Oxford's 47 percent was never a count of lost jobs. Ten years later, its authors reread the paper: the work they had called safest had come within reach.
27 September 2026
The paper is seventy-two pages long and dated 17 September 2013. Two Oxford researchers, an economist named Carl Benedikt Frey and a machine-learning researcher, Maike Osborne, who then published as Michael A. Osborne, took 702 American occupations, scored each one for how easily a computer could do it, and counted how many people held the jobs that scored high. On page 38 they reported the total.
"According to our estimate, 47 percent of total US employment is in the high risk category."
That is the sentence the world remembers. For the decade that followed, it was the number people meant when they said the robots were coming.
What did not travel with it was the caveat. Four pages later, in a section headed Limitations, the authors wrote: "We make no attempt to estimate how many jobs will actually be automated." They repeated it in their conclusions. They listed what they had left out: wages, the price of machines, regulation, political resistance, and the ordinary difficulty of predicting technology. They noted that their method could not see a job that changes shape without disappearing.
Nor did the rest of the headline sentence. The occupations in that category, it continues, "are potentially automatable over some unspecified number of years, perhaps a decade or two." The authors offered their probability scale as "a rough timeline," on which the high-risk jobs were "likely to be substituted by computer capital relatively soon."
The paper gave a loose horizon. What it declined to give was a count. The number everyone quoted was not a prediction that 47 percent of jobs would go, and its authors said so twice in the paper itself.
What the paper measured was exposure. Take the jobs that existed in 2010. Ask of each one whether its tasks could, in principle and given enough data, be done by the kind of machine that engineers could see coming. Count the people in the jobs where the answer was probably yes.
That is a real question with a real answer. It is not the same as asking how many people will lose their jobs. A job can be exposed for decades and never be automated, because labor is cheap, because a regulator says no, or because the work reorganizes around the machine.
The public argument trampled that distinction from both sides. One side read the 47 percent as a countdown to mass unemployment. The other side went after the number itself.
Some answers engaged the paper on its own ground. In 2016, three economists writing for the Organisation for Economic Co-operation and Development (OECD) redid the exercise task by task instead of job by job, on the grounds that people with the same job title do different things. For the United States they found 9 percent of workers in highly automatable jobs, not 47. In 2018 another OECD study, covering thirty-two countries, put the figure at 14 percent, with a further 32 percent at risk of significant change in how the work is done. That is a fair fight about how much exposure there is, and the paper's number lost ground in it. Frey and Osborne still contest the method. Judging a job by "the fraction of tasks that can be done by machines," they wrote in 2023, "will lead to flawed estimates," because work is usually restructured before it is automated.
Others missed. In September 2013, the month the paper was dated, the United States had 136.8 million payroll jobs and an unemployment rate of 7.2 percent. In August 2026 it had 159.1 million and 4.1 percent. More than 22 million jobs have been added since, in the years when, on the popular reading, nearly half of all jobs were supposed to go. Every figure there is correct. None of it touches what the paper claimed. The paper said what machines might become able to do. It did not say what employers would do with them, or how many people would be let go.
An estimate of what could happen is not refuted by what has not happened yet.
The paper made a second claim, bolder than the headline and easier to test.
Frey and Osborne did not ask whether computers could do everything. They asked what computers could not yet do, and why. Drawing on the robotics and machine-learning literature and on a workshop at Oxford's engineering department, they settled on three obstacles they called engineering bottlenecks: the tasks where, in their judgment, machines would stay stuck. They then hand-labeled seventy occupations as automatable or not and let a statistical model spread those judgments across the rest.
The first bottleneck was perception and manipulation: handling irregular objects in cluttered, unpredictable places. Warehouses and hospitals are built for wheeled machines; most homes are not.
The second was creative intelligence. The obstacle, they wrote, was not generating novelty, which machines could already do. It was "stating our creative values sufficiently clearly that they can be encoded." Their conclusion: "In the absence of engineering solutions to overcome this problem, it seems unlikely that occupations requiring a high degree of creative intelligence will be automated in the next decades."
The third was social intelligence: negotiation, persuasion, care. Here they chose a precise test. "Even simplified versions of typical social tasks prove difficult for computers," they wrote, "as is the case in which social interaction is reduced to pure text." In the annual Turing-test competitions, where judges chat by text with hidden humans and programs, "sophisticated algorithms have so far failed to convince judges about their human resemblance."
From the bottlenecks came a sequence. A first wave would reach transport and logistics, office and administrative support, and production work. A plateau would follow while engineers chipped away at perception and manipulation. A second wave would "mainly depend on overcoming the engineering bottlenecks related to creative and social intelligence." Creative and social work would be the last within reach.
The appendix shows the sequence occupation by occupation. Telemarketers sat at the very end of the ranking, 702nd of 702, one of a dozen occupations at a probability of 0.99. Tax preparers were another, and paralegals scored 0.94. At the other end, writers and authors scored 0.038, application software developers 0.042, editors 0.055, and graphic designers 0.082.
This was still a claim about capability, not about layoffs. The same model produced the 47 percent. But a sequence of capabilities has one property a headline count lacks: you can later see which part of it gave way first. Ten years on, its authors went back to look.
In 2023 Frey and Osborne reread their own paper in public. The result appeared in 2024 in the Brown Journal of World Affairs under the title "Generative AI and the Future of Work: A Reappraisal." They did not soften what they had believed:
We firmly believed that in the absence of major leaps, tasks requiring creativity and social intelligence, as well as unstructured manual work, would remain safe havens for human workers. The jobs of journalists, scientists, software engineers, art directors, and architects, we documented, were all at low risk of being automated.
Then they asked the question directly: "did we underestimate the near-term scope of automation?"
On social intelligence, their answer was largely yes, for work that can be done remotely, though "key bottlenecks to the automation of social tasks still persist." In 2013 they had written that a computer making a subtle joke would need "a database with a richness of knowledge comparable to that of humans." Ten years later they recalled the requirement as "a large database of human-generated jokes and methods of benchmarking the algorithm's performance," and added: "We now have both." Social tasks, "which we deemed non-automatable in 2013," now looked different: "AI may be able to replace human labour in many virtual settings, meaning that if a task can be done remotely, it can also be potentially automated." Text, the medium they had picked in 2013 as the hard case, now fell within what they pictured as "a world where AI excels in the virtual space."
Their assessment of creativity was more mixed. Generative systems, they argued, recombine existing work well and leap poorly. Such a system "is best suited for creating sequels rather than new narratives," and it competes "with average human talent rather than with superstars."
The physical bottleneck, which their 2013 model expected to erode first and only gradually, largely held. They acknowledged that "advancements in computer vision have paved the way for automating more perception tasks," and pointed to "numerous robotaxi trials, from San Francisco to Shenzhen." But jobs built on complex perception and manipulation "remain relatively safe from automation, as we deemed that they were in 2013," because in high-stakes public settings "we cannot accept mistakes."
Set the two documents side by side, and the sequence has run the other way. In 2013, creative and social intelligence were to be the last bottlenecks to give way, and the physical world was to yield first, gradually. In 2023, by the authors' own assessment, social tasks done at a distance and creative work built by recombination had come within reach in principle, while the physical bottleneck was holding.
They do not call this a reversal. They write that "significant bottlenecks to automation remain," and they expect the bottlenecks of 2013 to keep constraining automation. The reading is ours. Every step in it rests on their own pages.
Nor do they read the shift as a verdict on the jobs they once called safe. Treating generative AI "as an automation technology is, in its current form, a mistake," they write; its immediate effect will be "the transformation of existing content-creating jobs, making them easier to perform." Or, more bluntly: "ChatGPT won't replace journalists, just like Github's Copilot won't replace coders."
They then drew a new line, narrower than the old one and again stated with conviction. Human advantage, they now say, lies in being present in person. The large language models behind chatbots such as ChatGPT, as they put it, "don't have bodies." It also lies in genuine novelty rather than recombination, and in physical work where errors are unacceptable. They closed with an expectation: "in the absence of major breakthroughs, we also expect the bottlenecks we outlined in our 2013 paper to continue to constrain our automation possibilities for the foreseeable future."
Frey and Osborne could have left their 2013 paper alone. They reread it in print instead, and said which parts had moved.
The strongest reply comes in two parts. First, it is early. What a machine can do in principle is not what employers do with it, and the labor market has barely registered any of this. The Budget Lab at Yale, which tracks the occupational makeup of the American workforce, found in September 2026 that "measures of AI usage show no connection to changes in employment or unemployment." Payroll records studied by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen at Stanford's Digital Economy Lab show "no evidence of widespread, economy-wide job displacement," and one narrow signal: workers aged twenty-two to twenty-five in the occupations most exposed to AI now stand 19 percent behind their less-exposed peers, mostly because fewer of them are being hired. The authors call their findings "early, descriptive indicators," not causal estimates. The pattern weakens once education is taken into account, part of it began before generative AI, and it is stronger in their payroll sample than in national surveys.
Second, the new line (in person, novelty, and high stakes) may hold for a long time, and its authors expect the old bottlenecks to keep binding.
Both points stand. The labor data cannot yet say jobs have gone in the wrong order, and the argument does not need the new line to fail. The new line may hold, and part of the old one did: in 2013 the authors judged that perception and manipulation were "unlikely to be fully resolved in the next decade or two," and ten years later, by their own account, they were not. What moved was the other side of the map. Its authors redrew it around social and creative work, the ground they had thought most secure.
Since 2013, estimates of AI's reach have mostly come as percentages, read as counts of jobs to be lost, and the argument over the size of the number rarely settles. The question that decides who is exposed first is a different one: which work, and in what order.
Forty-seven percent of what, then? Of the jobs that existed in 2010, measured against the machines an Oxford workshop could see coming in 2013. It was an honest measurement of exposure, read for ten years as a forecast of losses. The part of the paper that deserved the attention was the map of where machines would not go. That line has now been drawn twice in ten years, with conviction both times. Anyone who chose a career by the first map in 2013 chose by a line its own authors have since moved. Anyone reading the second today is choosing a career by it, and the choice is no more certain.
The ideas, the argument and the conclusions in this essay are mine. I develop each essay in conversation with an AI assistant: I set the question and the direction, it turns my thoughts into prose, and I review and rework every draft until it has the shape and form I want. Every figure is sourced, and the sources are listed below.