Comparing a published projection with what actually happened is the only real test of the method, and it is possible because every release is archived. The pattern is consistent: direction is usually right, magnitude frequently is not, and the misses cluster where an assumption about technology or a shock did the work rather than demographics.
Paralegals were the textbook case
The most useful thing you can do with a projection is check an old one. The record is more mixed than either the enthusiasts or the critics claim, and paralegals are the clearest illustration. Every release is archived, so the test costs nothing to run.
Document review was the example everyone reached for. It was repetitive, rule-governed, high-volume and expensive, and software genuinely did absorb most of it. The prediction that followed was that the occupation would collapse.
The occupation is projected at 376,200 people in 2024 and 376,800 in 2034, which is flat rather than gone. What happened is that the automatable part left and the rest stayed: client contact, case management, knowing which document matters and why, and being the person an attorney can ask. The bundle survived even though one item in it did not.
Three more that were written off
Radiologic technologists were supposed to be an early casualty of image recognition. They are projected to grow 4.3 percent to about 237,800 people at a median of $80,110. Reading an image turned out to be a smaller part of the job.
Travel agents were written off entirely when booking moved online, and the logic looked airtight at the time. They are projected to grow 2.2 percent, having settled into complex, high-value and multi-leg trips that a search box handles badly. Complexity became the product once simplicity was automated.
Bank tellers are the classic example and the one where the prediction was directionally right and badly wrong about pace. They are down 12.9 percent over the coming decade from 347,400 people, which is decades after the cash machine arrived. Being right about direction says nothing about being right about timing.
Where the forecasts did land
Some predictions were accurate and then some. Word processors and typists are projected down 36.1 percent, from 40,000 people to about 25,600, and data entry keyers down 25.9 percent. Both occupations shrank roughly as sharply as anybody warned.
Those occupations really did shrink the way the confident write-offs said they would. Ignoring that record would be as dishonest as ignoring the misses. The method deserves credit where it earned some.
The pattern separating the hits from the misses is clean. The forecasts were right where an occupation’s entire output was a single automatable task, and wrong wherever the job was a bundle of several. Typing is a task; being a paralegal is not.
Why the misses are systematic rather than unlucky
Scoring a whole occupation assumes it is one thing. Most occupations are five or six things at once, and they do not automate at the same rate. Averaging across those parts produces a number describing none of them.
The automatable parts leave first while the rest consolidates into the remaining time. The person still holds the job title and spends their week differently. The statistics record continuity while the work quietly changes.
That is the same methodological criticism that took headline automation risk estimates from 47 percent down to single digits when the analysis moved from occupations to tasks. The forecasts and the risk indices made the same error for the same reason. Occupation-level scoring was the shared mistake underneath both.
What survives is consistent
In every case above, the occupation kept the part requiring judgment, accountability or a relationship. It shed the part that was mechanical and rule-governed. The division fell along specifiability rather than difficulty.
That consistency is the most useful finding in this whole exercise. It is a pattern rather than a series of lucky exceptions, and it predicts which parts of your own week are durable. Apply it to your own tasks rather than to your job title.
No exposure score can represent it, because a score attaches to a title and the title is exactly what stays constant while everything underneath it moves. The measure is fixed to the one thing that does not change. That is a structural flaw rather than a calibration problem.
The uncomfortable half of the good news
An occupation surviving in a changed form still requires the person in it to change with it. That is the part reassuring coverage leaves out. Survival of the category is not survival of the person.
Somebody who was excellent at document review and indifferent at client management is worse off, even though paralegal is a stable occupation on paper. The aggregate held and their particular position did not. Averages protect nobody in particular from anything at all.
Aggregate survival is not individual security, and the aggregate figure is what gets quoted reassuringly. The right question is never whether your occupation survives; it is which of your tasks are the ones that stay. That question has an answer you can actually act on.
How to run the check yourself
Take a projection published a decade ago for a year now past, find the same occupation in current employment data, and compare the two. Every release is archived, which makes this a genuinely available exercise rather than a rhetorical one. An afternoon with two spreadsheets settles it properly.
Watch for occupational reclassification, which accounts for a share of apparent misses that are really definitional. An occupation split into two is not comparable with its earlier self and will look like a dramatic error. Definitional change masquerades as forecasting failure surprisingly often.
Expect the demographic projections to have held up better than anything resting on how fast a technology was adopted. That difference in reliability is consistent enough to plan around. Demographics beat technology assumptions in every comparison available.
What this justifies believing
That occupations rarely vanish outright and frequently change beyond recognition. Those two facts together describe almost every case in the record. Neither the alarmists nor the dismissers described it correctly.
That the confident write-offs are usually about a task and get reported as being about a job. And that the useful planning question is about your week rather than your title. Titles are durable in a way that daily work simply is not.
It also inoculates against the opposite error, which is treating all of this as guesswork. These projections were right about typists and data entry, and being right about the clear cases is what makes them worth reading carefully on the unclear ones. A method with a real track record deserves careful reading.
Common questions
Can I check an old projection?
Yes. Every release is archived, so a projection made for a year now past can be compared with what actually happened.
What is the general pattern?
Direction is usually right and magnitude frequently is not, with misses clustering where technology assumptions or shocks did the work.
Which projections held up best?
Demographically driven ones, because the driver was already visible when the projection was made.
What about occupations written off and still here?
That happens in both directions. Adoption is sometimes slower than expected, or the work is harder to automate than it appears from outside.
What is the trap when comparing?
Occupational reclassification. A split or merged occupation is not comparable with its earlier self, and that explains some apparent misses.
Did paralegals get automated away?
No. Projected at 376,200 in 2024 and 376,800 in 2034 — flat, after software genuinely did absorb most of document review.
Which write-offs were accurate?
Word processors and typists, projected down 36.1 percent to about 25,600, and data entry keyers down 25.9 percent.
What separates the hits from the misses?
The forecasts were right where an occupation's whole output was one automatable task, and wrong wherever the job was a bundle of several.