Report Digest - 8 minute read, or 2 if you are busy
Is AI taking jobs yet?
Anthropic's economists built a new way to measure which jobs AI is actually touching, then checked it against US employment data. The honest answer so far is "not that we can see, with one early warning". Here is the short version.
Digest of Labor market impacts of AI: A new measure and early evidence by Maxim Massenkoff and Peter McCrory, published by Anthropic on 5 March 2026. The findings and numbers are theirs. The summary and the opinions are mine.
The 2-minute version
Four numbers, then one sentence.
In computer and maths jobs, the share of tasks AI could do in theory, against the share actually covered by observed use.
Task coverage for computer programmers, the most exposed occupation on this measure.
Of workers have zero coverage. Cooks, mechanics, lifeguards and bartenders are examples.
Estimated drop in the rate at which 22 to 25 year olds start jobs in exposed occupations. The authors call it barely statistically significant.
My one-line summary
So far there is no clear sign that AI has pushed exposed workers into unemployment. There is a faint signal that fewer young people are being hired into exposed jobs, and the authors are careful not to overstate it.
A new measuring stick
Earlier studies asked what AI could do in theory. This one adds what people are actually using it for at work. The authors call the result "observed exposure".
What goes in
A US database of the tasks in about 800 occupations, earlier estimates of which tasks a language model could speed up, and Anthropic's own data on how Claude is used.
What counts more
Work use counts, personal use does not. Fully automated use gets full weight. Use where AI assists a person gets half weight.
What comes out
A coverage score for each job: the share of its work, weighted by time, where AI is really being used in an automated way.
Of observed Claude use falls on tasks rated fully feasible for a language model alone.
Falls on tasks rated not feasible.
Of observed tasks fall in categories rated theoretically feasible. Theory and practice point the same way.
Theory against practice
Scroll slowly. Picture 100 tasks in a typical computer and maths job.
Start
One hundred tasks
Every job is a bundle of tasks. Take 100 of them from the computer and maths category.
In theory
94 could be sped up
Earlier research judged that a language model could, in theory, speed up 94% of these tasks.
In practice
33 actually are
Observed use on Anthropic's platform covers 33% of the tasks. That is real, and it is a fraction of what is possible.
The gap
The gap is the thing to watch
The authors expect the covered area to grow as tools improve and adoption spreads. Tracking how fast this gap closes is the purpose of the measure.
Who is most exposed?
Coverage on the new measure. The top ten list in the report also includes customer service representatives and financial analysts.
Observed coverage
Zero coverage does not mean AI can never touch the job. It means the tasks appeared too rarely in the usage data to count.
17.4% hold a graduate degree. Compared with the unexposed group they are 16 percentage points more likely to be female and earn 47% more on average.
In short: older, more educated, higher paid. Not the usual picture of who automation hits first.
Has anyone lost their job?
This is the question everyone asks. The report's answer is careful.
Unemployment rates. The left side is the level the report cites. The right side is a scenario the authors describe to show what their method could detect, not a forecast. They estimate a differential rise of about 1 percentage point would be visible.
What they found
No systematic increase in unemployment for highly exposed workers since late 2022. The trend lines for exposed and unexposed workers have stayed similar.
The early warning
Fewer 22 to 25 year olds seem to be starting jobs in exposed occupations: an estimated 14% drop in the job-finding rate, barely statistically significant. No such drop for workers over 25.
Other explanations
Young people not hired may have stayed in their current jobs, taken other jobs or gone back to study. Surveys also measure job moves imperfectly.
The limits the authors state
A good report tells you where not to lean on it. This one does. I have added one of my own.
It will not catch everything
The authors say the approach will not capture every way AI could reshape work.
The theory scores are dated
The "possible in theory" ratings reflect what language models could do in early 2023.
It is one company's data
Usage comes from Anthropic's own product. Other tools may be used differently. The authors say the measure can take in other usage data.
It is US data
Tasks, jobs and unemployment figures are American. My own note: none of this tells us directly about Kampala or Nairobi.
Projections are weakly linked
For every 10 point rise in coverage, the official 2024 to 2034 growth projection falls by 0.6 points. The authors call the relationship slight.
What this means for you
Choose the hat you wear most.
Do not panic and do not coast. The data shows no wave of job losses, and it shows AI use growing inside jobs like yours.
List your own tasks. Mark which ones AI already helps with. Get very good at the rest.
Become the person on your team who knows how to check AI's work.
Check yourself
Five quick questions.