.png)
The Job Stays, The Task Goes: Why do the numbers on AI and Jobs never agree?
Back in 2013, two researchers at Oxford University, Carl Benedikt Frey and Michael Osborne, estimated that about 47% of the jobs in the United States were at high risk of being automated. A few years after that, a team of economists at the OECD, made up of Melanie Arntz, Terry Gregory and Ulrich Zierahn, asked the same question using data from 21 countries, and they arrived at a figure of roughly 9%.
Illustrataion by The Geostrata
The economy was the same and so was the technology, but the two answers were five times apart. The difference came from what each team decided to measure. Frey and Osborne rated entire occupations, while the OECD researchers looked at the individual tasks that people actually carry out, and those can vary a great deal even between two workers who share the same job title.
That study is now more than a decade old, but the problem it revealed is still with us. In 2026, forecasts about artificial intelligence are once again pulling in opposite directions, with some experts warning that half of all jobs could disappear and others pointing out that almost nothing has changed so far.
A good deal of the best evidence now comes from Anthropic, the company that makes Claude, which has begun to measure how AI is used at work one task at a time, and the company’s own chief executive and its head of economics have said rather different things about what that evidence means.
WHAT ANTHROPIC MEASURED
Most earlier studies asked what AI could do in theory. In March 2026, Anthropic economists Maxim Massenkoff and Peter McCrory tried a different approach, which they call “observed exposure.”
They took the government’s lists of tasks for about 800 occupations and checked which of the tasks that language models could speed up were actually being done with Claude in the workplace, and they gave extra weight to the tasks that Claude handles on its own.
Their main finding is that real-world use is still far behind what the technology could do. For computer and math jobs, language models could in theory speed up 94% of the tasks, yet Claude’s observed use covers only 33% of them.
Programmers turned out to be the most exposed group, with about 75% of their tasks covered, and customer service and data entry workers were not far behind.
When it came to jobs, the result was fairly undramatic. Since late 2022, the researchers found no systematic rise in unemployment among the most exposed workers when compared with the least exposed ones. They did add two cautions, though.
Government projections show slower job growth through 2034 for the more exposed occupations, and people between 22 and 25 are being hired into those occupations about 14% less often than before, although the authors describe that result as barely significant.
A GROWING FOOTPRINT
Anthropic’s Economic Index looks at the same question using a broader set of usage data. In its early reports, about 36% of occupations had AI involved in at least a quarter of their tasks, and by early 2026 that share had grown to 49%.
It is worth remembering that this figure counts tasks and not whole jobs, so it shows that Claude now handles a real part of the work in about half of all occupations, but it does not mean that half of all jobs have gone.
In a June 2026 survey of about 9,700 users, only around 10% said they thought they were likely to lose their job within a year, although these people already use Claude and may be more optimistic than most workers.
WHAT OUTSIDE RESEARCHERS
Researchers outside Anthropic describe a similar picture, but with an important twist. The Yale Budget Lab, whose tracker was updated in October 2026, says that the mix of occupations is not changing in a way that clearly follows the spread of AI, and that AI use shows no link to changes in employment or unemployment.
Stanford researchers, however, do see something happening at the entry level. In the August 2026 revision of their paper “Canaries in the Coal Mine,” they report that employment among 22-to-25-year-olds in the most exposed occupations is about 19% below where it would have been if it had kept pace with their less exposed peers, up from a gap of 15% a year earlier.
Workers with more experience do not show a gap like this, and the reason seems to be that companies are hiring fewer young people, and not that they are laying off more of the people they already employ.
The declines are also concentrated in those jobs where AI takes over human tasks, whereas in jobs where AI mostly helps people to do their work.
Some news from just the past few days shows how murky this picture can be. On October 26, Microsoft offered voluntary exit packages to around 60,000 of its employees in the US, Canada, India and the UK, which, by one estimate, is roughly a quarter of its entire workforce.
The company describes the offer as a time-limited, voluntary program that gives teams a way to “adapt their workforce to changing needs,” and it has been careful to say that this is not a layoff. Certain groups, such as those working in data centres and in sales and marketing, were left out of it.
Microsoft’s own announcement does not mention AI at all, but press coverage has linked the move to the company’s effort to rework its workforce for the AI age. It also remains unclear how many people will actually accept.
When Microsoft made a similar offer to retirement-age staff earlier this year, over 30% of the roughly 8,750 eligible workers took it. So the headline number tells us how many people were offered a way out, and not how many jobs will really disappear.
WHERE THE FORECASTS SPLIT
This is the point at which Anthropic’s own people begin to disagree with one another. In May 2025, the company’s chief executive, Dario Amodei, warned that AI could wipe out as many as half of all entry-level white-collar jobs, and that unemployment might rise to somewhere between 10% and 20% within the next one to five years.
In July 2026, however, Peter McCrory wrote an essay that summed up eighteen months of the company’s research, and the tone of it was quite different. He argued that AI had not yet made a real dent in the US labor market.
He noted that unemployment stood at 4.2% in June, and he said that he didn’t expect it to be noticeably higher a year from now, at least not because of anything AI was doing. He did concede, though, that companies have been hiring fewer young workers into the roles that AI affects the most.
Then, in September, the Anthropic Institute published a report titled Economic Scenarios for Transformative AI. It comes closest to being a forecast from the company, although its authors are careful to say that the scenarios are not predictions.
Looking ahead to 2030, the report suggests that GDP could be roughly 1.6%, 8.3% or even 32% higher than it would be without AI, depending on whether things turn out to be modest, substantial or extreme.
As that happens, the share of the economy’s output that goes to workers slips from 59.4% to 56.1% and then down to 45.2%. In the substantial scenario, unemployment stays at about 4.6%, which is not far from where it is now, but in the extreme scenario, press reports put it at around 12% overall and nearly 18% among white-collar workers.
The drop in labor’s share is the figure that caught my attention most, because it hints that even when AI works extremely well, the bigger problem might be who ends up with the gains, and not just whether people still have work.
WHY THE ANSWERS DIFFER
The gap between 47% and 9% came down to how the researchers measured things, and that is still true today. For a start, studies don’t count the same thing, since a job can lose a quarter of its tasks and still carry on.
They also differ on whether they look at what AI could do or at what it is really doing. Another difference is the outcome they follow, because unemployment, hiring and wages are not the same, and that is how Anthropic can see no change in unemployment while Stanford sees a large fall in the employment of young workers, even though both can be right.
And then there is the time frame, because the numbers we have today say very little about what 2030 will look like.
We should also remember who is behind some of this data. Anthropic’s measures only cover how its own product is used, and critics have pointed out that a company with something to sell is the one studying what that product does to the job market. That is why the independent work from Stanford and Yale carries so much weight, since it makes the overall picture easier to believe.
WHAT TO TAKE AWAY
As things stand in October 2026, the reality is a good deal less dramatic than the loudest headlines on either side. AI already does a meaningful share of the tasks in many occupations, and it hasn’t led to any visible wave of unemployment so far.
It does seem to be holding back the hiring of young people in the fields it affects most, and the range of possible outcomes for 2030 remains very wide. So the better question to ask isn’t just whether AI will take our jobs, but how much of each job it will take over, how fast that will happen, and who will benefit once it does.
BY SHOUNAK MAJUMDER
TEAM GEOSTRATA
Comments