What Happened to the Jobs Earlier Automation Was Meant to End

The claims were specific, confident and public

Paralegals were going to be automated away a decade ago. There are 376,200 of them and the projections show the number flat through 2034. The occupation that was supposed to vanish did not move.

That is worth dwelling on, not to score a point against forecasters, but because the same kind of claim is being made again right now about a different set of occupations. The value of the old predictions is that we can check them. Checkable claims are rare enough to be worth using properly.

Almost nothing else in this subject can be checked. Every current claim about artificial intelligence and employment is a statement about the future, and the only evidence available for judging such statements is how the last round of them performed. Track record is the only external evidence available here.

What actually happened, briefly

Radiologic technologists were supposed to be early casualties of pattern recognition and are projected to grow 4.3 percent. Travel agents were written off completely and are projected to grow 2.2 percent. Both occupations found work the technology handled badly.

Bank tellers did decline, at 12.9 percent from 347,400 people, which is a slow contraction rather than the disappearance that was forecast. The full accounting, including the occupations where the forecasts were accurate, is set out in the growth and decline pillar. This article is about what the record teaches instead.

What matters here is not the scoreboard but the mechanism. Three things stood between the capability and the outcome, and all three are still operating today. None of the three has been solved by better technology.

Mechanism one: jobs are bundles

Document review really was absorbed by software, exactly as predicted. The prediction was correct about the task and wrong about the job. That distinction is the entire lesson in one sentence.

A paralegal also handles client contact, case management, knowing which document matters and why, and being the person an attorney can ask a question. Removing one component of a bundle changes the bundle rather than eliminating it. The remaining components absorb the freed-up time instead.

This is the mechanism that defeated the most confident predictions, and it defeats them in the same way every time. An occupation is not a task with a name attached. Treating it as one is what produced the failed forecasts.

Mechanism two: verification governs the pace

The capability to perform a task and the willingness to rely on it are different things. What separates them is how quickly somebody can tell that the output is wrong. Detection speed governs willingness to rely on anything.

Where errors surface immediately and cost little, adoption is fast. Where a plausible wrong answer is dangerous and expensive to detect, adoption is slow regardless of how good the technology is. Capability arrives long before permission to use it does.

That is most of why imaging did not go the way the predictions said. The capability arrived and the accountability question did not resolve, and the accountability question moves on a much slower clock. Institutions revise their rules over decades rather than quarters.

Mechanism three: somebody has to redesign the work

Adoption requires a person inside an organization to redesign a process, and that person already has a job. This is the least discussed of the three and probably the largest. Deployment is a project rather than a decision somebody makes.

Technology does not deploy itself into an organization. It gets deployed by people with competing priorities, limited budgets and an existing arrangement that mostly works. Competing priorities defeat more projects than technical limits do.

Predictions written from the outside consistently underestimate this, because from the outside the change looks like a decision rather than a project. Inside, it is a project competing with every other project. From outside it looks like a switch somebody could flip.

What did disappear, and what it had in common

Some predictions were entirely correct, and any honest account has to say so. Word processors and typists are projected down 36.1 percent to about 25,600 people, and data entry keyers down 25.9 percent. Both figures come from the current projection round.

Those occupations shrank because their entire output was a single automatable task. There was no bundle to survive, no verification problem worth solving, and no redesign required beyond buying software. Single-task occupations offer no resistance at any of the three points.

So the record is not that predictions fail. It is that they succeed on single-task occupations and fail on bundles, which is a distinction anybody can apply to a new claim in about a minute. The test is cheap enough to run on everything you read.

Applying the record to today’s claims

When you meet a claim that an occupation is about to be automated, run the three mechanisms against it rather than arguing about the technology. That is the whole method and it takes very little time. Three questions settle most claims you will encounter.

Ask whether the occupation is a single task or a bundle, and if a bundle, which parts are actually exposed. Ask how quickly somebody would notice a wrong answer and what it would cost. Ask who inside an organization would have to redesign the work and whether they have any reason to.

A claim that survives all three deserves to be taken seriously. Most claims do not survive the first one, which is why the record looks the way it does. The failures cluster exactly where the bundles are.

The trap in the reassuring version

None of this supports the conclusion that nothing changes, and that conclusion is as poorly supported as the alarming one. The occupations above survived and the work inside them did not. Continuity of the title concealed a complete change of content.

Somebody who was excellent at document review and indifferent at client management is worse off, even though paralegal is a stable occupation and the headline number looks fine. Aggregate survival is not individual security. The statistics and your position are different questions.

The reassuring reading and the alarming reading make the same mistake in opposite directions. Both treat the occupation as the unit, when the unit that matters to you is your own week. Occupations do not have careers and people do.

What the record cannot tell you

It cannot tell you that this time is the same. Language models are genuinely different from earlier automation in what they can do and in which occupations they touch, and the previous pattern is evidence rather than proof. Evidence is still considerably better than intuition here.

What the record does establish is that the three mechanisms exist and have mattered more than capability every previous time. That is a reason to apply them rather than a reason to dismiss the question. Applying a test is not the same as assuming an answer.

The honest position is that exposure is well measured, displacement is not yet observable, and the mechanisms that governed every earlier wave have not gone anywhere. Anybody claiming certainty in either direction is ahead of what can currently be known. Uncertainty is the honest description of the present moment.