The jobs AI will replace first are not whole occupations but slices of work: routine writing, basic coding, image production and other tasks a machine can do cheaply. So far the clearest damage falls on freelancers and on young people trying to enter exposed occupations. Experienced workers and the wider labor market show little change.
Most public talk about this question rests on three assumptions. The first is that a job is one solid block that a machine either can or cannot do. The second is that once software can perform a task, employers will soon stop paying people to perform it. The third is that the damage, when it comes, will arrive as a wave of layoffs that everyone can see. Each assumption sounds reasonable. Each one breaks down when you look at the research.
A job is a bundle of tasks, and most bundles contain some work that software handles well and some that it handles badly. Technical possibility is not the same as a business decision. A firm automates only when the machine is cheaper and good enough, and when customers, regulators and the rest of the workflow allow it. The newest payroll evidence also suggests that the first effects are quiet ones. They show up as jobs that are never advertised rather than as people who are shown the door. This article works through the evidence in the order it was produced, from the early forecasts to labor market records collected in 2026, and it states plainly where findings conflict.
Where Did the Idea That Half of All Jobs Would Vanish Come From?
The figure most people remember comes from an influential 2017 study by two University of Oxford researchers. They started with expert judgments about whether a sample of occupations could be automated with the technology of the time, then used a statistical model to extend those judgments to 702 detailed occupations in the United States. Their headline result placed about 47% of total United States employment in a high risk category. Office and administrative support, transportation and logistics, and production work sat near the top of the risk list. Management, business and finance, education, health care, the arts and creative specialist work sat near the bottom.
The study answered a narrow question: could the typical work done in this occupation be performed by a computer? It did not try to predict how many people would actually lose their jobs. It treated a whole occupation as automatable if its typical tasks looked automatable, so a receptionist who also calms upset visitors and fixes scheduling conflicts counted the same as one who only answers the phone. It set aside cost, customer acceptance, regulation, and the new tasks that tend to appear when old ones are mechanized.
Those choices were reasonable for the question the researchers asked. The trouble is what happened next. The 47% figure travelled far beyond its question. It was repeated as a forecast of job loss in speeches, board presentations and career advice, and it became the starting point for a decade of anxiety about automation. When the newer chatbots built on large language models arrived in late 2022, the same habit of mind returned: if a machine can write an email, surely the people who write emails are finished.
The gap between that belief and the evidence is the subject of the rest of this article. The useful question is not "can a machine do some of this job?" The useful question is "what happens to employment, hiring and pay when it does?" Those are different questions, and the research gives quite different answers to them.
Why Is Exposure of Tasks Not the Same as Replacement of Jobs?
The first serious challenge to the 47% figure came quickly. A 2016 working paper written for the Organization for Economic Cooperation and Development looked at what individual workers in 21 member countries actually do in their jobs, rather than assuming that everyone with the same job title does the same work. Using this task based approach, the authors estimated that on average only 9% of jobs were automatable, ranging from about 6% in Korea to about 12% in Austria. This was a working paper rather than a peer reviewed article, so it deserves some caution, but the same researchers later tested the core idea in a later peer reviewed journal article. When they allowed for the wide variation of tasks within each occupation, the estimated share of United States jobs at high risk fell from 38% to 9%.
The reason is simple once you see it. Two accountants, two teachers or two clerks may share a job title and spend their days in very different ways. If even part of a job involves face to face problem solving, judgment calls, or adapting to unusual situations, the whole job is far harder to hand to a machine. Measuring at the level of the occupation hides that variety and inflates the risk.
The arrival of large language models, which are computer programs trained on huge amounts of text to predict and produce language, renewed the exposure debate with a different technology. A 2024 Science article estimated that about 80% of the United States workforce could have at least 10% of their work tasks affected by these models, and that about 19% of workers could see at least half of their tasks affected. In this research, a task counts as exposed when the technology could cut the time needed to complete it by at least half without lowering quality. Unlike earlier automation waves, the exposure reached well into professional and higher paid work.
Those numbers are often quoted as if they were job loss forecasts. They are not. Exposure means overlap: the technology touches the task. It says nothing about whether the worker is replaced, helped, or simply given a new tool that changes how the hours are spent. The same is true of a popular exposure measure published in 2021, which links advances in specific machine capabilities, such as image recognition and language processing, to the human abilities each occupation requires. It produces exposure scores for occupations, industries and even counties. Its authors built it so that others could study the effects of artificial intelligence, not to declare those effects in advance, and an occupation with a high score could gain from the technology just as easily as lose from it.
Whether overlap becomes replacement depends on at least four things that exposure scores do not capture. The first is how much of the job is exposed. A job with one exposed task out of 10 is in a different position from a job with eight. The second is whether the exposed tasks can be separated from the rest. If a loan officer's paperwork can be automated but the paperwork is woven into client conversations, the job changes shape rather than disappearing. The third is how demand for the product responds when the work becomes cheaper, which history shows can push employment up as well as down. The fourth is cost: a firm will keep paying a person when that remains cheaper or safer than buying, checking and maintaining a system. Exposure is the start of the analysis, not the end of it.
What Does History Say About Machines and Employment?
Fear of technological unemployment is old, and so is the puzzle of why it has rarely come true at the scale people expected. The economist David Autor addressed this puzzle directly in a 2015 journal essay. His argument was that automation does replace human labor in some tasks, but it also complements workers in other tasks, raises output, and lifts demand for labor. Machines that take over routine, rule following work tend to make the remaining human tasks, such as problem solving, adaptability and creativity, more valuable. One example he discusses is bank tellers, whose numbers did not collapse when cash machines spread, because branches became cheaper to run and the tellers' work shifted toward relationship tasks. He also noted the cost of this pattern: in recent decades employment and wage gains concentrated at the top and bottom of the skill ladder, squeezing middle wage work.
Whether automation adds or removes jobs in a given industry depends heavily on what happens to demand. A two century study of the United States textile, steel and automobile industries found that automation first spurred job growth, because falling prices led people to buy far more cloth, steel and cars. Only later, once people already had most of what they wanted, did further productivity gains bring falling employment. The lesson for today is that the same technology can grow one industry and shrink another, depending on whether customers would buy much more of the product if it became cheaper.
History also shows where new work comes from. A 2024 study built a database of new job titles across eight decades and linked it to census records and patents. It found that the majority of current United States employment is in job specialties that did not exist in 1940. New work tends to appear when innovations complement what an occupation produces. The same study carries a warning, though. Innovations that automate tasks reduce demand for the occupations they touch, and the job eroding effect of automation has grown stronger over the past four decades while the job creating effect of augmenting technology has not. Optimism drawn from history needs that caveat attached.
Not every technology wave has ended well for the workers in its path. Research on industrial robots in United States local labor markets found that each additional robot per thousand workers was linked to a fall of about 0.2 percentage points in the share of adults in employment and a 0.42% fall in wages. Those are real losses for real communities. A parallel study using German administrative records found a different overall picture. Manufacturing jobs lost to robots were fully offset by new jobs in services. Workers already employed in exposed plants mostly kept their jobs and took over new tasks, often of higher quality. The burden fell on young people just entering the labor force, who found fewer openings in manufacturing and shifted from vocational training toward colleges and universities.
That German finding matters for the present debate. It shows that a technology can reduce employment in exposed work without anyone being laid off: firms simply stop hiring newcomers into the old jobs. Keep that pattern in mind, because the newest evidence points in the same direction.
What Happened to Jobs Before Chatbots Arrived?
Machine learning, the branch of computer science in which programs learn patterns from data rather than following hand written rules, spread through firms well before the public chatbots of late 2022. A vacancy study between 2010 and 2018 traced what happened. Establishments whose work involved tasks that the technology could perform posted rapidly growing numbers of vacancies for specialists in artificial intelligence. At the same time, those establishments cut hiring for other positions and changed the skills they asked for in the jobs that remained. So the substitution was real at the level of the workplace. Yet when the researchers looked at occupations and industries as a whole, the effects on employment and wage growth were too small to detect. Change was happening inside firms long before it showed up in national statistics.
A cross country analysis reached a similar conclusion. Research across 23 countries, conducted by economists at the Organization for Economic Cooperation and Development and published in a peer reviewed journal, applied an exposure measure to employment data from 2012 to 2019. Across all occupations, it found no clear link between exposure to artificial intelligence and employment growth. In occupations where people use computers heavily, higher exposure was associated with faster employment growth, possibly because workers with good digital skills could move toward higher value tasks. In occupations with little computer use, there was suggestive evidence of fewer average hours worked. The study also found that highly educated workers were concentrated in the most exposed occupations: 82% of them worked in the more exposed half of occupations, compared with 16% of workers with low education.
Before 2022, then, the evidence pointed to reorganization inside firms rather than mass displacement. The obvious objection is that earlier systems were narrow, built to recognize images or predict prices, while large language models handle a broad range of language tasks. That objection is fair, and it is why the research since 2023 deserves careful attention.
Which Jobs AI Will Replace First: What Do Freelance Markets Reveal?
Online freelance platforms are the cleanest place to watch substitution happen. A client posts a task, such as writing a product description, translating a page, or fixing a piece of code, and pays a freelancer to complete it. There is no employment contract to unwind, no team to reorganize and no severance to pay. If a chatbot can do the task well enough, the client can stop posting it the next day.
One study of freelancers on a large online platform examined what happened after the release of a widely used chatbot in November 2022 and several image generation tools. Freelancers in the most affected occupations experienced fewer jobs and lower earnings than comparable freelancers in less affected work. A strong track record did not protect them: past performance and employment history did not soften the blow, and there was suggestive evidence that the best freelancers were hit hardest, perhaps because they had specialized in exactly the work that became cheap.
A larger platform study found a 21% fall in job posts for writing and coding work that could be automated, compared with jobs that required manual skills, within eight months of the chatbot's release. After image generators appeared, posts for image creation fell by 17%. Two further findings complicate the simple replacement story. Competition among freelancers increased, because more people chased fewer posts. And the posts that remained were more complex and paid more. Clients stopped buying the simple tasks and kept buying the hard ones.
This is the best evidence available that generative tools can replace paid human work, and it should be taken seriously. It also has limits. Freelance platforms account for a small share of total employment. The work traded on them is already split into separate tasks, which is exactly the condition under which substitution is easiest. Inside a firm, the same writing or coding task is usually tied to meetings, reviews, accountability and knowledge of the business, which makes it harder to remove cleanly. The freelance results show what happens when tasks are fully separable. Most salaried jobs are not built that way.
Are Young Workers the Canaries in the Coal Mine?
The most discussed new evidence comes from Erik Brynjolfsson and colleagues at Stanford, in an early careers paper first released in August 2025 and revised in August 2026. It has not yet completed peer review. The researchers used anonymous monthly payroll records from a large payroll processing company, covering between 3.5 million and 5 million employees each month from January 2021 to June 2026, and grouped occupations by how exposed their tasks are to generative tools.
The headline finding is about age. For workers aged 22 to 25, employment in the most exposed occupations fell by about 11% between November 2022 and June 2026, while employment for the same age group in the least exposed occupations grew by about 10%. The researchers estimate that employment of young workers in exposed occupations now stands about 19% below where it would have been had it kept pace. That gap was about 15% in the data available in mid 2025 and has widened since. Software development and customer service are among the occupations highlighted. More experienced workers in the same exposed occupations show no comparable decline.
Three details make this study more useful than a simple headcount comparison. First, the decline came mainly through reduced hiring of young workers, not through more layoffs or separations. This is the German robot pattern again: the door to the occupation narrows for newcomers while people already inside stay put. Second, the declines were concentrated in occupations where people use generative tools mostly to automate work, handing over a task to be completed. In occupations where people mostly use the tools to augment their own work, checking, refining and learning with the tool, employment was flat or rising, especially for experienced workers. Third, the adjustment showed up in the number of jobs more than in base pay.
The authors are careful about what their data can and cannot show, and readers should be too. The results are correlations, not proof that the technology caused the decline. Some of the divergence began before the chatbot was released. Adjusting for education weakens the estimates, and the gap in this payroll sample is larger than in national survey data. Other forces, such as rising interest rates after 2022 and a pullback in technology hiring, hit the same young workers at the same time. Even so, the finding fits a plausible mechanism. Young workers bring knowledge learned from books and courses, which is exactly the kind of codified knowledge that large language models have absorbed. Experienced workers bring judgment, relationships and knowledge of how this particular firm works, which the models do not have.
Why Do Danish Records Show Almost No Change?
If the early career finding is a warning, the Danish evidence is a check on panic. A Danish working paper published by the National Bureau of Economic Research, which has not yet completed peer review, linked surveys of about 25,000 workers in about 7,000 workplaces to national administrative records on earnings and hours. The surveys covered 11 occupations thought to be exposed to chatbots, including accountants, teachers and software developers, and tracked outcomes from the chatbot's release in November 2022 to the end of 2024.
Adoption was widespread. In workplaces where employers had encouraged use, provided an approved chatbot and offered training, about 93% of workers had used one, compared with roughly 40% in workplaces with none of these measures. Workers reported time savings. Yet when the researchers compared earnings and recorded hours between workers and workplaces with more and less adoption, they found effects close to zero. Their estimates rule out effects on earnings or hours larger than 2% two years after the chatbot appeared. Employers mostly responded by reorganizing tasks, and some workers moved into better paid occupations where chatbots were more relevant, though not enough to change average earnings.
A 2025 preprint, which has not been peer reviewed, found something similar in United States data. Comparing occupations with different levels of exposure, it found that unemployment rates in highly exposed occupations did not change after the chatbot's release, while earnings in those occupations rose.
How can the Danish and United States evidence sit beside the early career decline? The studies measure different things. The Danish research follows people who already had jobs, and asks whether their pay and hours changed. The early career research looks at who gets hired into exposed occupations in the first place. Both can be true at once: existing workers absorb the tool and carry on, while employers quietly hire fewer beginners. The Danish data also end in 2024, while the payroll data run to mid 2026, and Danish labor institutions differ from those in the United States.
A wider view supports caution about dramatic forecasts in either direction. Daron Acemoglu, using a model built on task level exposure and cost savings, estimated in a macroeconomic analysis that the technology would raise total factor productivity, the economy's output for a given amount of labor and capital, by no more than 0.66% over 10 years, and possibly by less than 0.53%. Early gains come from easy tasks with clear right answers, and many remaining tasks depend on context and lack objective measures of success. A technology with gains of that size is unlikely to trigger mass unemployment within a decade. It is also unlikely to deliver the boom some of its promoters promise.
So Which Jobs AI Will Replace, and What Does That Mean for You?
Start by dropping the idea that your job title carries a single risk score. What carries risk is the mix of tasks you do, how easily those tasks can be separated from the rest of your work, and whether you are already inside the occupation or trying to get in.
Write down the 10 or 15 tasks that take most of your week. For each one, ask two questions. Could a chatbot produce a first version that someone could check quickly? And is that task bought and sold on its own, the way freelance writing or basic coding is, or is it tied to relationships, accountability and knowledge of your organization? Tasks that are both easy to check and easy to separate are the ones under real pressure. The freelance evidence shows that such work can lose a fifth of its demand within months.
Next, look at how people in your occupation use these tools. If they mostly hand over whole tasks, your occupation looks like the ones where young workers are being hired less. If they mostly use the tools to work better, checking, refining and extending what they already do, your occupation looks like the ones where employment has held steady or grown.
If you are early in your career, the evidence says the risk sits less in being fired and more in not being hired. Your advantage is not the knowledge you learned in a course, which the models already have, but the judgment and context you build on the job. Get into roles where you work alongside experienced people and handle real exceptions as early as you can.
If you are experienced, the current evidence is more reassuring. Your earnings and hours, on the best data available, have barely moved. That can change, and the job eroding effect of automation has grown over recent decades, so treat the calm as time to adapt rather than proof that nothing will change.
Key Takeaways
- A high exposure score means a technology overlaps with some of your tasks. It is not a forecast that your job will disappear, and early forecasts that treated it as one overstated the risk several times over.
- When researchers measured what individual workers actually do, rather than assuming everyone in an occupation does the same work, the share of jobs at high risk fell from roughly four in 10 to about one in 10.
- Freelance markets, where tasks are bought one at a time, show the clearest replacement so far: demand for simple writing, coding and image work fell sharply within months, while harder, better paid tasks remained.
- The first sign of displacement in salaried work is not layoffs. It is fewer young people hired into exposed occupations, especially where the tools are used to hand over tasks rather than to help people do them.
- Workers already in exposed jobs have, so far, seen almost no change in earnings or hours, according to the strongest national records available.
- History offers real comfort, because new kinds of work keep appearing, but also a warning: automation has been destroying demand faster in recent decades while the creation of new work has not kept pace.
Implications for Practice
Stop using occupation level risk scores to make headcount decisions. The gap between the occupation based estimate of 38% and the task based estimate of 9% is a warning about what happens when you treat a job title as a single unit. Workforce planning for this technology should start with a task inventory for each role: what people actually spend their time on, which tasks could be handed to a tool, and which tasks are bound up with judgment, relationships and accountability. A role in which two of 12 tasks are exposed needs redesign, not removal.
Watch your hiring pipeline, not only your layoff numbers. Both the German robot study and the recent payroll evidence show that technology tends to cut employment by closing the entry door rather than pushing people out. If your organization is quietly hiring fewer graduates and junior staff into exposed functions, you may be weakening the pipeline that produces your future experienced people. Ask each function head to explain any fall in junior hiring and whether it reflects a deliberate decision or simply drift. An organization that stops training beginners today will find that it has no mid career specialists within a decade.
Decide deliberately whether a tool is meant to automate a task or to help people do it better. The payroll evidence links employment declines to automating uses and stable or rising employment to augmenting uses. Human resources leaders can shape that choice through how roles are redesigned, how tool use is introduced, and what managers are rewarded for. Where the aim is augmentation, protect the parts of the work through which junior staff learn, and make sure experienced employees pass on the judgment that the tools lack.
Review spending on contractors and freelancers first. The platform evidence suggests that separable tasks bought from outside, such as simple copywriting, translation and routine coding, are where substitution happens fastest. That is also where the human cost to your own workforce is lowest. It is a reasonable place to test tools, measure quality and learn, before changing any internal role.
Expect the payroll effects to be slow and modest. The Danish evidence shows that even with high adoption, encouragement and training, earnings and hours barely moved in the first two years, and the macroeconomic analysis points to modest productivity gains over a decade. Business cases that assume rapid headcount savings from chatbots are not supported by the evidence so far. Plan for task reorganization, track it, and revisit assumptions as new data arrive.
When a role does shrink, follow the German example and redeploy before you release. In that study, workers already employed in plants that automated mostly stayed and took over new, often better, tasks in the same plant. Redeployment preserves knowledge of the organization that new hires and machines cannot replace, and it is far cheaper than recruiting the same people back later.
Related Reading on The Human Capital Hub
- Thirty scientific findings on employee productivity
- Uses of artificial intelligence in human resources with evidence
- Taking back control of your career in 2026
