# AI-Proof: What Machines Can Never Be — Volume I: Awakening, Part II: The Machine at the Gate, Chapter 6: What AI Will Replace

By Nithin Palal · 9 November 2026 · 36 min read
Section: Artificial Intelligence · Series: AI-Proof: What Machines Can Never Be
Source: https://nithinpalal.com/ai/ai-proof/chapter-6-what-ai-will-replace

> A machine found the spot on the X-ray that the radiologist missed. Nobody was fired. Three posts will simply never be filled. Replacement rarely looks like a machine taking your chair. Usually it is a chair that is never offered.
*Part II turns to the machine. Chapter 5 asked what stays scarce when intelligence becomes cheap. This sixth chapter asks what the machine is replacing now, and what the word “replace” hides.*

← Previous: [Chapter 5: When Intelligence Became Cheap](https://nithinpalal.com/ai/ai-proof/chapter-5-when-intelligence-became-cheap) · Series hub · Next: [Chapter 7: What AI Cannot Replace](https://nithinpalal.com/ai/ai-proof/chapter-7-what-ai-cannot-replace) →

**A machine found the spot on the X-ray that the radiologist missed. Nobody was fired. Three posts will simply never be filled. Replacement rarely looks like a machine taking your chair. Usually it is a chair that is never offered.**

**In Brief:** *For readers who see their own work on the list of jobs AI will change. Chapter 6 asks what a machine actually replaces when it does your work. It names five forms of replacement, from the post that is never advertised to the reader who never clicks through, and follows three people inside them: a radiologist in Mumbai, a junior solicitor in London, a copywriter in Makati. Marx’s worker who “steps to the side”, Paul David’s slow dynamo and the Gītā’s svadharma help it ask the harder question: when a machine does the work better, what is lost, and to whom?*

**Replacement takes tasks before jobs, and quietly before loudly. When the work serves another, the better performer should do it; when it was also forming you, its loss is real even if the machine does it better.**

## Opening: The Nodule Behind the Rib

The machine read the chest X-ray in four seconds. Dr Sanjay Rao had read the same image the evening before, at the end of a long shift, and he had missed what it found.

A six-millimetre nodule in the upper part of the right lung, half hidden behind a rib. The kind of finding that a good radiologist might pass over in a long shift: hour nine, a backlog of cases, the canteen coffee gone cold beside the monitors. Not a bad radiologist. A good one, having a bad moment. The machine does not have moments. It found the nodule the way a calculator finds a sum, without effort and without fatigue, and without a daughter’s school fees or a sore back to think about.

Sanjay did not take it well.

He checked the finding. The nodule was there. He pulled up his own report and saw, with the clarity that only comes afterwards, exactly where his eyes had passed over the shadow. He had been thinking about something else. Or about nothing: the low hum of a mind that had read images for twenty-two years and had learned to trust its side vision a little too much.

Then he did the part that was his. He wrote an addendum to his report, under his own name: what he had missed, where it was, and that a CT scan was needed. He telephoned the patient’s doctor himself, instead of letting the correction wait in the system. The patient was a retired bus conductor of sixty-one from Jogeshwari. The doctor asked whether the man needed to be told that the first reading had missed the spot. Sanjay said yes, and that he would tell him himself if the man wanted to hear it from him. The evening’s miss was his, and so was the correction.

The hospital’s management sent a note round at the end of the month. In its first month, it said, the new system had flagged findings that human readers had missed in seven per cent of the cases reviewed. The note was cheerful. It spoke of partnership and of supporting clinicians. It did not mention the evenings, or the telephone calls that followed them. Nobody wrote the word that everyone heard.

Sanjay is fifty-three. He works at a private hospital in Andheri, in Mumbai. He chose radiology twenty-five years ago because it let him help people without sitting with them. Images, not faces. Patterns, not feelings. He is precise and careful, and more comfortable reading a scan than reading a room. He built a career on that preference, and it served him well. In February 2025 it began to become a problem.

This chapter follows three people whose work is being replaced, though none of them has been fired. Sanjay is in Mumbai. Eleanor Marsh is a junior solicitor in London. Rosario dela Cruz is a copywriter in Makati, in Metro Manila. Their industries share almost nothing. What they share is a discovery: the ground they built on was not as solid as they thought. And what its removal reveals about them is not always what they hoped. So the chapter asks one question. **When a machine can do your work, what exactly has it replaced?**

## 6.1 What “Replace” Actually Means

The word suggests a clean swap. A person walks out, a machine rolls in, and the work goes on. Almost nothing happens like that. Replacement comes in at least five forms, and the loudest is not the most common.

**Pre-emption** is the quietest. Sanjay’s hospital has not fired a radiologist and does not plan to. Three of its radiologists will retire in the next few years, and their posts will not be filled. The machine and the remaining doctors will absorb the work. In a hospital that cannot find enough radiologists to hire, this looks like good sense, and it may be. But nobody loses a job, because the job is never offered. The young doctor who trained for a decade discovers, when she looks for work, that the seat she was preparing for was never advertised. While radiologists are scarce, she may find a seat elsewhere. What shrinks first is the number of places where a young doctor learns beside someone like Sanjay. A shortage can also close this way, with the machine doing the work and nobody new being formed. This is the form Chapter 4 found in the data. Young workers in the jobs most exposed to AI are falling behind their peers, “primarily through reduced hiring” rather than through dismissals.**¹** Chapter 4 called it the missing rung. Here it appears as a kind of replacement: the replacement of someone who never arrived.

**Compression** keeps the work and shrinks the team. The same number of claims, tickets or reports is handled by fewer people, each working with the machine. It shows up in budgets and headcount reviews rather than in the news. Klarna’s announcement in 2024, that its assistant was doing “the equivalent work of 700 full-time agents”, was compression stated proudly. Its partial reversal fifteen months later, as Chapter 4 told, was compression discovering its limit.**²**

**Redefinition** keeps the job and changes what it is. Sanjay is still called a radiologist. What he does each day is starting to look less like what he spent two decades learning. The pattern is older than the current machines. In 2014 the Associated Press began publishing company earnings reports written by software. Before, its reporters had written about 300 such stories a quarter. A year later the software was producing about 3,700. No jobs were cut as a result. The agency said the change had freed about a fifth of the time its staff used to spend on earnings reports. It put that time into stories the software could not write.**³** The reporters kept their title. Their work moved.

In 1858 Karl Marx, in notebooks he never published, described this movement more exactly than most writers since. As machinery takes over production, he wrote, the human being “comes to relate more as watchman and regulator to the production process itself.” And then: “He steps to the side of the production process instead of being its chief actor.”**⁴** That is Sanjay’s new day: fewer images read, more of the machine’s findings checked. Marx did not think stepping aside was only a loss. In the same notes he imagined machinery reducing “the necessary labour of society to a minimum”, freeing time for “the artistic, scientific etc. development of the individuals in the time set free, and with the means created, for all of them.”**⁴** Everything turns on the last four words. Chapter 5 asked who gets the gain when thinking becomes cheap. Marx asked the same question about machines more than a century and a half earlier. Whether stepping to the side is a demotion or a release depends on who owns the time it frees.

**Substitution** is the form people imagine: a task a person was paid for is now done by a machine, and the person is no longer paid. It happens. At the end of 2023 the language-learning company Duolingo cut about a tenth of its contractors as it moved content work to AI. In April 2025 its chief executive told staff that the company would “gradually stop using contractors to do work that AI can handle.”**⁵** That is substitution, said plainly. It is the most visible form, and so far it is not the most common.

There is a fifth form, and it is new. Here the machine does not replace the person who makes something. It replaces the road between that person and the people who use what they make. In March 2025, researchers at the Pew Research Center tracked the web browsing of 900 American adults. When a Google search showed an AI summary at the top, people clicked through to an ordinary search result in 8 per cent of visits. When it did not, they clicked in about 15 per cent. They clicked a link inside the summary itself in 1 per cent.**⁶** The searches were not alike. Longer searches and questions drew a summary far more often, so the gap does not measure the summary’s effect alone. The writer whose page supplied the answer was not replaced. Her reader was. Call this the replacement of **the path**.

All five happen at once, often inside one organisation. Public talk fixes on substitution because it makes good stories. The quiet forms may matter more, and they are harder to see, because each of them can happen without anyone being fired.

A form of replacement says how the work leaves. It does not say what leaves with it. Chapter 4 named four goods that a job bundles together: income, practice, standing and voice. They are the wage, the daily work in which qualities are worn, the sense of being needed, and a say in how the work changes. Any of the five forms can take one of them and leave the rest, or take several at once. Eleanor, in the next section, keeps her income and her title. What she is losing is the practice that formed her.

## 6.2 Repetitive Cognitive Work

Eleanor Marsh takes the Central line to Holborn five mornings a week. From the station it is a seven-minute walk to a building whose stairwell smells of old carpet and printer toner. She is a junior solicitor at a commercial law firm whose name she mentions at parties with a careful kind of casualness. She has been there two years. She reviews contracts.

Reviews, not interprets. She reads commercial agreements page by page, checks each clause against the firm’s standard wording, and marks where they differ. Then she writes up her findings for a partner, who reads her memo for about thirty seconds before approving it. The work needs good reading, close attention and a working knowledge of contract law. It is also repetitive, and Eleanor has never said so aloud. The contracts vary, but the ways they vary are limited. After two years she can feel a difference before she sees it, the way a proofreader feels a typo at the edge of her vision.

In the spring of 2025 the firm began running every incoming contract through an AI review tool. It does the first pass in minutes. It does not have Eleanor’s feel for a clause, but it is accurate enough for what a first pass is for. The first pass is a filter. Eleanor had always been the filter, not the judge.

What this chapter has to say about Eleanor is something she has not yet said to herself. The review was a hiding place.

She chose it, though not on purpose. Nobody joins a law firm asking for the most repetitive work available. But she drifted toward it, volunteered for it, and became known for it, because the review had clear edges. At the end of every day she knew whether she had done the job. The other work, advising, negotiating, sitting across from a client whose business is at risk and saying something that matters in real time, has no such edges. It is open, unscripted and frightening, and Eleanor has managed to avoid it for the whole first part of her career.

The machine has taken the hiding place. What it leaves behind is a junior lawyer who does not yet know whether she is any good at the part that remains.

And yet the hiding place was also a school. Two years of reading clauses is how she learned to feel a wrong one. Chapter 3 said that qualities are worn into a person by repeated choices, and the repetition that sheltered Eleanor was also wearing something into her. The machine takes both at once, the shelter and the schooling. It cannot tell them apart. Neither, yet, can she. This is Chapter 4’s missing rung, seen from inside a job that survives. The work best suited to a machine may be the very work through which people became able to do the rest.

This is the pattern across repetitive cognitive work. The paralegal who summarises witness statements. The analyst who compiles the quarterly report. The compliance officer who checks filings against the rules. Each role needs real expertise. Each follows a logic structured enough that, once a machine can handle language well, the human part shrinks to the judgement calls. And the judgement calls are a small share of the hours. What the removal exposes is different for every person. For some, the routine hid a capacity much larger than the routine. For others, the routine was the capacity. No performance review has ever asked which, and no machine, as of 2026, can answer it on your behalf.

## 6.3 Predictable Creativity

The sentence needed one more beat. Rosario dela Cruz could feel it, the way a drummer feels a bar that ends too soon: the slight wrongness of a phrase that does everything it should and nothing more. She deleted three words, wrote four new ones, deleted two of those, and arrived at the version that breathed. The whole change took nine seconds. By any outside measure it was trivial. No client would notice. The number of people who clicked “buy” would not move. But Rosario noticed, and the noticing was what she had spent seven years training.

She writes marketing copy at a content agency in Makati: product descriptions, email campaigns, social posts, the words on shopping pages for online sellers across South-East Asia. She has an ear for rhythm, and for the different ways English is read in Manila, Kuala Lumpur and Singapore. She hears the Taglish, the mix of Tagalog and English, that her Manila readers actually speak. She has the discipline to cut a sentence in half when half is enough. She is good at this.

She would also admit, if you asked carefully and she trusted you, that the work follows patterns. A product description for a skincare brand has a structure: benefit, ingredient, proof, an invitation to buy. The execution varies. The architecture does not. An email subject line that works follows rules so well documented that they could be written down. The creativity is real. The range of possible results is not endless.

AI-written marketing copy in 2025 is not great. It misses what Rosario catches in her nine-second changes: the beat, the breath, the difference between words that function and words with someone behind them. But it is adequate. And for most of her clients, small online sellers with thin margins in crowded markets, adequate and fast beats good and expensive.

Clients who once sent the agency ten briefs a month now send three. They handle the rest with AI tools and the occasional freelancer to tidy up. The pattern is not Rosario’s alone. Researchers studied a large online marketplace for freelance work after ChatGPT was released. Freelancers in the most exposed occupations, writing above all, got about 2 per cent fewer jobs each month, and earned about 5.2 per cent less.**⁷** Those are averages, measured early. For the Philippines, whose outsourced service industries employed about 1.82 million people in 2024, many of them working in English, averages like these are not abstract.**⁸**

Rosario has been offered a new role: editing and choosing among AI-written drafts instead of writing her own. It pays about four-fifths of what she earns now. She has not accepted.

The money is clear enough. Her objection is that the role feels like a demotion dressed in new words, and Rosario has too much pride to pretend otherwise. She is a writer. The new job would make her an inspector of machine text, checking for errors the machine usually does not make, smoothing edges the client has already decided are smooth enough. Seven years of training her ear, and the market’s answer is quality control.

Chapter 5 said that when a machine can write a hundred drafts, the scarce act is to say “this one”, and to say why. Rosario’s offer is exactly that act, and it pays less than the work it replaces. Scarce is not the same as paid. This book will not pretend otherwise for her.

Then, one afternoon in May, editing a colleague’s batch, she found a line the machine had written for a coconut-oil brand, and it stopped her. It was not adequate. It was good. The beat was there, and the breath. She checked the file’s history to see whether a colleague had touched it. Nobody had. She read it again, and closed the file without telling anyone.

A different person might take the editing role and find something valuable in it: that taste matters more, not less, when everything around it is generated. Rosario cannot see that yet. Perhaps it is not true, or not true at a price she can live on. The pride that keeps her from the role is the same pride that made her good at the work. It is her best quality and her biggest obstacle, and this chapter will not settle what she has not settled herself.

**Predictable creativity** is work whose output takes skill but follows patterns established enough that a machine can approximate it: most copywriting, most design from templates, most routine illustration. It does not include work whose point is that nobody has seen it before. But the line between the two is drawn by the work, not by the job title, and Rosario’s afternoon showed that the machine sometimes crosses it. Whether anyone will pay for the difference between adequate and alive is not settled by taste. It is settled by markets, and Chapter 5 has said what markets do with things that are rare.

## 6.4 The Death of the Simple App

Tomas Villanueva built a grocery price tracker in eight months, with three developers in a shared office in Cebu City. It compared the prices of rice, cooking oil and tinned sardines across the stalls of Carbon Market and the supermarkets around it. By the end of 2024 it had forty thousand users. His mother used it. His landlord used it. He had left a call-centre job to build it.

What hurt him was not a better app. In early 2025 a cleaner competitor appeared, built by one developer with AI coding tools in a few weeks, and Tomas survived it. His users trusted his prices, because his team walked the market every Saturday to check them. What he did not survive so easily was the assistant. Shoppers began asking the AI assistant on their phones, in the same chat window where they talked to their families, where rice was cheapest that week. The assistant answered. Sometimes its answer used prices that Tomas’s team had collected on foot and published on his website. The shopper got the answer. Tomas did not get the visit. Over six months, the number of people opening his app each day fell by half.

This is the fifth form from 6.1, the path, at the size of one small business. Tomas was not out-built. He was routed around. For the shopper it was a gain: the answer came faster. For Tomas it was a loss. Nobody aimed it at him. But someone did decide that the assistant would answer from pages like his without sending the shopper to them. And some of what the assistant said had been his, gathered on Saturday mornings in a crowded market. There is a further turn. The assistant’s answer was good because Tomas’s team walked the market. If the visits that paid for the walking stop, the walking will stop too, and the answer will get worse. A path that routes around its sources can starve them. Whoever owns the assistant now owns the road between the shopper’s question and the stall. Whose the Saturday work is, once a machine repeats it, is a question for Chapter 20.

Chapter 5 told the story of Diane, whose speed stopped being an advantage when every rival had the same tools. Tomas’s story is different. The cost of building fell for him too, but that was not what nearly ended him. A simple app depends on being the place people go. When the question can be asked anywhere and answered there, the simple app loses its reason to be opened. What survives, as of 2026, is what the assistant does not reach: the Saturday walk through the market, and the stallholders who tell Tomas’s team the real price because they know them. They survive as things only his team can do. Whether anyone will pay for them once the visits have gone is a different question.

## 6.5 Thin AI Wrappers and Why They Die

In late 2023 Adaeze Nwosu raised four hundred thousand dollars in Lagos for an AI writing assistant tuned to Nigerian English. It knew the formality of a Lagos business letter and the warmth of a Yoruba greeting carried into English. It kept the idioms that a model trained mostly on American text flattened into something bland. She had the insight, the market and the timing. The product was a well-designed screen on top of another company’s model, with careful instructions that shaped its answers.

Within a year the model’s makers improved how it handled Nigerian English and let any user ask for a register in plain words. Two competitors offered the same thing for less. With every update she did not control, her product’s reason to exist grew smaller.

A **thin wrapper** is a product whose main value is access to something the underlying model already does. When the model improves, the wrapper is replaced by the layer it was built on. It is replacement from below. Chapter 14 returns to these businesses, and to what they can build instead. Here the point is diagnostic. Adaeze confused access to intelligence with owning something. Her insight about Nigerian English was real. It was not hers alone, and nothing protected it. What she knew lived in text, and text is what these models learn from. As of 2026, a difference that can be written down is a difference a model can be taught.

## 6.6 Organisational Middle Layers

Eleanor’s firm restructured in March 2025. She was not on the list. She was too junior to be worth cutting, and at first too junior to understand what the list meant. But she watched. Three senior associates and two junior partners were moved to “advisory roles” that everyone understood to be the waiting room for leaving. They were the people who occupied the space between the partners who brought in the business and the juniors who processed it. They gathered information, put it into shape, presented it upward and carried decisions downward. They were the **middle layer**.

AI can compress this layer. How far it has done so is not yet well measured. It does not replace good managers, who do far more than pass information along. They read a room. They know which associate is struggling before the billing records show it. They hold memory that no dashboard captures. But the passing along, the gathering, formatting, summarising and presenting, is exactly the kind of work a machine now does competently. The question for everyone in the middle layer is how much of their work is passing information along, and how much is something a person must do and answer for. The honest answer decides whether their role survives.

Eleanor watched the restructuring and felt two things she had not expected to feel together: sympathy and relief. Sympathy for the people being moved out. Relief, because the years of slow progress through the middle ranks that had always stood between her and the client’s table had suddenly grown shorter. The ladder was being compressed along with everything else.

But a shorter ladder has fewer rungs, and Chapter 4 said what rungs are for. Eleanor would reach the client’s table sooner than the people above her had, with less practice behind her. Whether she was ready to be in the room was a question nobody had asked, least of all the firm. The work of the middle manager who stays, the turn from controller to cultivator, belongs to Chapter 19. Here the point is structural. The layer is thinning, and it may also be changing what it carries: less information, more responsibility. The people who remain in it will be those whose value was never mainly in passing information along, but in being the one others can ask why.

## 6.7 The Messy Middle

The electric dynamo is the best guide we have to the years we are in. In the early 1880s the first central power stations opened in New York and London. By 1899, electric motors supplied less than 5 per cent of the mechanical power in American factories. Only by 1919, nearly four decades after those first stations, was just over half of it electric. And the large gains in factory productivity came in the 1920s, not the 1890s.**⁹**

Why so slow? The economic historian Paul David gave the answer in 1990. At first, factory owners did the obvious thing. They added electric motors to the old system of long shafts and belts, in the same tall buildings, often leaving the steam engines in place. Little changed. The gains came when engineers redesigned the factory around the new power. They put a small motor on each machine, built single-storey buildings, and arranged the machines by the flow of the work instead of by the line of the shaft. That took decades of learning by architects and engineers, and a generation of new buildings.**⁹**

The lesson is about where the time goes. The technology arrived quickly. The redesign of work around it was slow. And the years in between were confusing: new machines in old layouts, gains that did not appear, investments that looked like mistakes, and some that were.

Whether the AI transition will follow this pattern is an open question, and anyone who says they know the answer is selling something. The case that it will be faster: AI improves in months, and it arrives through software, not through new buildings. The case that it will be slower than it looks: organisations, contracts, laws and habits change at their own speed. Chapter 4’s three clocks, the machine’s, the institution’s and the slow clock of forming people, still run at different rates, and the gap between them is where the confusion lives.

What is already clear is that the in-between years will hold overshoot and undershoot at the same time. Some organisations will adopt too fast and pay for it. Others will adopt too slowly and lose ground. Some people will retrain for roles that have changed again by the time they qualify. Others will cling to roles that are changing under their feet. But too fast and too slow are always relative to someone. Sanjay’s hospital adopted fast. For the patients whose nodules were found, that was not too fast at all. For Sanjay, moved toward a room he chose his whole career to avoid, and for the three doctors who will not be hired, it may have been. Rosario’s agency was slow, and by the time it offered AI-assisted work, three competitors had taken its clients. That was too slow for the agency, and perhaps the right speed for the shoppers who now pay less. Klarna is the cleanest case, because when it changed course it named what it had lost: customers who wanted a person. When someone says a change came too fast, ask: too fast for whom?

In the messy middle, prediction is worth less than the ability to act before the direction is certain, and to answer for the action if the direction turns out to be wrong. A machine can become very good at the first half of that sentence. The second half is the Prologue’s second never, that a machine’s makers and users will never be free of what it does.

## 6.8 Consumer Work vs. Enterprise Work

Not all cognitive work faces the same timeline. A useful line runs between work that faces consumers and work that faces institutions.

Work that faces consumers lives where people tolerate imperfection and care a great deal about cost. The social post. The product description. The first reply to a customer’s complaint. If the machine’s output is eighty per cent as good at five per cent of the cost, the economics decide. Rosario’s clients did not leave because the machine was better than her. They left because it was cheap enough.

Work that faces institutions lives under different rules. When Sanjay’s hospital uses AI to read images, the tolerance for error is rightly low. When a drug company uses AI to analyse a clinical trial, a mistake means harm to patients, action by regulators and legal liability, not an unhappy customer. When a bank uses AI to decide who gets a loan, every error carries costs far larger than the money saved. Here the timeline is measured in years, not months.

A third kind of work answers neither to customers nor to clients, but to citizens: courts, welfare offices, tax departments, the police. Here an error is not only a cost. It can be a wrong done by the state, in the name of everyone. From 2016 the Australian government used an automated system to raise welfare debts by averaging people’s yearly income into fortnightly amounts. It was a fixed rule, not AI in today’s sense, and its lesson does not depend on how clever a system is. A class action ended in 2020 with a settlement covering about 430,000 of its debts. In 2023 a Royal Commission called the scheme “a crude and cruel mechanism, neither fair nor legal.”**¹⁰** A bank that makes a bad loan loses money. A state that wrongly accuses its citizens loses something it needs in order to govern at all: the belief that its decisions are **legitimate**, made by someone with the right to make them. This is where the Prologue’s third never applies with most force: being intelligent will never give a machine the right to decide for people. Knowing best does not give anyone that right either.

But slow is not never. The institutional world is not protected from these forces. It is delayed. As the systems become more reliable, as the rules adapt and as organisations gain experience, the room given to machines in high-stakes work will grow. The time this buys is real. It is best spent preparing, not resting.

## 6.9 The Uncomfortable List: Jobs You Thought Were Safe

You already know what is on this list. You may have known before you opened the chapter.

Start with what the evidence says, as of 2026. In 2025 the International Labour Organization estimated that one worker in four, worldwide, is in an occupation with some exposure to generative AI. In the most exposed group, it put 3.3 per cent of all jobs. Clerical work was the most exposed of all. And in that top group women were more exposed than men: 4.7 per cent of women’s jobs, against 2.4 per cent of men’s. The ILO’s own conclusion was that transforming jobs, not ending them, is the most likely effect.**¹¹** Chapter 4 recorded where the first damage has fallen: on young people entering exposed work, through hiring that does not happen. Chapter 5 recorded that in customer support the machine helped new workers most. The list below should be read with all of that in mind. It is a list of work under pressure, not of work that is ending. And read every name on it as Chapter 4 read the radiologist: as a bundle of tasks, some of which the machine takes and some of which it leaves.

**Radiologists.** Not disappearing. Chapter 4 told how a famous forecast that they would vanish failed. In its 2025 census, the Royal College of Radiologists counted a shortfall of about a third of the consultant radiologists the UK needs.**¹²** What is changing is the work: the machine finds more, the radiologist checks, connects and signs, and more of the job is talking to other doctors and to patients. Sanjay is not losing his job. He is losing the job he chose, and what remains is one he did not choose. Many redefinitions push the human part toward people: explaining, reassuring, persuading. That is not a safe harbour either. Chapter 2 recorded that health professionals, comparing written answers to patients’ questions, rated a chatbot’s answers as more empathetic than doctors’.

**Lawyers**, though not the advocates and advisers. The associates whose work lives in documents: research, review, analysis, first drafts. Eleanor is one of them.

**Accountants** in compliance roles. **Financial analysts** producing routine research. **Journalists** on predictable beats, such as company earnings, sports results and council meetings. The Associated Press began handing earnings reports to software in 2014, and said so in public. **Software engineers** writing routine code. **Copywriters, translators, paralegals, technical writers, junior consultants** and **customer-support agents**. The list reaches deep into work that takes real training and real skill.

**Teachers**, though this one is more complicated than a list allows. The part of teaching that delivers content is being supplemented by adaptive software. The part that is human formation is not. That part notices when a student is struggling, shows how a thoughtful person meets difficulty, and wakes a curiosity the student did not know they had. Most teaching jobs are both. The economics push toward automating the first, and that risks wearing away the conditions that support the second.

The list does not fall evenly, even on the three people of this chapter. Sanjay, senior and hard to replace, is being redefined. The exposure falls hardest on the junior and the freelance: on Eleanor’s review and on Rosario’s briefs. If you recognised your own work on the list, that recognition is information, not a verdict.

The question it raises is not whether AI will affect your work. It is what share of your work is pattern, structured and repeatable, and what share is the part that someone must choose and answer for. Not your knowledge. Not your speed. Not your qualifications. The part where a decision becomes yours, and you are the one who will be asked why. The answer is different for every reader. Only you can give it honestly. If the honest answer is that most of it is pattern, that is a fact about the job as it is arranged now, not about your worth. A large share that is yours to answer for would not guarantee your pay either.

There is a counter-argument worth taking seriously, because it is not wrong. Some economists argue that AI will increase demand for many of the jobs on this list. Cheaper imaging leads to more imaging. Faster legal research makes legal help affordable to more people. Chapter 5 described how cash machines changed bank tellers’ work without ending it, and why a machine can complement a person as well as substitute for her. The pattern has precedent.

Whether it holds this time is unknown. The honest answer is that the hopeful argument may be right about totals and wrong about individuals. Even if the number of radiologists grows, the work they do will be strange to Sanjay. Even if legal services expand, the junior associate’s role will not look like the one Eleanor trained for. The profession survives. The professional must become someone new.

### What would prove this wrong?

The first sentence at the head of this chapter is a claim about order: replacement takes tasks before jobs, and quietly before loudly. Beside it sits a claim about the fifth form, the path. Both describe 2026, and both can be checked.

Three findings would tell against them. If the exposed occupations began to shrink mainly through dismissals rather than through posts left unfilled, the loud form would have overtaken the quiet ones. The payroll study cited in 6.1 already separates hiring from separations, so its later updates are the place to look.**¹** If whole occupations shrank while the tasks inside them barely changed, jobs would be going before tasks. And suppose a fair study compared like searches, with and without an AI summary. If it found no fall in visits to the pages the summaries drew on, the fifth form would be smaller than I have said. The study in 6.1 did not make that comparison, and longer searches and questions drew summaries far more often.**⁶**

None of these is a forecast. Each is a measurement someone can make, and this chapter should be read against it.

## 6.10 Whose Work Is It?

Everything in this chapter has rested on one question: who does the work best? It is the market’s question, and in many cases it is also the right one. Ask any patient whose X-ray shows a nodule. She does not want her radiologist’s practice protected. She wants the spot found. One evening proves little either way. A fair test of who reads better counts many cases. It counts the shadows a machine flags that turn out to be nothing, as well as the ones it finds. It asks whether the machine reads every kind of patient as well. The hospital’s cheerful note counted only the finds. If the machine passes that test, and finds what Sanjay on a tired evening does not, it should look first. Sanjay should then answer for what is done with what it finds. That is delegation, as the Prologue described it, and it is good.

Hold that answer firmly, because there is an old voice that asks a different question. The market asks who performs the task best. This voice asks which path of action is yours. In the third chapter of the *Bhagavad Gītā*, Kṛṣṇa tells Arjuna:

> “One’s own duty, though defective, is better than another’s duty well performed. Death in (performing) one’s own duty is preferable; the (performance of the) duty of others is dangerous.”**¹³**

The Sanskrit begins *śreyān svadharmo viguṇaḥ*: better is one’s own *dharma*, though it lacks merit. *Svadharma* is one’s own duty, one’s own path of action. *Paradharma* is another’s.

Notice first to whom the verse is spoken. Arjuna does not want to fight the battle in front of him. He has already said it would be better “to live even on alms in this world”. Now he asks Kṛṣṇa why, if knowledge is higher than action, he is being pushed toward “(this) fearful action”.**¹³** The verse answers a man who wants to leave his own hard task for another path, one that looks cleaner and that he might perform well.

I must be more careful with this verse than with any other in this book so far. For centuries it was read as an order to stay in the work of one’s birth, and it was used to keep people there. Chapter 4 recalled how long the idea that work is fixed by birth lived in India’s law books and villages. I do not take that reading. I am applying the verse to a question it was not asked: what is lost when someone, or something, does your work better than you? Its second line, about death, belongs to the battlefield where it was spoken, and I leave it there. I use only the first.

Read that way, the verse lets us see a loss the market does not register. Work is not only output. It is also the path along which a person becomes who they are. Your own work, done imperfectly, does something for you that another’s perfect performance cannot do: it forms you. Rosario’s seven years formed her ear. Remove the imperfect doing, and you remove the forming, however good the replacement is.

But the verse’s own scene adds a cut that this chapter needs. Arjuna’s temptation was not to do his work badly. It was to leave it for a path with cleaner edges. Formation through work you will stand behind is one good. Shelter in work you chose because you always knew when you had finished it is another thing. Eleanor’s review was both, and the verse does not bless the shelter. Read in its scene, it would point her toward the room where her voice shakes, if that room is her own work. Whether it is, she does not yet know. The same cut applies to Sanjay. The dark room was his preference, not his path. The consultation room, which he is bad at, may be the work that is now his.

So the market’s question and the verse’s are different questions, and each matters in its own place. Where the work exists for someone else’s sake, the patient’s scan, the client’s contract, the shopper’s question, the market’s answer holds. The better performer should do it, even if the better performer is a machine, and a person should still answer for it. Where the work is also forming the person who does it, the verse lets us see what the market cannot: the loss is real, even when the machine does the task better.

Most work is both. That is why replacement is so hard to judge, and why the people inside it feel two things at once.

Notice also what the *Gītā* does not say, read this way. It does not say the patient must accept a worse reading so that Sanjay can go on being formed. It does not tell Rosario’s clients to pay more for copy they cannot tell apart. The forming of a person is a real good, but no customer owes it to the people who serve them.

Who does owe it? Sanjay’s hospital gives one answer by what it does. It uses the machine for the patient, rightly. And it will not fill three posts. That is the debt declined in advance: the patients get the better reading, and the doctors who would have learned beside Sanjay are never hired. There is a party to that decision that no chair in the room speaks for: the people who will need radiologists in twenty years. Chapter 4 called this the missing rung. Chapter 5 asked the plainer question of who gets the gain. This chapter adds that paying the debt need not mean keeping people doing work a machine does better. A hospital could use the machine for the patient and build the school separately. Young doctors could read cases first and then compare their reading with the machine’s, as Chapter 3’s practice asked of one person. Who pays for that school is a political question as much as a medical one. On this book’s reasoning, the hospital that saves three salaries takes part of the gain and owes part of the cost. A society need not preserve the work. It may need to preserve two things the work used to supply: the practice that forms people, and the authority that lets a person answer. A society that hands all the imperfect doing to machines, and keeps neither, will find in time that fewer of its people were ever formed by doing anything.

Four months after the machine arrived, Sanjay reads fewer images and reviews more. The machine flags; he checks, connects, decides and signs. He knows what signing costs, because in February he signed a correction to his own report. He also knows the danger Chapter 3 described. An eye that no longer reads first will weaken, and a signature from someone who can no longer check is a signature in name only. So each morning, before he looks at the machine’s flags, he reads ten images on his own. Nobody asked him to. He has been asked, first gently and then firmly, to spend part of each week in a consultation room, explaining findings to patients.

He is bad at it. Not careless, not unkind. But the consultation room asks for everything the dark room did not. It asks for steadiness in front of another person’s fear, and for the ability to stay with that fear without drowning in it or backing away. He chose a speciality that let him help from a distance. The distance has been closed, not by his choice, but by a machine that took the work he was comfortable with and left him the work he was not.

In June a woman of sixty-four came in with her son. That week the machine had found a small spot on her chest X-ray, and it needed a CT scan to say what it was. Sanjay explained it with numbers: its size, the guidelines, what usually happens next. Most spots of this size turn out to be nothing. Not all. He could not promise. She listened politely, and then asked the question his numbers had not answered. Her daughter’s wedding was in five weeks. Could the scan wait until after it?

It could, within what her own doctor thought safe, though Sanjay would have preferred it sooner. He said so. He said what he did not know. She chose to wait. It was her decision to make, and he wrote it in her notes in her own words.

Afterwards he sat in the dark room for ten minutes before he went back to the screens. He did not know whether he had done it well. He knew it was part of the work the machine had not taken, and that it was the part he had chosen radiology to avoid.

Eleanor’s firm finished its restructuring in the summer. Half the junior team was given new work. She was moved to a role facing clients, which the partners called an opportunity and which Eleanor experienced as exposure. One week in July she sat in on a meeting with a commercial property client. The partner leading the matter turned to her and asked for her view on an unclear clause.

She gave an answer. It was correct. Her voice shook.

She told a friend afterwards that the restructuring was exciting. She was lying. What she feels is closer to vertigo. The safety of the repetitive work was also its ceiling, and removing the ceiling has not freed her. It has shown her the height. She does not yet know whether her fear is the kind that fades with practice or the kind that means she is in the wrong place. She suspects the first. She is not certain.

Rosario has not accepted the editing role. She has not declined it either. She goes to work each morning and does the writing that remains, the briefs too complicated or too local for the machine, and she does them well. In the afternoons she reads the machine’s copy that her colleagues are editing. She sees every place where it is flat, and every place where flatness is all it will ever be. And now and then, a place where it is not.

### Practice

*Leave the hiding place.* Most jobs have a part with clear edges, where you always know whether you have done it. Name yours. Then, once this week, do the part without edges instead. Speak first in the meeting. Give your view on the open question. Call the client yourself. Say what you think before you are sure you are right, say what you cannot prove, and say what would change your mind. If your voice shakes, notice it, and do it again next week. The exercise will not fit every hiding place: Sanjay’s was a whole specialty, and Rosario’s obstacle is the pride that made her ear. It will not make the market pay, and it will not protect your job. It is practice in the sense of Chapter 4: the place where qualities are worn into a person, and the one place a machine cannot do your practising for you.

## Ending: The Coconut-Oil Line

In Makati, Rosario copied the coconut-oil line into a private file. She did not know why.

Three cities, three kinds of work, and none of the three has reached the other side. The chapter does not know where the other side is. Neither do they.

## Words to keep

- **compression:** the same work done by fewer people, each working with a machine.
- **legitimacy:** the belief that a decision was made by someone with the right to make it.
- **messy middle:** the confused years between a technology’s arrival and the redesign of work around it.
- **middle layer:** the people in an organisation who gather information, shape it, pass it up and carry decisions down.
- **overshoot and undershoot:** adopting a technology too fast and paying for it; adopting it too slowly and losing ground.
- **pre-emption:** replacement by a job that is never offered; nobody is fired, and nobody is hired.
- **predictable creativity:** skilled work whose output follows patterns established enough that a machine can approximate it.
- **redefinition:** the job keeps its title while its content changes.
- **substitution:** a task a person was paid for is done by a machine, and the person is no longer paid for it.
- **the path:** the road between the people who make something and the people who use it; a machine can replace the road without replacing the maker.
- **thin wrapper:** a product whose main value is access to something the underlying model already does.
- <strong><em>dharma</em></strong><strong>:</strong> duty; the right way of acting for a person in their situation.
- <strong><em>paradharma</em></strong><strong>:</strong> another’s duty or path of action.
- <strong><em>svadharma</em></strong><strong>:</strong> one’s own duty or path of action. In this book, applied to the work through which a person is formed.

## Sources

1. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Stanford Digital Economy Lab, “Canaries in the Coal Mine” update, 12 August 2026: the employment gap for workers aged 22–25 in the most AI-exposed occupations came “primarily through reduced hiring of young workers rather than increased separations”; see Chapter 4, 4.1.
2. Klarna, “Klarna AI assistant handles two-thirds of customer service chats in its first month,” press release, 27 February 2024; Charles Daly, “Klarna Slows AI-Driven Job Cuts With Call for Real People,” Bloomberg News, 8 May 2025; see Chapter 4, 4.1.
3. “How AP is using automation to increase story output tenfold,” Journalism.co.uk, 4 June 2015, reporting Lou Ferrara of the Associated Press: automated earnings reports went to the AP wire from October 2014; about 300 earnings stories a quarter before automation and about 3,700 after. Associated Press, “Automated earnings stories multiply,” January 2015: the change “freed up about 20 percent of the time that we had spread throughout the staff in producing earnings reports each quarter.” Benjamin Mullin, Poynter, 29 January 2015: “No jobs have been cut as a result of the automated earnings reports.”
4. Karl Marx, *Grundrisse: Foundations of the Critique of Political Economy* (written 1857–58; first published 1939–41), Notebook VII, trans. Martin Nicolaus (Harmondsworth: Penguin, 1973), pp. 705–706; online at marxists.org.
5. Fiona Jackson, “Duolingo to Replace Human Contractors With Bots, Become ‘AI-First’,” TechRepublic, 29 April 2025: Duolingo “cut 10% of its contractor workforce” (offboarded at the end of 2023 and reported in January 2024); Luis von Ahn’s memo: “gradually stop using contractors to do work that AI can handle.”
6. Athena Chapekis and Anna Lieb, “Google users are less likely to click on links when an AI summary appears in the results,” Pew Research Center, 22 July 2025: browsing data from 900 U.S. adults, March 2025; a traditional result was clicked in 8 per cent of visits with an AI summary and about 15 per cent without; a link inside the summary in 1 per cent. Longer searches drew a summary more often (53 per cent of searches of ten words or more, against 8 per cent of one- or two-word searches), as did searches beginning with question words such as “who”, “what”, “when” or “why” (60 per cent).
7. Xiang Hui, Oren Reshef and Luofeng Zhou, “The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidence from an Online Labor Market,” *Organization Science* 35, no. 6 (2024): 1977–1989: on Upwork, freelancers in the occupations most affected by ChatGPT saw a 2 per cent fall in monthly jobs and a 5.2 per cent fall in monthly earnings.
8. Louella Desiderio, “IT-BPM sector seen to hit $40 billion revenues this year,” *The Philippine Star*, 1 May 2025: the Philippine IT and business-process-management industry had 1.82 million employees in 2024 and revenue of $38 billion.
9. Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” *American Economic Review: Papers and Proceedings* 80, no. 2 (1990): 355–361. Warren D. Devine Jr., “From Shafts to Wires: Historical Perspective on Electrification,” *Journal of Economic History* 43, no. 2 (1983): 347–372: electric drive was “less than 5 percent of total capacity for driving machinery” in 1899, and in 1919 electric motors “represented over 53 percent”.
10. “Crude, cruel and unlawful: Robodebt Royal Commission findings,” *Law Society Journal* (NSW), lsj.com.au, 9 August 2023, on the *Report of the Royal Commission into the Robodebt Scheme* (7 July 2023): “a crude and cruel mechanism, neither fair nor legal”; income averaging of Australian Taxation Office data; the 2020 class-action settlement covering about 430,000 debts.
11. Paweł Gmyrek, Janine Berg, Karol Kamiński, Filip Konopczyński, Agnieszka Ładna, Balint Nafradi, Konrad Rosłaniec and Marek Troszyński, *Generative AI and Jobs: A Refined Global Index of Occupational Exposure*, ILO Working Paper 140 (Geneva: International Labour Organization, 20 May 2025).
12. The Royal College of Radiologists, *Clinical Radiology Census 2025*: a 32 per cent shortfall of consultant clinical radiologists in the UK, projected to reach 40 per cent by 2030 without action.
13. *Bhagavad Gītā* 3.35, trans. Kashinath Trimbak Telang, *The Bhagavadgîtâ with the Sanatsugâtîya and the Anugîtâ*, Sacred Books of the East 8 (Oxford: Clarendon Press, 1882), p. 56. Sanskrit: *śreyān svadharmo viguṇaḥ paradharmāt svanuṣṭhitāt | svadharme nidhanaṃ śreyaḥ paradharmo bhayāvahaḥ*. The thought returns at 18.47. Arjuna’s wish to withdraw: 2.5, “it is better to live even on alms in this world” (Telang, p. 43); 3.1, “then why, O Kesava! do you prompt me to (this) fearful action?”

*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 a chest CT nodule was “partially obscured by a rib”; that radiologists are “contracting”; that automated journalism was “a fact the industry has been reluctant to discuss publicly”; “Makati, Manila”; that the automobile took “roughly fifty years”, digital communication “three decades”, cloud “more than fifteen years” and mobile “nearly a decade” (unsourced); that Priya’s “team of twelve becomes a team of three”; that Tomas lost to a competitor built “in three weeks”; that “the market, eventually, will” tell Rosario’s nine-second changes from generic copy; that adapting to the messy middle “requires something a machine does not possess: the awareness that the ground is moving”; that Sanjay’s missing skills are “qualities of consciousness”; and “Here is the thing about Eleanor.”
- *What I now hold:* replacement comes in five forms, not four: pre-emption, compression, redefinition, substitution, and the replacement of the path between makers and users. The X-ray, not the CT, hides a nodule behind a rib. Radiology is a case of redefinition, not of contraction, and Chapter 4 tells why. The AP automated earnings stories in public in 2014 and cut no jobs over it. The UK’s radiologist shortfall is about a third (RCR, 2025). Marx’s watchman describes redefinition, and his “for all of them” asks Chapter 5’s question about the gain. The dynamo shows why the in-between years are confused: the work has to be redesigned around the technology. Tomas is routed around, not out-built, so his story no longer repeats Diane’s. Scarce is not paid: Rosario’s editing offer pays less, and the machine sometimes writes a line she admires. Where work serves another, the better performer should do it; where it also forms the worker, its loss is real (Gītā 3.35, applied, with its history of use for caste named and refused). Read in its own scene, the verse is spoken to a man tempted to leave his hard task, so it does not bless a hiding place. Sanjay answers for his miss with an addendum and a telephone call. Too fast and too slow are relative to someone. Public work answers to citizens, and its risk is legitimacy (Robodebt). A society need not preserve the work; it may need to preserve the practice that forms people and the authority that lets a person answer. A fair test of a machine counts its false alarms and asks whether it serves every kind of patient. A loss that nobody aimed at a person was still someone’s decision. The hospital that takes the gain owes part of the cost of the school.
- *What changed my mind:* the audit’s corrections; the Robodebt Royal Commission; Chapter 4’s radiology evidence; the Pew study of AI summaries; the AP’s own account; the Upwork study; the ILO’s index; Paul David’s dynamo; the *Grundrisse*; the *Gītā*’s *svadharma*, read against the market.
- *What survives:* the four forms of 6.1, now five; the messy middle, without its forecast of “five years” of chaos; Eleanor’s hiding place, and “She gave an answer. It was correct. Her voice shook.”; Rosario’s nine seconds and her pride; the messy middle and its overshoot and undershoot; the uncomfortable list and “right about totals and wrong about individuals”; Sanjay, who chose the dark room.

*Next in the series:* [*Chapter 7: What AI Cannot Replace*](https://nithinpalal.com/ai/ai-proof/chapter-7-what-ai-cannot-replace)*.*