It Took Twelve Years to Show — Save The 1 Percent

In 2013, Carl Benedikt Frey and Maike Osborne gave 702 American occupations a probability of being done by computers. Five years later, Michael Coelli and Jeff Borland of the University of Melbourne checked the ranking against what actually happened to the jobs, and found that it told economists almost nothing they did not already know from an older rule: machines take routine work first.

This video scores the list against what the Bureau of Labor Statistics counted in 2013, 2018, 2022 and 2025: six jobs one by one, then every job at once. It repeats the 2019 test, and with data to 2025 the result reverses. The list caught something the older rules missed, and it took twelve years to show. The limits matter as much: the scores still explain less than a sixth of the differences between jobs, the counts cover wage jobs only, and they show which jobs grew, not why.

Related reading, different angle on the same subject: 47% of What?

Check the numbers

The figures in this video are the channel's own calculation, not a peer-reviewed study. These two scripts produce every one of them from public data, downloaded fresh from the original sources each time they run.

fo_vs_oews.py: the six-job scoreboard, and employment growth for the three risk bands on the 604 occupations counted under the same code at every date.
fo_vs_rti.py: the 2019 test repeated with data to 2022 and 2025. Do the Oxford scores still say anything once jobs are sorted by how routine their work is?
LICENSE: MIT. Reuse them freely, with credit.

To run them you need Python 3, uv and pdftotext (part of Poppler). The Bureau of Labor Statistics refuses downloads without a contact address, so set your own first:

export [email protected]
uv run --with pandas --with openpyxl --with xlrd fo_vs_oews.py --end 18
uv run --with pandas --with openpyxl --with xlrd fo_vs_oews.py --panel 18,22,25
uv run --with pandas --with openpyxl --with xlrd fo_vs_rti.py

Keep both files in the same folder: the second uses the first.

Sources

Frey, Carl Benedikt, and Michael A. Osborne. "The Future of Employment: How Susceptible Are Jobs to Computerisation?" Working paper, Oxford Martin School, University of Oxford, 17 September 2013. www.oxfordmartin.ox.ac.uk. Appendix ranking of 702 occupations.
Coelli, Michael, and Jeff Borland. "Behind the Headline Number: Why not to Rely on Frey and Osborne's Predictions of Potential Job Loss from Automation." Melbourne Institute Working Paper 10/19, October 2019. melbourneinstitute.unimelb.edu.au.
United States Bureau of Labor Statistics. Occupational Employment and Wage Statistics, national estimates, May 2013, May 2018, May 2022 and May 2025. www.bls.gov/oes.
Task data: David Autor and David Dorn (2013), routine, manual and abstract task scores and the occupation crosswalk, from David Dorn's data page (files A8 and B1); United States Census Bureau, 2010 Occupation Code List. Four task groups after Daron Acemoglu and David Autor (2011), as used by Coelli and Borland.
Brynjolfsson, Erik, Bharat Chandar, and Ruyu Chen. "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence." Stanford Digital Economy Lab, August 2026 version. digitaleconomy.stanford.edu. Workers aged 22 to 25, November 2022 to June 2026: about 11 percent fewer in the most exposed jobs, about 10 percent more in the least exposed.
The Budget Lab at Yale. "Tracking the Impact of AI on the Labor Market." Updated 15 September 2026. budgetlab.yale.edu.
Image credits: every image in the video is an original line drawing or a typeset graphic. One drawing is made after a photograph: Roy Amara, Rochester Institute of Technology, News & Events, 25 September 1980, RIT Digital Archive.