# AI-Proof: What Machines Can Never Be — Volume I: Awakening, Part III: The First Awakening, Chapter 11: Judgement: The Skill That Compounds

By Nithin Palal · 11 November 2026 · 34 min read
Section: Artificial Intelligence · Series: AI-Proof: What Machines Can Never Be
Source: https://nithinpalal.com/ai/ai-proof/chapter-11-judgement-the-skill-that-compounds

> A designer with perfect taste built a beautiful website for guests the hotel no longer had. When a machine can weigh options as well as you, judgement is not the override. It is learning honestly from mistakes, and knowing when to yield.
*Part III looks at the person doing the work. Chapter 9 asked what the work is for, and Chapter 10 what the person doing it is made of. This eleventh chapter asks how judgement grows once a machine can weigh the options, and when it should give way.*

← Previous: [Chapter 10: Building Deep Human Capital](https://nithinpalal.com/ai/ai-proof/chapter-10-building-deep-human-capital) · Series hub · Next: [Chapter 12: Becoming AI-Leveraged](https://nithinpalal.com/ai/ai-proof/chapter-12-becoming-ai-leveraged) →

**A designer with perfect taste built a beautiful website for guests the hotel no longer had. When a machine can weigh options as well as you, judgement is not the override. It is learning honestly from mistakes, and knowing when to yield.**

**In Brief:** *For readers who wonder what their judgement is for when a machine can weigh the options as well as they can. Chapter 11 follows a designer in Mumbai whose taste failed a client, an editor in Lisbon whose self-scrutiny made her slow, a manufacturing director in Pune deciding what happens when every factory buys the same inspection model, and a logistics owner in Lagos whose best decision was luck. Aristotle on practical wisdom, the Bhagavad Gītā on the clear intellect and on action without fruit, the Tirukkuṟaḷ on delay, the Pañcatantra’s lion-makers and Plato’s ship ask how judgement grows, how to judge a decision by what was known rather than by how it turned out, and when the machine should decide.*

**Judgement is not the override: it grows from mistakes judged by what you knew rather than by how things turned out, and it includes knowing when the machine should decide, and who answers for letting it.**

## Opening: The Website for Other Guests

Arjun Mehta lost the most important job of his year on a Wednesday in October 2024, and the worst part was that he had been warned.

The client was a hotel chain in Goa: four small properties, with three more planned, and the kind of quiet luxury that draws Europeans tired of Bali. The year before, Arjun had designed their brand from his studio in Parel, in Mumbai, two rooms above a printing shop. The identity was beautiful: clean lines, a muted palette, letters that spoke softly. It sat at the top of his portfolio. Much of his growing name rested on it.

When the chain came back for a new website and booking system, Arjun assumed he knew what they needed. He had made the world they lived in. At the first meeting, the chain’s operations director, Vandana, who had not been part of the branding work, mentioned that the guests had changed. More Indian families. Fewer couples from Europe. More bookings arriving through the big sites that compare hotel prices. Arjun noted it and moved on. The design problem, as he saw it, was the same as before: elegance, restraint, taste.

He spent three weeks building something beautiful. The type was perfect. The pages moved smoothly from one to the next. He presented it with the calm of a person who knows he has done excellent work.

Vandana looked at it for ten minutes. “This is beautiful,” she said. “It will not work for us.”

The families wanted to know about room sizes, the distance to the beach, and whether the pool was safe for small children. The people comparing prices needed to see the price quickly, not behind a slow and elegant animation. Vandana did not say any of this unkindly. She said it like someone explaining weather to a person who had dressed for a different season.

Here is what the print edition did not say. The evidence had been sitting in the chain’s own booking system for a year: the ages of guests, the rooms they booked, the sites they came from. A design tool given that data and the brief would probably have started with families. Arjun had the evidence from a person’s mouth and did not hear it. A machine would not have needed to hear it. It would have read it.

The chain did not drop him at once. They gave him a second chance, with conditions: Vandana at every stage, and every design choice justified by what their actual guests needed. He accepted. But something more than the contract had been damaged: his belief that seeing clearly was the same as seeing completely. Chapter 7, set three months later, mentioned that Arjun had once trusted his eye so completely that he missed what a client needed. This was that time.

Machines now weigh options too. Ask one to compare three designs against a hotel’s booking data, and it will, quickly and often well. Chapter 3 defined **judgement** as deciding well when the evidence does not settle the question, and Chapter 7 gave it a date as a skill. This chapter asks what comes next: **If a machine can weigh the options as well as a person, how does a person’s judgement grow, and when should it give way?**

## 11.1 What Judgement Actually Is

Taste says: this is good.

Judgement says: this is good given everything that is true about this situation, including the things I would prefer were not true.

Arjun’s taste was real. Chapter 7 showed it at its best, in a rose-coloured label that left almost everything out. Taste told him the website was beautiful, and it was. Nothing in his taste told him it was aimed at guests the hotel no longer had. That needed something else: the willingness to hold his own preferences to account against a reality larger than them. For a gifted person, that is hard. Clients, colleagues and prizes have told him that his instincts are his greatest asset, and they are, until a problem is wider than the instinct can see. Then taste becomes a refined kind of blindness. Arjun’s eye was fine. He was not looking wide enough.

Aristotle saw why this happens to the young and gifted. In W. D. Ross’s translation of the *Nicomachean Ethics*, he notes that “young men become geometricians and mathematicians and wise in matters like these”, but “it is thought that a young man of practical wisdom cannot be found”. The reason is that practical wisdom “is concerned not only with universals but with particulars, which become familiar from experience”, and “it is length of time that gives experience”.¹ Chapter 3 met his word for it, *phronēsis*, practical wisdom. Here is its condition. Arjun at twenty-nine was something like Aristotle’s young geometer: brilliant at the universals of design, still new to the particulars of a hotel’s guests.

A few pages later Aristotle draws the conclusion that seems to settle the matter for experience. We should attend to the unproved sayings of “experienced and older people or of people of practical wisdom not less than to demonstrations; for because experience has given them an eye they see aright.”² This is the strongest case for the experienced human judge, and it meets an equally strong case against. Chapter 3 told how, since the 1950s, simple statistical rules have matched or beaten experts in field after field, and how experts allowed to overrule the rules find too many exceptions. Both cannot be true in the same place. The way through is in Aristotle’s own sentence. Experience gives an eye only where it has been the right kind of experience: long, in the same kind of case, and with honest news of when the eye was wrong. Chapter 10 called the last of these feedback. Where they are missing, years give confidence instead of sight.

The *Bhagavad Gītā* places the trouble somewhere else. Near its end, Kṛṣṇa sorts the human intellect, *buddhi*, by its quality. In Kashinath Trimbak Telang’s translation: “That intelligence, O son of Pṛthā! is good which understands action and inaction, what ought to be done and what ought not to be done, danger and the absence of danger, emancipation and bondage.” The intelligence marked by passion is the one “by which one imperfectly understands piety and impiety, what ought to be done and also what ought not to be done.”³ The verses belong to a whole theory of the three qualities of nature, which this chapter does not need. What it needs is the diagnosis. The Gītā does not say that poor judgement comes from too little intelligence. It says the passionate intelligence understands what ought to be done “imperfectly”, and elsewhere it says what passion is: it “consists in being enamoured, and is produced from craving and attachment.”⁴ Put the two together, as I read them, and poor judgement is intelligence bent by wanting. The passionate mind sees what ought to be done, but through what it wants.

The two lamps disagree, and the disagreement is useful. For Aristotle, judgement grows with time and experience. For the Gītā, judgement is cleared, not only grown: the same experience, seen through wanting, can make a person more confident and no wiser. Arjun had years of design. He also wanted this website to be the next jewel of his portfolio. On this reading, wanting can bend what a person hears. Whether it bent what Arjun heard, the Wednesday cannot settle; that would be the kind of reason Chapter 1 warned we supply too easily for what moved us. There is a plainer account. Last year’s problem had made his name, and he assumed this was the same problem. Success gives a person a working model of what wins, and a model that has won before is the hardest one to question. Either way, the evidence he ignored did not need a theory of the mind to be evidence.

What does the machine have here? The print edition said that a machine “cannot weigh competing values” and that “The machine has no stakes. It optimises. The human judges.” That is withdrawn. Tell a model what the hotel values (beauty, families, margins) and it will weigh them, propose trade-offs and explain its choice. Chapter 3 said that a machine’s stakes, where it has any, were put there by its makers and can be copied. The difference this chapter can defend is not that the machine cannot weigh. Chapter 3 put it in one line: a score is uncertainty with the person removed; judgement is uncertainty with someone attached. The someone is a named person who will still be there if the call turns out wrong, and who can be asked why. The rest of this chapter asks how that someone gets better, and when that someone should step aside.

## 11.2 The Education of Error

Judgement is not grown by instruction alone. It grows when expectations meet evidence: when decisions have consequences, and the person can tell what the decision contributed from what chance supplied. Success teaches little, because success fits too many explanations. You succeeded because your judgement was sound, or because the market was kind, or because someone else’s mistake made your ordinary decision look brilliant. Failure seems a better teacher. But failure can teach the wrong lesson too, and that is the harder part of this section.

Beatriz has bought books for a literary publishing house in Lisbon for sixteen years. She has found writers who went on to win prizes. She has also lost the house money, and the losses are where her story begins.

She keeps a journal. Sales live in spreadsheets; the journal is a record of reasons. At each decision to buy a book, she writes down what she believes and why. Not the polished version she gives at the editorial meeting: the real one, with its half-formed hunches and unexamined assumptions. Two or three years later, when the sales and the reviews are in, she reads the entry again.

The exercise is often brutal. She found patterns in her errors that no outside analysis had shown her. She overrated beautiful sentences and underrated whether a story moved. She mistook her own excitement for readers’ readiness. And she often missed what she came to call the book behind the pitch: the real reason the writer had written it. That reason was sometimes different from the story the writer told her, and sometimes from the one the writer told themselves.

Then the journal began to cost her. Over the years she grew so good at finding bias in her own reasoning that she found it everywhere, even in convictions that deserved trust. In 2022 a first novel arrived that she loved at once. Her love felt exactly like the old bias: beautiful sentences again. That year the house had also begun to use a model that scores manuscripts against the sales of similar books. It scored this one high. She distrusted that too, because it agreed with the excitement she had trained herself to distrust. Another house made an offer. On the last day before the book would go to them, her publisher pointed out that she had talked herself into doubt by sheer rigour. She bought it. It became the house’s most successful first novel in ten years.

Notice the temptation in that last sentence. It is exactly the kind of story her journal was built to distrust, and here it is, wearing a success. The sales do not prove the decision; a book that sells is compatible with a good decision, a bad one, and a lucky season, as 11.4 will argue. Nor was the model’s agreement independent proof. The model scores what resembles what has sold, and her taste had been shaped by the same books, so their errors may run together. It is often wrong, mostly about books that sell slowly for years and last. What the episode shows is narrower and more useful. Her caution was no longer tracking evidence. Two witnesses she respected, one human and one not, both saw a strong book, and she could not say what she knew that they did not. The correction that had made her wise was making her slow.

It took her two years to find her way through, and she did not give up the journal. She added a question to every entry: *Am I analysing, or am I hiding?* And she added a date: the day by which she will decide, written before the reading begins. The *Tirukkuṟaḷ*, the Tamil book of conduct, has a couplet on the conduct of affairs that she would recognise. In V. V. S. Aiyar’s translation: “The end of all deliberation is to arrive at a decision: and when a decision is come to, it is wrong to delay the execution thereof.”⁵ Deliberation that never ends is not care. It is another way of not deciding.

Why did honest attention to her errors make her worse? Three reasons, and each is a warning for anyone who keeps such a record. First, her world gives slow and noisy news. Chapter 3 described the two conditions under which skilled intuition can be trusted: clear signals, and enough chances to learn them with feedback. Publishing has neither. A book’s fate turns on a prize, a review, a season, years after the decision. In such a world a record of outcomes teaches as much about luck as about judgement. Second, her journal records the reasons she gave. Chapter 1 told how people often give confident reasons for choices that had little to do with what actually moved them. Her journal may have been teaching her about her explanations, not her decisions. Third, a remembered mistake teaches caution, and caution has no natural stopping point. Chapter 3’s surgeon, Renata, carries in her memory three men whose stents failed and one operation she advised that was not needed, and so far they have made her careful. Beatriz’s errors made her slow. The same memory can do either.

The aim, then, is not to become less confident. It is to become better calibrated, in the sense Chapter 10 gave the word. Confidence is a feeling about a decision. Calibration is a record of how often that feeling has been right. A person who has learned to distrust every instinct is no wiser than one who trusts every instinct that once paid. The test is whether confidence moves when the evidence moves.

So the education of error needs something outside the learner: a second witness who can say when the lesson has gone too far. For Beatriz it was her publisher, and, uncomfortably, the model. The question she added is the inside half of that check. The date is the outside half.

Arjun, after the Goa meeting, did something he had never done with a client. He asked Vandana for a second meeting, not to defend his work but to understand what he had missed. It lasted two hours. She walked him through the business from the operations side: how guests booked, which kinds of guests were happy and which were not, what each kind of guest was worth to the chain. None of it was design. All of it bore on the design decisions he had made without it. He left with pages of notes and the beginning of a different idea of his job. The problem had never been “make this beautiful”. It was “make this beautiful for these guests, with these numbers, for this business”. The first problem needed taste. The second needed judgement, and judgement began with the discomfort of finding that his frame had been too small.

## 11.3 Second-Order Thinking: The Consequences of Consequences

Most decisions are thought of in one step: if I do this, that follows. The world does not stop there. What follows causes something else, and that changes the conditions of the next decision. A person who looks only at the first step arrives at the last one surprised.

The French economist Frédéric Bastiat made this the whole difference between good and bad economics, in an essay of 1850. An act, a habit, an institution or a law, he wrote, “gives birth not only to an effect, but to a series of effects”. Only the first is seen at once. “Between a good and a bad economist this constitutes the whole difference”: the bad one sees only the visible effect, and the good one also sees the effects “which it is necessary to foresee.”⁶

India has an older and funnier version. The *Pañcatantra*, a collection of animal fables and stories made, as its own frame says, to teach worldly wisdom to three slow-witted princes, has a book on ill-considered action. In one story four Brahmins set out together. Three are great scholars with no sense. The fourth has little learning and a great deal of sense; in the Sanskrit text I have used, his name is Subuddhi, “Good Sense”. In a forest they find the bones of a dead lion. The first scholar puts the skeleton together. The second gives it flesh and skin. As the third prepares to give it life, Subuddhi stops him. In the Sanskrit he says, in my rendering: “Wait. This is a lion. If you bring it to life, it will kill all of us.” They laugh at him. He climbs a tree. The lion wakes and kills the three scholars.⁷ The story’s verdict is *varaṃ buddhir na sā vidyā*: better sense than that kind of learning. In Arthur Ryder’s English: “Scholarship is less than sense; / Therefore seek intelligence.”⁸ The verdict is smaller than it sounds. It is not a charter for the experienced eye. The scholars could put a lion together and could not imagine it eating them. Subuddhi could. That is second-order thinking, and nothing more is claimed for it here.

The two lamps see the second step from different places. In the fable, the unseen effect falls on the people who acted: the lion eats its makers. In Bastiat, it usually falls on others: the person who loses a sale because money went elsewhere, the worker never hired. Real decisions have both. The *Pañcatantra* is also not a book of ethics. Its wisdom, *nīti*, is the prudence of survival and success, and some of its heroes are cheats. I borrow its picture of a second step, not its morals.

Mr Deshmukh is the manufacturing director of a company in Pune that makes machined parts for car makers. He has spent thirty years on factory floors. In 2025 a vendor proposed an AI system that inspects every part with cameras and flags defects. In a three-month trial on one line, it caught more surface defects than his inspectors did, and caught them faster. He did not argue with the trial. On the first step, the machine was better.

Two of his competitors already used the system. That was the first-order argument for buying it: they had, and their costs had fallen. Then he learned something the sales presentation had not stressed. The same vendor’s model also inspected parts at the car maker’s own receiving dock, where his parts and his competitors’ parts were checked before they went into cars.

That was the second step. Until now, a defect that slipped past his inspectors might be caught by the car maker’s inspectors, who looked differently, at different times, for different reasons. Their mistakes did not line up. If every gate used the same model, their mistakes would tend to line up. The model would meet different parts and different light at each gate, so its errors would not be identical everywhere. But a kind of defect it had never learned to see would be likely to pass every gate, in every factory that used it. Bastiat’s unseen effect fell on others; the fable’s fell on the actors. Deshmukh’s was a third kind: what happens when many people make the same decision with the same method. Two computer scientists, Jon Kleinberg and Manish Raghavan, showed something similar for hiring in 2021. When firms use the same algorithm, even one more accurate than each firm’s own judgement, the decisions of all of them together can get worse. They all see the candidates through the same ranking, and they lose the advantage of independent views.⁹ A factory is not a hiring market, so this is my extension of their model. The plain version is old: when everyone uses the same test, everyone makes the same mistakes.

Deshmukh bought the system. He did not keep his inspectors as a protest against it, and he did not assume that people would catch what the machine missed. A human check that copies the machine’s method adds little. What he wanted was a different way of looking. He kept a small team doing a different kind of check on a rotating sample of parts, with other instruments and other habits. And he agreed that after a year the check itself would be judged by one question: had it found anything the system missed? He asked the vendor in writing for the kinds of defect its model was known to miss. He told the car maker what he was doing and why. What happens to his inspectors’ own skill, when a machine does most of their looking, is a separate question, and Chapter 13 takes it up.

The arrangement cost more than the vendor had projected, and the board asked why. He said what he could not prove. “I cannot show you the defect we will miss,” he told them. “I can show you that if we miss it, everyone else will miss it too.” The board approved the second check for a year, without enthusiasm. Its reluctance was not foolish. It was the shape of its incentives. The cost of the check would appear in this year’s accounts. The failure it might prevent would never appear anywhere, because it would not happen. Judgement is exercised by people, but it is shaped by what their institutions reward, and an institution that rewards only visible savings will cut its safeguards in a good year.

He may be wrong. The model may be good enough that the second check finds nothing for years, and then it will look like waste, and someone will cut it. That is the nature of a judgement about a second step: the cost is visible now, and the benefit, if it ever comes, arrives as something that did not happen. The difference between Deshmukh and a first-order thinker is not certainty. He looked one step further, and then he chose, knowing that looking further had made him less sure, not more.

The print edition said that deciding which downstream effects matter “requires someone who cares which future arrives”. Machines can model downstream effects, and as of 2026 they can list risks like this one when asked about a whole supply chain. The vendor’s own analysis had modelled one factory, because one factory was its customer. Deshmukh’s contribution was to look at the network his factory sat in. His standing was that the board could hold him to the answer.

## 11.4 The Portfolio of Decisions

Judgement compounds. That is the claim in this chapter’s title, and the title promises more than this chapter can give. What compounds is not what most people think, and it is not quite a skill.

The usual version goes like this. Good decisions build trust, trust brings bigger decisions, you make those well too, and so on. That describes the compounding of outcomes. Something else compounds underneath it. Every working person carries a portfolio of decisions: not written down, not on a CV, but real, in the record of choices made under uncertainty and what followed. The value of the portfolio is not its win rate. It is how well you learn from it. A record is not judgement. A person can collect decades of decisions and learn nothing from them.

Nkechi runs a logistics company in Lagos. She started in 2011, at twenty-three, with one truck on the Lagos–Ibadan expressway. Fourteen years later she runs a fleet. Her formal education ended with a secondary school certificate. Her other education is fourteen years of deciding under pressure. In her business many agreements are spoken rather than written, rules change without warning, and the gap between profit and ruin can be one wrong call.

Her best decision looked like her worst for three years. Some years ago she began running long-haul freight to the north, on routes that everyone in Lagos trucking called a graveyard. Delays were unpredictable. The roads were dangerous. Two established operators had already pulled out. Nkechi went in anyway. If she is honest, it was not because her analysis was better. A client she trusted asked her to, and she was young and stubborn. For three years the routes lost money. Drivers quit. Her partners questioned her judgement, and so did she. Then two things happened that she had no part in. A competitor collapsed, and a large manufacturer, needing a reliable carrier, signed a five-year contract with the only one still running. The northern routes became her most profitable. They paid for much of the company’s growth since.

She knows, and this honesty is what makes her more than a success story, that the line between her stubbornness and good judgement became visible only afterwards. Had the competitor survived, the same decision would have been the one that broke her company.

Psychologists have a name for the mistake most people make with stories like hers. In 1988 Jonathan Baron and John Hershey gave people the same decisions, made with the same information, ending well in one version and badly in another. People rated the decision itself as worse when the outcome was bad.¹⁰ They called this **outcome bias**. The fair way to judge a decision is by what the person knew, or could have known, when they made it. By what Nkechi can now reconstruct, her northern bet was less well grounded than its success made it look. Luck did much of the work she had once credited to judgement. Whether the bet was poor when she made it, nobody can now say for certain; the client relationship she protected may have been worth more than the numbers she remembers.

The *Bhagavad Gītā* has a line that says something close to this, from the other side. In Telang’s translation: “Your business is with action alone; not by any means with fruit.”¹¹ In the poem it is advice about motive: act without being driven by what you will gain. Read as a rule for judging decisions, which is my application, not the poem’s purpose, it says: judge the act, not the luck. It also hides a trap, and Chapter 3 named its cousin in the Gītā’s eighteenth chapter. Read carelessly, “not with fruit” becomes “the results are not my business”. This book cannot accept that. To answer for a decision is to explain it, to undo it where it can be undone, and to put right what can be put right. Nkechi’s drivers who quit, and the money her partners lost in those three years, were results she had to answer for, her share of them at least, whatever the luck. Judging a decision by what was known is fair to the decider. It does not excuse her from what followed.

So what compounds? Not the wins. What compounds is a habit: keeping the decision apart from the outcome, and learning from both. Call it a habit of not lying about luck. It is a discipline more than a skill, and its gains are conditional. They depend on honest feedback, and on the next problem resembling the ones that taught you. Experience can pile up while its usefulness falls, when the rules of the road or the market change. What compounds most reliably is not certainty but the ability to examine, revise and sometimes discard what experience has taught. A decision can be good and end badly; another can be poor and end well. Nkechi learned what it feels like to hold a position for years under uncertainty. She also learned that, from the inside, conviction and stubbornness feel almost the same. So she changed one thing. Before any new bet, she now writes down what would make her stop, and by when. Chapter 9 asked readers to decide in advance what evidence would change their frame. This is the same habit, paid for in a harder currency.

The machine’s side of this is real. Nkechi gave a model fourteen years of her records of the customers she had given credit to. It found that she gave credit too easily to customers recommended by friends, and that those customers paid late more often than others. It was right. Credit, in a business with many customers and quick results, is exactly the kind of world in which patterns can be learned. A machine with her whole record learned them faster than her memory had. A machine can keep the portfolio better than a mind. The print edition said that “AI cannot build this portfolio” and that it is “among the most valuable things you own”. The first is withdrawn: the record can be kept, and read, by a machine. The second promised a value no market has agreed to pay. What stays is narrower. The decisions in the record were hers. She is the one who must answer for them, and she is the one the record can change. One more thing follows from Chapter 10. A person who sees only the cases a system flags never builds a portfolio of their own. A machine can improve today’s decision while removing tomorrow’s chance to learn how to make it. Automation does not make that learning unnecessary. It changes where it has to happen: on cases set aside for practice, where a person writes a first answer before seeing the machine’s.

This habit is worth having in the sense Chapter 10 gave depth: for its own sake, and because it is how a person becomes less dishonest about their own record. No market has agreed to pay for it, and this chapter does not promise that one will.

## 11.5 When to Yield

Here is an uncomfortable truth that this chapter has to earn: sometimes the machine’s judgement is better than yours.

Not in every field, and not for every decision. In a large and growing set of decisions, which the next paragraph describes and which are bounded, measurable and recoverable, insisting on the human override is not judgement. For those decisions, and only for them, the insistence is not consciousness. It is ego wearing consciousness as a costume.

The decisions where the machine should usually decide share some features. There is a large body of data with stable, well-understood patterns. The results can be measured. The things that matter are limited in number. Mistakes can be recovered from. And the decision benefits more from consistency than from attention to the special case. Routine credit checks, like Nkechi’s. Stock levels. Inspecting parts against a clear specification, like Deshmukh’s first step. In these decisions, some people override the system not because they have seen something it missed, but because they cannot bear not being the one who decides. They are confusing judgement with control.

Even here, better average performance is not enough. A system can do well on average and badly for a small group whose cases were rare in its data. The thing it measures may be only part of what matters. A decision can be reversible for the institution and not for the person: a loan refused in error can be granted on appeal, after the house has been lost. Before yielding, ask not only whether the system does better, but on what measure, for whom, and what happens to the person it fails.

This is the point where Plato’s ship, which the Prologue described, has its best day. The sailors fight over the helm though none has studied navigation; the true navigator is mocked as a useless stargazer. Where a decision is like navigation, a matter of skill applied to a known sea, Plato is right. Let the better navigator steer. Chapter 3 told how simple rules have matched or beaten experts for seventy years. Chapter 7 reported that people working with machines did worse, on average, than the better of the two alone. The losses were clearest when the machine was the stronger partner, and in tasks that end in a decision. Two more findings explain why people resist. In 2015 three researchers at the Wharton School found that people lose confidence in an algorithm faster than in a human after seeing both make the same mistake. So they choose the human, even when the algorithm has done better.¹² In 2018 the same team found that people were far more willing to use an imperfect algorithm if they were allowed to adjust its forecasts, even slightly.¹³ We accept the better navigator more easily when we can keep one hand on the wheel. The hand is often for us, not for the ship.

The trap catches experienced people most. A surgeon dismisses a diagnostic system, not because of clinical evidence but because deferring feels like being made smaller. A designer overrides the evidence of how people use a page because it offends his eye. A manager reads every analysis the machine produces, not to check its substance but to keep feeling necessary. Part of judgement is knowing when your judgement is not the point.

### Plot twist

The Wharton findings are not the whole story, and the rest of it runs the other way. In 2019 Jennifer Logg, Julia Minson and Don Moore gave people estimates and forecasts to make, with advice to use. Lay people relied more on the same advice when told it came from an algorithm than when told it came from a person. The authors called this algorithm appreciation. The difference from the Wharton studies is timing. In the 2015 study, people had first watched the algorithm err. Here, they had not.

The appreciation faded in two cases. It faded when people chose between the algorithm’s estimate and their own. And it faded with expertise. Seventy national security professionals in the United States, whose government work made them experts in forecasting world events, discounted all the advice they were given, the algorithm’s included. With the same algorithmic advice in hand, the lay people ended up more accurate than the experts.

This sharpens the trap described above. An expert’s danger is to give a better machine too little weight. A newcomer’s danger can be to give it too much, before seeing it fail. Yielding well needs a record of the machine’s errors as well as its successes, kept by someone who will look at it.

There are other decisions where the person must not yield, and they are the ones that shape lives and institutions. They involve the uncertainty that Frank Knight, in Chapter 7, distinguished from risk. These are not odds that can be estimated. They are chances nobody can put a number on, because cases like them have not happened often enough to count. They involve values in conflict that no single measure can settle. They depend on context the data does not hold: the history of a relationship, the politics of a team, what a decision will mean in a particular community. They cannot be undone. And some of them are not anyone’s to decide but the person whose life it is. Chapter 3’s mechanic chose his own operation, and the Prologue’s third never applies here in full: being intelligent will never give a machine the right to decide for people. Even where the machine knows best, the decision may belong to someone else. This is where this book parts from Plato. He used the ship to argue that a city should be ruled by those who know what is good. Knowing what is best is not the same as having the right to decide it.

And some decisions are not prediction problems at all. A hospital model may predict, better than any doctor, which patients will benefit most from a scarce treatment. That does not decide who should get it. The patient most likely to survive? The one who has waited longest? The one in greatest need? Those are questions about what the institution is trying to protect and whose claims count. A machine can help think them through. Its ability to do so does not settle who has the right to answer them.

Notice that yielding is itself a decision, and someone makes it. So is the earlier call that a decision is bounded enough to yield. That call can be wrong, and the person who makes it answers for it. The Prologue drew the line that matters here. To **delegate** is to choose the machine, watch it, and keep the power to take the reins back. To **abdicate** is to say “the machine decided” and let nobody answer. Yielding well is delegation. So a fair test of any arrangement in which a machine decides is three questions. Who decided to let it decide? What would make them take the decision back? And who answers if it goes wrong? If nobody can answer the first, the arrangement is abdication, however well the machine performs. And the power to take a decision back is real only while someone can still make it without the machine, at least on a sample.

A name on the chart is not enough to answer the third. Picture a manager formally responsible for a system that approves thousands of transactions a day. She cannot review them. A vendor updates the model, and she cannot see inside it. She is measured on speed and cost, not on appeals. She is the name; she is not the answer. Chapter 7 said a position is full only when the person in it can see the evidence, has the time and authority to refuse, is not punished for a good refusal, and must explain. Chapter 1 called the person who carries blame without control a moral crumple zone. Where a system runs at a scale no one person can watch, answering has to be shared. It belongs to the people who design the system, approve it, maintain it and put its failures right, arranged so that responsibility cannot fall into the gaps between them. Delegation does not require one person to inspect every output. It requires an arrangement in which errors can be found, challenged and corrected.

Priya Chandran, the test lead in Bengaluru whom Chapter 9 left with a card in her pocket, found her own rough line in January 2025. She calls it the sleep test. If the machine makes this decision and gets it wrong, will I lose sleep? If the answer is no (a false alarm on a minor bug, an order of tests that is slightly inefficient), she lets the system decide. If the answer is yes (whether a critical module is safe to release, whether a failure is a real risk or a ghost), she steps in.

The sleep test is good at one thing: it matches her attention to the stakes. But it tells her where she feels the weight of a decision. A person can feel that weight because the stakes are high, or because she fears blame, or because the decision is painful. It does not tell her how much risk there is, who is more likely to be right, or who has the right to decide. In a decision where much is at stake and the machine is the better judge, stepping in should mean checking and being ready to answer, not overriding. Stakes are a reason to be awake. They are not a reason to take the wheel. When she does disagree with the machine, the rule from Chapter 3 holds: she says what she cannot prove. The sleep test is a signal to look harder, not a verdict to obey, and later chapters will show what it misses. The three questions above are the better method.

## 11.6 Judgement and the One Who Answers

The print edition said that judgement “is what happens when consciousness meets a decision that matters”, and that the machine “does not feel the weight of a decision, because weight requires a self to press upon.” Both claims are withdrawn. Whether a machine experiences anything is an open question, which Chapter 1 set out and this book does not need to settle. Machines now weigh options, choose among them and explain their choices. In bounded decisions they often choose better than we do. If judgement meant only choosing well, this chapter would be a chapter about a skill with a date.

What this chapter has found is different. Judgement grows when mistakes are judged by what was known, not by how things turned out (Nkechi), and when someone outside can tell you that your caution has become hiding (Beatriz). It looks one step further than the decision in front of it (Deshmukh). It includes knowing when to give way, and making sure that someone still answers for giving way (Priya). A machine can help with every one of these. It can keep the record, find the pattern, model the network, and in many decisions make the better call.

Four things have run together in this chapter, and they should be pulled apart before the end. There is the quality of a decision: was it good, given what could be known? There is judgement as a capacity: can this person make better decisions over time? There is the authority to decide: whose decision is it? And there is the duty to answer: who must explain, undo and put right? A decision can be good without the decider having the right to make it. A person can have the right to decide without being the best predictor of what will happen. And someone can be held responsible without having had the power to act well. A machine can improve the first and help build the second. Neither settles the third, and neither guarantees the fourth.

So what a machine’s performance does not settle is who has the right to decide, who can challenge the decision, and who owes a remedy when it fails. Here the chapter rests on two of the Prologue’s nevers. The first is true by meaning: a machine will never be you. The third is a commitment: being intelligent will never give a machine the right to decide for people. Chapter 3 put the difference in a line: a score is uncertainty with the person removed; judgement is uncertainty with someone attached. And the open question stays open. A machine might one day form itself over years, with its own aims and its own record of failures, changed in ways its makers did not decide. Then the argument of this chapter would have to be made again.

Arjun’s second design was ready in March 2025. Before he showed it, he did something he now does with all his work. He wrote down the three strongest arguments against it. He did not write them as a critic would. He wrote them in the voices of the people who would live with it: a family trying to choose a room; the night clerk checking guests in at midnight; the owner watching the margins. Answering them made the design less striking. It was plainer, more crowded with things families needed to know, and the price came sooner.

The chain renewed his contract in April 2025, for one year, with Vandana’s name beside his on every approval. He does not know whether the second design is better, only that it is better aimed. The bookings that summer were good, and so was the season, and nobody can yet separate the two. The renewal did not prove his judgement. It gave him another year in which to use it.

### Practice

*Three voices and a date.* Choose one decision you must make this month that matters to other people. Before you decide, write the three strongest arguments against your preferred choice, each in the voice of someone who will live with the result. Then write the date by which you will decide, and keep it; the *Kuṟaḷ*’s couplet applies. Then write one paragraph about what you know now, what you do not, and what would make you change your mind. Seal it, or put it where you will not reread it soon. In six months, judge your decision by that paragraph, not by how things turned out. If a machine is helping, ask one more thing. Is this a decision where the machine is likely to be the better judge (much data, clear results, mistakes you can recover from)? If so, let it decide, and write down who decided that, and what would make you take it back. None of this will make anyone pay you more. It will tell you, over time, what kind of judge you are.

## Ending: The Page on the Wall

On the wall behind Arjun’s desk, among other people’s work he admires, there is now one piece of his own. It is the first website, the one that would not work, printed and pinned up in October 2025, a year after Vandana’s ten minutes.

It is still the most beautiful thing he has made. He has not decided whether he keeps it there as a warning or as a consolation.

## Words to keep

- **abdication:** letting a machine decide while nobody answers for it (“the machine decided”).
- <strong><em>buddhi</em></strong><strong>:</strong> the intellect; in the *Gītā*, the faculty that judges what ought and ought not to be done.
- **calibration:** how well a person’s confidence matches how often they turn out to be right.
- **delegation:** choosing a machine to decide, watching it, and keeping the power to take the decision back.
- **judgement:** deciding well when the evidence does not settle the question (Chapter 3); in this chapter, also learning from decisions honestly and knowing when to give way.
- <strong><em>nīti</em></strong><strong>:</strong> worldly prudence; the practical wisdom the *Pañcatantra* teaches.
- **outcome bias:** judging a decision by how it turned out rather than by what was known when it was made.
- **override:** a person setting aside what a machine decided or recommended, and putting their own decision in its place.
- <strong><em>phronēsis</em></strong><strong>:</strong> Aristotle’s practical wisdom, which concerns particular cases and comes with experience.
- **second-order effect:** a consequence of a consequence.

## Sources

1. Aristotle, *Nicomachean Ethics* VI.8, 1142a12–16, trans. W. D. Ross (Oxford, 1925; Internet Classics Archive).
2. Aristotle, *Nicomachean Ethics* VI.11, 1143b11–14, trans. W. D. Ross.
3. *Bhagavad Gītā* 18.30–31, trans. Kashinath Trimbak Telang, *The Bhagavadgîtâ with the Sanatsugâtîya and the Anugîtâ*, Sacred Books of the East 8 (Oxford: Clarendon Press, 1882), p. 125. The “dark” intelligence of 18.32 “understands impiety (to be) piety, and all things incorrectly”. Telang’s transliteration (“Prithâ”) is normalised here to IAST.
4. *Bhagavad Gītā* 14.7, trans. Telang (1882), p. 107.
5. *Tirukkuṟaḷ* 671 (chapter 68, “The Conduct of Affairs”), trans. V. V. S. Aiyar, *The Kural or the Maxims of Tiruvalluvar* (Madras, 1916; Standard Ebooks).
6. Frédéric Bastiat, “That Which Is Seen, and That Which Is Not Seen” (*Ce qu’on voit et ce qu’on ne voit pas*, 1850), in the nineteenth-century English translation reprinted by the Foundation for Economic Education and Bastiat.org.
7. *Pañcatantra* V, “The Lion-Makers” (*siṃhakāraka-mūrkha-brāhmaṇa-kathā*); Sanskrit text as given by learnsanskrit.org, with Subuddhi’s warning and the closing *varaṃ buddhir na sā vidyā*. Recensions differ in names and wording. The English of the warning is my rendering.
8. Arthur W. Ryder, trans., *The Panchatantra* (Chicago: University of Chicago Press, 1925), translator’s introduction (“Scholarship is less than sense; / Therefore seek intelligence”); the story is Book V, “Ill-considered Action”, pp. 442–444.
9. Jon Kleinberg and Manish Raghavan, “Algorithmic monoculture and social welfare,” *Proceedings of the National Academy of Sciences* 118, no. 22 (2021): e2018340118.
10. Jonathan Baron and John C. Hershey, “Outcome bias in decision evaluation,” *Journal of Personality and Social Psychology* 54, no. 4 (1988): 569–579.
11. *Bhagavad Gītā* 2.47, trans. Telang (1882), p. 48.
12. Berkeley J. Dietvorst, Joseph P. Simmons and Cade Massey, “Algorithm aversion: People erroneously avoid algorithms after seeing them err,” *Journal of Experimental Psychology: General* 144, no. 1 (2015): 114–126.
13. Berkeley J. Dietvorst, Joseph P. Simmons and Cade Massey, “Overcoming algorithm aversion: People will use imperfect algorithms if they can (even slightly) modify them,” *Management Science* 64, no. 3 (2018): 1155–1170. On advice before any error is seen (for “Plot twist”, 11.5): Jennifer M. Logg, Julia A. Minson and Don A. Moore, “Algorithm appreciation: People prefer algorithmic to human judgment,” *Organizational Behavior and Human Decision Processes* 151 (2019): 90–103: “lay people adhere more to advice when they think it comes from an algorithm than from a person”; experienced professionals “relied less on algorithmic advice than lay people did, which hurt their accuracy” (abstract, IDEAS/RePEc); Experiment 4’s seventy US national security professionals, and the difference from Dietvorst et al. (advice “prior to receiving any performance accuracy feedback”), in the authors’ manuscript on eScholarship. Checked 31 October 2026.

*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 judgement is “the quality AI cannot replicate”, and “That ground is held”; that a machine cannot “stand behind a choice”, “cannot weigh competing values”, and that “The machine has no stakes. It optimises. The human judges.”; that judgement about one’s own judgement “is the kind of thing that only a conscious being can do”; that judging which effects matter “requires someone who cares which future arrives”; that “AI cannot build this portfolio”, which is “among the most valuable things you own”; that the sleep test “captures something no algorithm can compute”; that “judgement is what happens when consciousness meets a decision that matters”, and that the machine does not feel the weight of a decision “because weight requires a self to press upon”; that good decisions deposit “into an account that no disruption can deplete”, and “The machine cannot make these deposits”; that the hotel chain “renewed because they trusted his judgement”. It also set Arjun’s failure in January, made the editor a venture capitalist in São Paulo, gave the manufacturing director’s inspectors a “knowledge network” (now Chapter 13’s subject), credited the logistics owner’s turn to “a new trade agreement”, opened with “This chapter is about what grows in the crack”, and closed with “The next chapter examines”.
- *What I now hold:* judgement was defined in Chapter 3 and dated as a skill in Chapter 7. This chapter is about how it grows and when it gives way. Taste is not judgement; judgement holds preferences to account against what is true (Aristotle’s particulars, which come with experience; the *Gītā*’s intelligence bent by wanting). Experience gives an eye only where it had feedback; otherwise years give confidence instead. The education of error can teach timidity: records of reasons may record explanations, slow feedback teaches luck, and caution has no natural stop, so a second witness and a date are needed (*Kuṟaḷ* 671). Second-order thinking looks at the network a decision sits in (monoculture: shared models line up errors). Judge decisions by what was known, not by outcome; that fairness never excuses anyone from answering for results (*Gītā* 2.47, read with Chapter 3’s warning on 18.17). In bounded, measurable, recoverable decisions, yield, because there Plato’s navigator should steer; yielding is delegation only when someone decided it, can take it back, and answers. Four things are kept apart: the quality of a decision, judgement as a capacity, the authority to decide, and the duty to answer. A machine can improve the first and help build the second; it settles neither the third nor the fourth. A name on a chart is not answering: a person cannot answer for what they could not see, challenge or reverse, and at scale answering is shared. Average performance is not enough to yield (for whom, on what measure, with what remedy), and a prediction does not decide who should get a scarce good. What compounds is a habit of not lying about luck, conditional on honest feedback and on the next problem resembling the last; it is worth having for its own sake and promises no wage. Calling a decision bounded is itself a call someone answers for, and the power to take a decision back is real only while someone can still make it without the machine. The open question stays open.
- *What changed my mind:* the Prologue’s rule against defining people by what machines cannot yet do; Meehl and the meta-analysis of human–machine teams (Chapters 3 and 7); Dietvorst and colleagues on algorithm aversion; Baron and Hershey on outcome bias; Kleinberg and Raghavan on monoculture; Aristotle read to VI.11; the *Pañcatantra*’s lion-makers; Chapter 13’s ownership of the inspectors’ story.
- *What survives:* Arjun and Vandana, and “She said it like someone explaining weather to a person who had dressed for a different season”; “Taste says: this is good. Judgement says…”; the education of error and Beatriz’s journal, her slowness and her question, “am I analysing, or am I hiding?”; second-order thinking and Deshmukh’s choice under uncertainty, now judged after a year by what it finds; Nkechi’s long bet and her honesty about luck; the chapter’s title, kept and argued with; the portfolio of decisions; “ego wearing consciousness as a costume”; the bounded-domain criteria; Priya’s sleep test and its limit; Arjun’s three arguments against his own work; the renewal.
- *Section titles in the print edition:* 11.6 was “Judgement as Consciousness Applied to Decisions”.

*Next in the series:* [*Chapter 12: Becoming AI-Leveraged*](https://nithinpalal.com/ai/ai-proof/chapter-12-becoming-ai-leveraged)*.*