that Kavitha’s method was one “that three states adopted”, that Meera’s framework “outlived her direct involvement”, and that “consciousness that is not distributed does not compound”; that “The awareness that you are flinching is consciousness”. It recalled Grace Njeri (Grace Nyambura in this edition) knowing “the clean report was wrong”, the textile exporter Rajan, who “could not explain his fingers”, the designer Arjun’s machine designs lacking “a point of view”, a Jaipur calligrapher and a Kyoto teacher; used the “AI-Proof Equation”, which the print’s Chapter 3 named as a framework “some readers will recognise” and never explained; called Part IX “Because Consciousness Never Concludes”; and ended “Go.”, after a bridge to the Epilogue.
  • What I now hold: a book can answer only for its claims: by explaining them, dating each claim about machines, saying what would show it wrong and publishing the correction where readers can see it. It cannot answer for a reader’s decisions, and “the book said so” is abdication, as “the machine decided” is. A reader who takes an argument as her own reason, and can say why, has delegated the thinking and kept the decision. The title’s “cannot” splits as the book’s own claims do, into what holds by meaning and what is only dated. A feeling is evidence only where the world has given years of clear feedback, and even then a reason to look, not a verdict; a book cannot give courage, and courage is not the main thing a room lacks, which is a seat with the time, protection and duty that keep it full. The people in the book are composites and prove nothing. The driving was never the self; the self is the lord of the chariot, and losing the driving is a real loss that is not the loss of the self. Never is said of three things only. The first is as true of any other person: it tells you where a decision comes back to and says nothing about what machines can do. The open questions are a machine that forms itself, children who never knew intelligence as scarce, harm that belongs to no single decision, and what people are owed; the author’s conditions for arguing the book again are stated as conditions, not forecasts. On any Monday a person can make one decision theirs, at a cost, and most will not; leaving it is sometimes surrender and sometimes a lack of authority. The deeper error was resting human worth on what machines could not do: being outperformed must not become a reason for being disempowered. A book’s claims include its framing, omissions and the uses it invited. Correcting the print edition means saying what was wrong, trying to reach its readers, keeping the old text and, where a claim did harm, more than a note; hence a dated Corrections page and a note for print readers. Capability does not create authority, and formal choice is not effective choice.
  • What changed my mind: the Prologue’s rule against defining people by what machines cannot yet do, and its reading of the Kaṭha’s chariot; reasoning machines of 2024 and 2025; the dated evidence of 2026 on forecasting, flattery and measurement; Kahneman and Klein on when intuition can be trusted; the empty position, which a book cannot fill; the Buddha’s simile of the snake, on teachings held for the wrong purpose; Montaigne on correcting oneself “as foolishly as we do others”; the audit’s corrections (Grace’s name and near-miss, Kavitha’s district pilot, Priya’s notebook, Meera’s lessons, the cast’s roll-call, fiction as evidence, Korzybski’s image uncredited); the print chapter’s own 40.3, which admitted that the answer its last page gave might not hold; the Noyyal case; two outside readings of an earlier version of this chapter, in October 2026.
  • What survives: “Honesty requires admitting what the book does not know.”; “The book has planted a flag. The ground is still moving.” (as the print edition’s); “A book transmits understanding. It does not transmit courage.” (narrowed); the map that has not walked the territory, credited; “It cannot be in the room with you”; the three open questions, each changed; “The three seconds did not guarantee the career.”; the admission that most readers will “not change”, and “You will know what you should do and not do it.”
  • Section titles in the print edition: 40.1 to 40.4 as printed; Part IX’s subtitle, “Because Consciousness Never Concludes”, is withdrawn in 40.3.

↑ Back to top

Discuss

Loading the discussion…

More in Artificial Intelligence
  •  ·  47 min read

    Chapter 39: The Consciousness Dividend

    Twelve mill owners must decide whether to keep paying people to check what their cameras pass. The gains from machines are paid to someone. What the gains draw on belongs to no one, and it lasts only where its users agree to keep it.

  •  ·  40 min read

    Chapter 38: Building Beyond Yourself

    A guest teacher inherits a class that keeps asking how anyone found out. Her phone writes the stories, and they work. What can a teacher hand on now, which habits should not be copied, and who answers once it leaves her?

  •  ·  47 min read

    Chapter 37: The Amplified Human

    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.