---
title: Jobs That Will Survive the Age of AI
url: https://oosioo.com/en/p/ai-%EC%8B%9C%EB%8C%80%EC%97%90-%EC%82%B4%EC%95%84%EB%82%A8%EC%9D%84-%EC%A7%81%EC%97%85
date: 2026-11-06T12:42:54+00:00
author: SYSOP
summary: All my notebook said was “Jobs that will survive the age of AI.” This piece starts with former colleagues who say they have nothing left to do after handing all their work to agents. It then uses company cases and 2026 research to look at what is actually being replaced right now. Next, it follows the market that sells anxiety and the prophecy that “machines will take our work,” repeated for 200 years, from the Luddites through telephone operators, ATMs, spreadsheets and radiology, and shows how each turned out. It also looks at how agents scored in tests that handed them whole pieces of real work. The conclusion: tasks change hands, not jobs. People who go into the field and solve problems with AI, and people who make new things, are still far better at it than AI. Also included: a playground for breaking your job into tasks and estimating the agent's share, 14 jobs that were supposed to disappear with the predictions and what actually happened, and a record of the new work that emerged. The thirty-first installment of A Page for Today.
---
# Jobs That Will Survive the Age of AI

![On the left of a light-gray background are “A Page for Today · 31,” the title “Jobs That Will Survive the Age of AI,” and the line “The prophecy that machines will take our work has been repeated for 200 years. What's different this time, and what's the same?” On the right, three black-and-white photos (telephone operators in 1945, a human computer at a calculator in 1952, and a typing pool in 1918) overlap with bar cards that split a week of work for a customer service agent and a field AI engineer into what goes to the agent and what stays mine.](/uploads/29aeba2fc711c26b.webp =720)

The notebook I opened today had just this one short line:

> Jobs that will survive the age of AI.

These days, when I meet friends I used to work with, I keep hearing the same kind of story. They've handed most of their work over to AI agents, and now they have nothing to do. The agent writes the report drafts, the meeting minutes organize themselves, and the agent writes the code and even runs the tests. The human skims the output and presses the approve button. At first it sounded like bragging, but after hearing it a few times I noticed the anxiety at the end of it. If I have this little to do, how much longer will the company keep me?

The articles and videos say the same thing. Every few months a new model arrives that is supposedly almost at AGI, and each time a list of jobs about to disappear comes with it. More people are also selling that anxiety as a product. Courses and e-books that open with “Before AI takes your job” fill my feed without end.

But what I see where I work looks a little different. Bringing in AI to solve real problems still only works with people involved. People are also still far better at creating something truly new. Today I want to use numbers and cases to fill in the space between these two views.

## The People Who Ran Out of Work

My friends aren't the only ones. Anthropic publishes a report called the “Economic Index,” which uses conversation records to show what Claude is actually used for. In the first report, in February 2025, “augmentation,” where people work with AI through back-and-forth, made up 57%, and “automation,” where they hand off the work entirely, made up 43%. In that September's report, directive conversations of the “hand it off and be done” kind rose to 39%, and automation overtook augmentation for the first time. Among conversations from businesses using Claude through the API, 77% followed automation patterns. In the January 2026 report, directive use fell back to 32%. People keep handing work off, pulling back, and handing it off again.

The companies' declarations are bolder. In February 2024, the Swedish fintech company Klarna announced that its AI assistant had handled 2.3 million conversations in its first month, two-thirds of all chats, doing the work of 700 agents. In April 2025, Shopify CEO Tobi Lütke circulated a memo telling teams to show why AI couldn't do a job before asking for more people. That same month Duolingo declared itself “AI-first” and said it would use AI for work that contractors had been doing. Salesforce's Marc Benioff said in September 2025 that customer support staff had been cut from about 9,000 to 5,000, explaining, “I need less heads.” In February 2026, Block's Jack Dorsey cut more than 4,000 employees, about 40% of staff, and said the reason was AI, not trouble at the company. The stock reportedly rose as much as 24% that day.

But we also need to look at what these same companies did next. Klarna's 700 was not a layoff count. It was the workload expressed as a number of people. And in May 2025, CEO Siemiatkowski told Bloomberg that cost had been too dominant a factor and the result was lower quality, and the company began hiring human agents again. Duolingo, facing a growing backlash, clarified that no full-time employees had been laid off. IBM CEO Arvind Krishna said in May 2025 that AI agents had taken over the HR work of several hundred people. Total headcount had actually grown, because the savings went into hiring developers and salespeople.

## Two Sets of Signals

So if you ask what is actually happening right now, you get two answers. Depending on which numbers you pick, you can write opposite stories.

![Under the title “Two Sets of Signals” are two columns of number cards. The left column, “Replacement is happening,” shows: employment of 22–25-year-olds in AI-exposed occupations −19% (Stanford, 2026.8); 94% of the 285,000 jobs lost by 15–29-year-olds came from highly AI-exposed industries (Bank of Korea, 2026.8); full-time entry-level postings at large companies −43% (Jinhaksa Catch); AI cited in 24% of U.S. job cuts in January–July 2026 (Challenger). The right column, “People are still needed,” shows: AI's effect on economy-wide employment and real wages indistinguishable from zero (Yale Budget Lab); 55% of organizations that cut staff because of AI say it was the wrong decision (Orgvue); AI cited in 4.5% of U.S. job cuts in January–November 2025 (Oxford Economics); FDE job postings up from 643 to 5,330 (Indeed).](/uploads/cdcdefede7a96707.webp =720)

Start with replacement. The most-cited study is the Stanford Digital Economy Lab's “Canaries in the Coal Mine.” It tracked employment by occupation using payroll data from ADP, the largest payroll company in the U.S. In occupations highly exposed to AI, employment of 22- to 25-year-olds fell well below the trend in less-exposed occupations. The gap was 13% in the original August 2025 version. It widened to 19% in the August 2026 revision, which added data through June 2026. Experienced workers showed no such gap, and the drop came from fewer entry-level hires, not from layoffs. It was concentrated in occupations where AI takes over the work entirely rather than ones where it helps.

Korea is heading the same way. According to press reports, an August 2026 issue note from the Bank of Korea found that jobs held by 15- to 29-year-olds fell by 285,000 between June 2022 and June 2026, and 94% of the drop came from highly AI-exposed industries. Over the same period, jobs held by people in their 50s grew by 230,000. Youth jobs fell 31.4% in information services and 27.4% in publishing. The Bank of Korea, however, read this less as a change caused by AI alone and more as AI speeding up an existing preference for experienced hires. By the count of the hiring platform Jinhaksa Catch, full-time entry-level postings at large companies fell 43%, from 3,741 in January–November 2025 to 2,145 over the same period in 2026. In IT and telecom they fell 67%.

The numbers on the other side are just as strong. Since ChatGPT came out, Yale's Budget Lab has run repeated analyses of occupational mix, AI usage, and difference-in-differences. Even in its October 2026 update, it sees no clear sign of AI in the economy as a whole. In its May analysis, the effect on employment and real wages could not be told apart from zero. Layoff statistics can also be read in different ways. By the count of the U.S. outplacement firm Challenger, companies named AI as the reason for about 24% of layoffs in January–July 2026, and AI was the top reason for five months in a row. But these are the reasons companies themselves give. Oxford Economics pointed out that layoffs blamed on AI made up 4.5% of the total in January–November 2025, less than a quarter of those blamed on the economy or market conditions. It also noted that if AI were really replacing people in large numbers, productivity growth should be speeding up, but it had actually slowed. Saying “because of AI” sounds better to investors than saying demand fell or you hired too many people during the pandemic. This is called “AI washing.” Even OpenAI's Sam Altman admitted in India in February 2026 that some companies blame AI for layoffs they would have made anyway.

Companies are also going back on these cuts. When the consulting firm Orgvue surveyed 1,000 executives in September 2025, 55% of organizations that had cut staff because of AI said it was the wrong decision, and 32% had rehired.

Put the two sides together and the picture gets a little clearer. The economy as a whole isn't collapsing. But the first rung that people just entering the workforce used to step onto really is getting narrower. I don't want to brush past this second fact. I'll come back to it at the end.

## The People Selling Anxiety

When anxiety runs high, a market grows up to sell it. The loudest voices come from the AI companies themselves. In a May 2025 interview with the U.S. outlet Axios, Anthropic's Dario Amodei warned that AI could wipe out half of entry-level white-collar jobs within one to five years and push unemployment to 10–20%. The remark is often shortened to “half of white-collar jobs will vanish within five years,” but the original sentence said “entry-level,” “one to five years,” and “could.” A month later in Paris, Nvidia's Jensen Huang shot back that he “disagreed with almost all of it.” At the same event in May 2024, Elon Musk had said that “probably none of us will have a job.” The prediction has no date, so there's no way to check whether it came true. A year later, in May 2026, Amodei shared a stage in New York with JPMorgan's Jamie Dimon and took a different line: if 90% of work is automated, everyone ends up doing the remaining 10%, and that 10% could become the whole of their job.

When a company that builds a technology says that technology is frightening, it's worth thinking twice. Saying the threat is huge is also a way of advertising how powerful the technology is.

One level down is a market that turns anxiety into courses. It usually works the same way. First, it scares you by saying AI will take your job. Next, it narrows the solution down to one thing: people who use AI will replace people who don't. Finally, it sells that solution as paid courses, e-books, and one-on-one coaching. The same people sell both the fear and the answer. Requests for relief over “online side-hustle courses” filed with the Korea Consumer Agency nearly quadrupled, from 11 in 2024 to 42 in 2025. AI courses aren't counted separately, but 89.8% of the payments were between 1 million and 4 million won, and only 33.9% of cases were resolved. Korea's Fair Trade Commission is taking public comments until November 11 on draft labeling and advertising guidelines. The draft would define ads that play up AI when there is little or no AI, or when AI only plays a supporting role, as “AI washing.”

## A Prophecy Repeated for 200 Years

The idea that machines will take people's work isn't new. In fact, it's one of the oldest stories of industrial society.

![A colored print from 1812. A figure in a skirt and hat stretches out one arm and points ahead, while behind them buildings burn and people gather in crowds. The caption below reads “THE LEADER OF THE LUDDITES.”](/uploads/9afffabf3b762b5d.jpg =560)

Starting in 1811, textile workers in England's Midlands gathered night after night to smash stocking frames and looms. They were called Luddites, after an imaginary leader named “Ned Ludd.” They are often remembered as people who hated machines, but they chose the machines of particular employers who cut wages or hired cheap apprentices. They were fighting wage cuts made in the name of machines more than the machines themselves. In 1812, Parliament changed the law so that breaking frames could be punished by death, and the next year several were executed after trials at York. Individual skilled workers really were ruined. The machines kept coming, and the textile industry grew larger.

In 1930, John Maynard Keynes used the phrase “technological unemployment” in “Economic Possibilities for Our Grandchildren.” It's unemployment caused when technology saves labor faster than new uses for labor can be found. The essay is often quoted as a warning, but Keynes saw this as a temporary maladjustment. He predicted that a hundred years later, the grandchildren's generation would live comfortably on a 15-hour work week. Living standards have mostly risen as predicted, but working hours haven't fallen as much as expected. There are four years left until 2030.

In March 1964, a memo called “The Triple Revolution,” signed by the Nobel laureate Linus Pauling and others, was delivered to President Lyndon Johnson. It argued that computers and automation would cause mass unemployment, so incomes should be guaranteed. Johnson set up a commission on automation that August. In its 1966 report, the commission concluded that technology eliminates jobs, not work.

![On the left, a black-and-white photo from the 1945 Potsdam Conference shows U.S. Women's Army Corps switchboard operators wearing headsets, seated in a row at a long switchboard and plugging in cables. On the right, a black-and-white photo from the 1918 flu epidemic shows a masked elevator operator standing at an elevator door.](/uploads/807878f8783e7473.jpg =720)

Some jobs really did disappear. In the early 1920s, AT&T was America's largest employer, and more than half its employees were telephone operators. As switching became automatic, the work shrank. AT&T's manual switching ended in 1978 on Santa Catalina Island, California. According to a 2024 study by the economists Feigenbaum and Gross, the operators working at the time were really hurt. The older they were, the less likely they were to find other work. But employment didn't fall for the next generation of young women. They moved into newly growing jobs such as secretaries and clerks. The job disappeared and the people went elsewhere. The generation caught in between paid the price.

The economist James Bessen tracked the 270 occupations listed in the 1950 U.S. Census. Only one of them disappeared completely because of automation: elevator operator. In the others, machines took over part of the work, but the occupations survived in a different form.

Bank tellers are a popular example, but the ending is a little different from the usual version. ATMs spread from the 1970s on. According to Bessen's analysis, though, between 1988 and 2004 the number of tellers per urban branch in the U.S. fell from 20 to 13. Meanwhile the number of branches rose 43%, so the total number of tellers stayed about the same. Branches became cheaper to open, so banks opened more of them, and tellers went from counting money to selling products. That's the version most people hear. U.S. teller jobs peaked at about 600,000 in 2007, then fell as mobile banking spread. By 2025 there were 339,000, according to the Bureau of Labor Statistics, and that number is projected to fall another 13% over the next ten years. ATMs didn't reduce the number of tellers, but phones did. Two technologies that do the same work can lead to different results, depending on how the work is reorganized around them.

![A black-and-white photo of the Navy Department typing pool in Washington, D.C., 1918–1919. Rows of desks fill a large office, and dozens of men and women sit at typewriters and piles of papers, looking at the camera.](/uploads/d3fdff813f4e4f33.jpg =720)

Offices went through the same thing. A hundred years ago, offices had large typing pools like this one. In the U.S., there were 37,200 word processors and typists left in 2023. When the spreadsheet program VisiCalc came out in 1979, people said clerks who worked out the books by hand were finished. By NPR's *Planet Money* count in 2015, the number of bookkeeping and accounting clerks had fallen by about 400,000 since 1980, while the number of accountants had grown by about 600,000. Once calculation was cheap, people did more work asking “What if we changed this?” and weighing the answers.

The number that launched the AI era is worth another look too. In 2013, Frey and Osborne of Oxford University reported that 47% of U.S. jobs were at high risk of automation. The figure is often quoted as “half of all jobs will disappear in 10 to 20 years.” In fact, it was the share of employment in occupations with a high chance of automation, and it gave no date for when those jobs would disappear. In 2016, the OECD measured again, this time by the tasks within jobs rather than whole occupations. It found an average of 9% across 21 countries, with Korea lowest at 6%. Since then, U.S. nonfarm employment has actually grown.

## What Happened to the Radiologists

The most famous prediction on the AI side came from Geoffrey Hinton, known as the father of deep learning.

![Under the title “After ‘Stop Training Radiologists’” is what Hinton said in Toronto in November 2016: “We should stop training radiologists right now. It's clear that within five years deep learning will do better than them.” Three cards below show Mayo Clinic's radiology staff growing from about 260 to more than 400, +55% (New York Times, 2025.5); a record 1,208 U.S. radiology residency positions in 2025 (Works in Progress, 2025.9); and average income of about $520,000, +48% since 2015.](/uploads/cbcd7a8199c5f3a7.webp =720)

At an event in Toronto in November 2016, Hinton said we should stop training radiologists right now, because deep learning would clearly do better than them within five years. Reading images to find problems looked like exactly the kind of thing deep learning did best. And AI that helps read scans did come into wide use afterward.

The opposite happened. According to the New York Times in May 2025, Mayo Clinic's radiology staff grew from about 260 to more than 400. U.S. radiology residency positions hit a record 1,208 in 2025, and average pay was about $520,000, up more than 48% from 2015. Hinton also took the claim back, saying it had been too broad. Reading images was only part of what radiologists do. The job also includes making the final call on unclear findings, doing biopsies and procedures, working with doctors in other specialties, and explaining results to patients. On top of that, the number of scans went up.

Go tells a similar story. After AlphaGo beat Lee Sedol 9-dan in 2016, many people said Go played by humans would lose its meaning. A 2023 study by Minkyu Shin and colleagues in the *Proceedings of the National Academy of Sciences* (PNAS) analyzed more than 5.8 million moves by professional players from 1950 to 2021. After superhuman AI appeared, the quality of human moves got better, and new moves that had never been played before became more common. AI didn't replace the players. It became something they studied with.

Not every story ends this way, though. Translation is a case we need to look at honestly. The number of translators in the U.S. was 75,300 in 2024. It hasn't fallen, and it's projected to grow a little. But a 2025 study by Frey and colleagues estimated that without machine translation, the U.S. would have about 28,000 more translator jobs. In Korea, there have been reports of clients sending AI translations along with requests to cut fees by 10–30%, and sometimes by more than half. The job still exists, but the work pays less. Surviving and living as you did before are not the same thing.

## When Agents Were Given Real Work

So how much real work can today's agents do? Several tests have come out recently that copy real work as it is, instead of using exam questions.

![A horizontal bar chart titled “When Agents Were Given Real Work” compares how often agents finished the job in tests that copy real work: freelance jobs handed over as-is (Remote Labor Index, 2025.10) 2.5%; household chores given to robots (BEHAVIOR-1K) 12.4%; office work at a simulated company (TheAgentCompany, revised version) 30.3%; expert tasks compared one at a time (GDPval, 2025.9) 47.6%. A note below explains that GDPval uses a different standard: the share of outputs judged as good as or better than a human expert's work.](/uploads/dd4cd6f2384d5767.webp =720)

In October 2025, Scale AI and the Center for AI Safety (CAIS) collected 240 jobs that had really been posted on freelance platforms and built the “Remote Labor Index.” These were jobs people had been paid to do, such as game development, product design, architectural drawings, and data analysis. The best agent produced work a client would accept 2.5% of the time, and no model got above 3%. There have since been reports that follow-up releases reached the mid-teens, but the figures differ from source to source and I haven't been able to confirm them.

TheAgentCompany, built by researchers at Carnegie Mellon University, sets up a fictional software company and has agents work there as employees. Its 175 tasks include asking coworkers questions on the company chat, finding documents, fixing code, and scheduling meetings. Agents fully completed 24.0% of tasks in the first version and 30.3% in the revised one. OpenAI's GDPval, released in September 2025, covers 1,320 tasks from 44 occupations. Experts compared model outputs with the work of human professionals without knowing which was which. The best model was judged as good as or better than the experts 47.6% of the time. The longer and more complex the task, the lower its win rate.

Physical work is even harder. Stanford's BEHAVIOR-1K tests robots on 1,000 household chores, and the top team in the 2025 competition succeeded 12.4% of the time. “Moravec's paradox” says that chess and calculus, which are hard for people, are easy for machines, while walking and picking things up, which are easy for people, are hard for machines. Forty years later it still holds. Waymo, which runs driverless taxis, keeps about 70 remote assistance agents on duty to give advice when a car can't decide what to do. This came out in February 2026, after a U.S. senator asked about it. The agents don't take the wheel. They tell stuck cars which way to go.

AI didn't always help even people who were good at their jobs. In a randomized controlled trial published in July 2025 by the nonprofit research institute METR, 16 experienced developers worked on open-source projects they had worked on for years. With AI tools, they actually took 19% longer. They themselves believed they had been 20% faster. A February 2026 follow-up pointed toward a speedup, but METR itself said the result was hard to trust. Developers who didn't want to work without AI had dropped out of the study, which skewed the sample. In a July 2025 report, MIT researchers said companies had spent $30–40 billion on generative AI, but 95% saw no measurable effect on profit and loss. The survey has limits, with a small sample and self-reported answers, but it isn't far from what I see in the field.

Agents are clearly getting better fast. The numbers jump several times over within a year. Still, these tests all show the same thing. Agents do well on short tasks with clear boundaries. They score worse when the work is longer, less clear, and more tied up with real workplaces and people.

## Tasks Change Hands, Not Jobs

This is where something economists have long said starts to make sense. AI doesn't replace whole jobs. It takes some of the tasks inside them. In the 2024 paper “The Simple Macroeconomics of AI,” MIT's Daron Acemoglu used this kind of task-by-task math. He estimated that AI would raise U.S. total factor productivity by only about 0.66% over the next ten years. AI can do fewer tasks than people think, and even fewer of those are worth automating to save money. That number is a long way from the hype.

So instead of asking whether a job title will survive, I thought it would be more useful to break my week into tasks. The playground's first tab splits a 40-hour week for ten jobs into six tasks each. It scores each task from 0 to 3 on six traits. Repetitive work with clear rules, and work that is done entirely on a screen, are easier to hand to an agent. Work that needs physical effort in the field, judgment calls you're responsible for, problems you've never seen before, and work that depends on people's trust are what humans keep. I chose the scores and weights myself, so this is a way to organize your thinking, not a forecast. You can click the bars to change the scores or add your own tasks.

```map /embed/ai-jobs-2026-en.html 3:4
Split my job: break the week of ten jobs into tasks and estimate the agent's share and what remains, with agent ability ranging from 2023 chatbots to far stronger hypothetical agents (scores and rule are mine; your own tasks are saved only in this browser) / Jobs that were meant to vanish: 14 predictions from the Luddites to Amodei and what actually happened / Work that appeared: work that arose where machines arrived, where AI goes looking for people, and new things people made with AI
```

Run it with today's agents, and a call center agent hands over almost all the time spent answering common questions, entering orders and refunds, and writing up call notes. That means the AI agent takes close to half the week. Turn the setting up to “a much stronger agent” and the share reaches 56%. Even then, calming angry customers, deciding on exceptions to the rules, and finding and reporting the causes of repeated complaints stay with the human. On the same “2026 · Now” setting, the share is 38% for translators, 32% for accountants, 26% for software developers and radiologists, and 21% for pharmacists and elementary school teachers. I know my math is rough. Still, jobs differ more clearly in what work remains than in how much gets handed over.

The jobs that hand over the least are electricians, who do physical work on-site (12%), and the “field AI engineer” I added last (16%). That's true for the field AI engineer even though the job is done at a screen. They share data cleanup and coding with the agent. But most of their week goes to visiting sites to see how people really work, deciding which problems to tackle, finding out why a pilot failed, and convincing and training the people there. Building AI and bringing AI into a workplace to solve problems are two different jobs.

## New Work Always Shows Up Later

![A 1952 publicity photo taken at the NACA facility in Muroc, California. At a desk by a window with the blinds down, a woman looks through the eyepiece of a measuring instrument and writes numbers on paper, with a large Friden mechanical calculator beside her.](/uploads/5bfdff6a9e840c3b.jpg =720)

In the 1930s and '40s, the U.S. National Advisory Committee for Aeronautics (NACA) had people whose job title was “computer.” They were women who worked out wind tunnel data by hand and on mechanical calculators. The number of women computers at the Langley laboratory grew from 5 in 1935 to about 400 in 1946. Today they're a classic example of a job that disappeared, but back then it was a new job. When electronic computers arrived, the job title became the name of a machine. Dorothy Vaughan, who led a team of computers, saw that IBM machines were coming, learned FORTRAN, and became a programmer along with the team. When John Glenn orbited the Earth in 1962, Katherine Johnson checked the orbital calculations from the IBM 7090 by hand.

![Under the title “New Work Always Shows Up Later,” a donut chart on the left shows that about 60% of U.S. jobs in 2018 were in kinds of work not on the 1940 Census list of occupations (Autor et al., QJE 2024). On the right, cards list: about 400 women computers at NACA Langley in 1946; about 400,000 fewer bookkeeping clerks and about 600,000 more accountants, 1980–2015; forward deployed engineer postings up from 643 to 5,330, 2025–2026; Google electrical apprenticeship spots up from 19,500 to 30,000 a year, 2026; about 70 Waymo remote assistance agents on duty, 2026.](/uploads/6235d56db4a6ff59.webp =720)

Disappearing jobs are easy to see because they have names. New work doesn't show up in the statistics until it gets a name. MIT's David Autor and colleagues went through the record of new job titles added to the U.S. Census list of occupations. In 2024 they reported that about 60% of U.S. jobs in 2018 were in kinds of work that didn't exist in 1940. Read the names of more than half of today's jobs to someone from 80 years ago, and they wouldn't know what those jobs were.

New jobs of the AI era are already getting names. The best known is “forward deployed engineer” (FDE). It started at Palantir, which sent engineers to client companies to adapt its software on-site. Now OpenAI and Anthropic are competing to hire them. FDE postings on the job site Indeed grew from 643 in April 2025 to 5,330 in April 2026. In Korea, LG CNS set up a new FDE team, and in September 2026 Samsung SDS announced that it would hire 200 FDEs by the end of the year. The companies that build the models say they are shortest on people who can go into customers' workplaces and solve problems with those models.

AI is creating new physical work, too. Building data centers and connecting them to power takes electricians. In March 2026, Microsoft President Brad Smith called the shortage of electrical workers the number one problem holding back data center expansion in the U.S. Google is putting $50 million toward expanding apprenticeships in the U.S. electrical workers' union from 19,500 to 30,000 spots a year.

People and AI working together are also producing new things. The 2024 Nobel Prize in Chemistry went to DeepMind's Demis Hassabis and John Jumper for predicting protein structures with AlphaFold. More than 2 million researchers in 190 countries use the roughly 200 million protein structures AlphaFold has predicted. In August 2025, James Collins's team at MIT used generative AI to create and screen 36 million candidate compounds. They found possible treatments for gonorrhea and MRSA that are built completely differently from existing antibiotics, and showed that they worked in mice. AI produced the candidates, but people decided what to look for, chose which ones to keep, and tested them in the lab.

Small teams can also do much bigger things than before. In October 2024, Maor Shlomo of Israel started building the vibe-coding service Base44 single-handedly. It reached 250,000 users in six months and was sold to Wix for $80 million in June 2025, when it had six employees. Sweden's Lovable, according to the company, passed $100 million in annual recurring revenue with 45 full-time employees. Examples can be found closer to home, too. According to press reports, a pilot apple orchard complex in Andong, North Gyeongsang Province, installed IT traps that count pests automatically and send alerts. Its pest damage rate fell from 16% in 2021 to 1.6% in 2025. The orchard workers were the ones who decided which trees to hang the traps on and what to do when an alert came in.

## Instead of “Surviving”

Back to the line in my notebook: jobs that will survive the age of AI. Based on the numbers and cases I've looked at today, it's hard to answer with job titles. Some jobs that looked likely to disappear first, like radiologist, have grown. Others, like translator, still exist but pay less. Some, like bank teller, held up against the first technology and then shrank with the second. The tasks I actually do in a job say more than the job title.

Once you break a job into tasks, it's fairly clear which parts get handed over and which parts stay. Work that is done on a screen, follows clear rules, and already has a right answer gets handed over. When my friends said they had run out of work, it was probably because this kind of work had been a big part of their jobs. What stays is going into the field to find out what the real problem is, making decisions you're responsible for, solving problems you haven't seen before, and convincing people. These were also the areas where agents failed most often in the tests that gave them real work.

So this is what I think. If you can become someone who uses AI to solve problems in the field, or someone who makes something that didn't exist before, you don't need to worry much, even as the world changes. Neither kind of person avoids AI. Both are the people who use it the most. Because agents take the repetitive work, they get to spend more of their day in the field and on new problems. That's also what I want to tell my friends who say they've run out of work. Your work hasn't disappeared. You now have time to move on to the next stage of it. What you do with that time will make the difference over the next few years.

One thing still bothers me, though. As the Stanford and Bank of Korea numbers show, the first thing getting narrower right now is the first rung for people just starting out. People usually learned to solve problems in the field by doing repetitive work there. If agents take that repetitive work, where will new hires learn the field? As in the era of telephone operators, the generation caught in between may once again pay the price. Individuals can't solve this alone. It takes companies willing to give new hires real work in the field and a society that creates those jobs. It's also worth noting that the courses selling anxiety never mention this part.

*A Page for Today* is a series where I open my idea notebook to a random page and write about the line I find there. Today, the line “jobs that will survive the age of AI” became the thirty-first installment. If you spot a mistake in the numbers or examples, please let me know and I'll fix it.

## References

- Erik Brynjolfsson et al., “[Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine/),” Stanford Digital Economy Lab (2025, revised 2026.8)
- Yale Budget Lab, “[Tracking the Impact of AI on the Labor Market](https://budgetlab.yale.edu/research/tracking-impact-ai-labor-market)” (updated 2026.10)
- Anthropic, “[Anthropic Economic Index: January 2026 Report](https://www.anthropic.com/research/anthropic-economic-index-january-2026-report)”
- Challenger, Gray & Christmas, “[July 2026 Job Cuts Report](https://www.challengergray.com/wp-content/uploads/2026/08/Challenger-Report-July-2026.pdf)”
- Fortune, “[AI layoffs are a convenient corporate fiction, Oxford Economics says](https://fortune.com/2026/01/07/ai-layoffs-convenient-corporate-fiction-true-false-oxford-economics-productivity/)” (2026.1)
- Orgvue, “[New research exposes the complexity of deploying AI systems in the workforce](https://www.orgvue.com/news/new-research-exposes-the-complexity-of-deploying-ai-systemsin-the-workforce/)” (2025)
- SBS Biz, report on the Bank of Korea's findings on youth employment and AI-exposed industries ([link](https://biz.sbs.co.kr/amp/article/20000329179), 2026.8)
- Fortune, “[Dario Amodei on the Jevons paradox](https://fortune.com/2026/05/05/dario-amodei-jevons-paradox-will-ai-wipe-out-white-collar-jobs/)” (2026.5)
- James Bessen, “[Toil and Technology](https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm),” IMF Finance & Development (2015.3)
- James Feigenbaum and Daniel Gross, “[Answering the Call of Automation](https://jamesfeigenbaum.github.io/research/tele),” QJE (2024)
- Melanie Arntz et al., “The Risk of Automation for Jobs in OECD Countries,” OECD Social, Employment and Migration Working Paper 189 (2016)
- Deena Mousa, “AI isn't replacing radiologists,” Works in Progress (2025.9)
- Minkyu Shin et al., “Superhuman artificial intelligence can improve human decision-making by increasing novelty,” PNAS 120(12) (2023)
- Scale AI and Center for AI Safety, “[Remote Labor Index](https://arxiv.org/abs/2510.26787)” (2025.10)
- Frank F. Xu et al., “[TheAgentCompany](https://arxiv.org/abs/2412.14161)” (2024–2025)
- METR, “[Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)” (2025.7)
- Daron Acemoglu, “[The Simple Macroeconomics of AI](https://www.nber.org/papers/w32487),” NBER (2024)
- David Autor et al., “New Frontiers: The Origins and Content of New Work, 1940–2018,” QJE (2024) · MIT News [summary](https://news.mit.edu/2024/most-work-is-new-work-us-census-data-shows-0401)
- Waymo, “[Advice, not control: the role of remote assistance](https://waymo.com/blog/shorts/advice-not-control-the-role-of-remote-assistance/)”
- Nobel Committee, “[The Nobel Prize in Chemistry 2024](https://www.nobelprize.org/prizes/chemistry/2024/press-release/)” · MIT News, “[Using generative AI, researchers design compounds that can kill drug-resistant bacteria](https://news.mit.edu/2025/using-generative-ai-researchers-design-compounds-kill-drug-resistant-bacteria-0814)” (2025.8)
- Photos: “The Leader of the Luddites” (1812) · U.S. Women's Army Corps switchboard operators at the Potsdam Conference (1945, U.S. National Archives and Records Administration) · Elevator operator wearing a flu mask (1918, Webster & Stevens/MOHAI) · U.S. Navy Department typing pool (1918–19, Naval History and Heritage Command NH 52909) · NACA woman computer with a Friden calculator (1952, NASA) — all public domain, via Wikimedia Commons

