For about seventy years, the word computer named a job rather than a machine. It was held by people, most of them women, who performed arithmetic in organized teams for observatories, insurance offices, artillery ranges, and eventually spacecraft. The category was created by a technology that generated more numbers than anyone could handle, and it was ended by a technology that handled them. Neither event was about the ability of the people involved. The occupation existed in the gap between what a machine could produce and what a machine could process, and it lasted exactly as long as the gap did.
The category the gap creates
Some jobs are not displaced by a technology. They are manufactured by its immaturity and then withdrawn by its completion. This is a narrower claim than the general one about automation, and the narrowness is what makes it useful. A blacksmith was displaced by a technology that had nothing to do with blacksmithing. A telephone operator was created by the telephone and ended by automatic switching. The same system produced the job and removed it, and it did both for the same reason, which is that the system was incomplete in a specific place and then was not.
Call this scaffolding labor. It has a recognizable profile. It appears suddenly and in volume. It is accessible relative to the sophistication of the system it supports, because the whole point is to do at scale what the system cannot yet do at all. It is often the entry point into an industry for people the industry would not otherwise have hired. And it is load-bearing right up to the moment it is not.
The Harvard College Observatory started employing women as computers under Edward Pickering from the 1880s, to classify stellar spectra arriving faster than the observatory's astronomers could handle. They were paid a fraction of what a male assistant earned, and the work was understood as clerical. Williamina Fleming developed a classification scheme. Annie Jump Cannon built the system still in use for stellar classification. Henrietta Swan Leavitt found the period-luminosity relation in Cepheid variables, which is the measuring stick that established the scale of the universe and made Hubble's work possible.
That is the second thing worth noticing about scaffolding labor. The work is categorized as routine because the system needs a great deal of it, cheaply, and the categorization is frequently wrong about what the people doing it are actually capable of.
- Pre-1450Professional copyistsUniversity growth creates sustained demand for hand-copied texts. The trade expands with the institution it serves and is ended by movable type.
- 1840s onwardTelegraph operatorsA skilled category created entirely by the telegraph, requiring years to reach speed. Dissolved by teleprinters and the telephone.
- 1880sThe Harvard computersWomen hired to classify stellar spectra arriving faster than astronomers could process. Cannon and Leavitt produced foundational astronomy from inside a job classified as clerical.
- 1878 onwardTelephone switchboard operatorsManual connection of every call. Became one of the largest employment categories for women in the United States, then was removed by automatic exchanges over several decades.
- 1886Linotype and the compositorMechanical typesetting creates a skilled, unionized, well-paid trade that lasts a century and is ended by phototypesetting and desktop publishing.
- 1935 to 1960sThe NACA and NASA computing poolsHuman computers at Langley, including the segregated West Area unit. Katherine Johnson was asked to verify the electronic computer's trajectory figures for John Glenn's 1962 flight, which describes the transition exactly.
- 2010s onwardData annotation and evaluationA category created by machine learning's appetite for labelled examples, currently at volume, currently accessible.
The compression
The end is rarely a collapse. It is a compression, and what survives it is different in kind from what preceded it. Automatic telephone exchanges did not eliminate every operator at once. They removed the common case, which was connecting a local call, and left the exceptions: emergency calls, complex routing, assistance. The residual role required more judgment, employed far fewer people, and was less accessible than the job it descended from.
The Glenn detail captures the transition better than any statistic. By 1962 the electronic computers at NASA could produce orbital figures, and the figures were not yet trusted. Katherine Johnson was asked to run the numbers by hand and confirm them. At that moment the human computer's job had already changed from producing the answer to certifying it. That is what the remnant phase looks like: the same person, a smaller category, and a task defined by the machine's residual unreliability rather than by its absence.
Each transition is driven by the same system maturing. Nothing external needs to arrive.
What survives compression
Three properties predict which parts of a scaffolding category persist, and they are consistent across every case above.
Judgment about edge cases survives. The machine handles the distribution's centre and fails at its tails, and the tails are where the remaining work is. This is why the residual role is always harder than the average case of the original job, and why headcount and difficulty move in opposite directions.
Verification survives longer than production. The last thing a maturing system gives up is a human confirming that its output is correct, because trust in the system lags its actual reliability by years. Johnson checking the trajectory is the canonical example, and it is a durable position for exactly as long as the trust gap persists.
Taste survives, and is never described as a skill until it is scarce. Compositors were paid for typesetting and valued, in the end, for knowing what a page should look like. That knowledge was invisible while the mechanical part of the job dominated, and became the entire job once the mechanical part left.
The counter-case, stated at full strength
The historical record does not support a straightforward story of destruction, and the strongest evidence against the pessimistic reading comes from the same cases.
Movable type ended the professional copying trade within a few decades. It also created typefounding, presswork, composition, proofreading, publishing, bookselling, and eventually a reading public large enough to sustain professional authorship. The number of people employed making books rose enormously. The specific job of copying by hand disappeared, and the industry it belonged to grew by orders of magnitude.
James Bessen documented a sharper version of the same effect in modern banking. After automated teller machines were introduced, the number of bank tellers in the United States did not fall as expected. It rose, because ATMs made branches cheaper to operate, banks opened more of them, and the teller's role shifted toward sales and relationship work. The machine took the transaction and the job absorbed something else.
Both cases share a mechanism worth stating explicitly: automation lowered the cost of the output, demand for the output was elastic, and the volume increase outran the labor saving. Where that condition holds, employment rises. Where it does not, it falls. It is an empirical question about a specific market, not a matter of temperament, and the honest answer for any given case is usually that nobody knows in advance.
- 1GrewMovable type against copyistsThe specific task vanished. Total employment in book production rose by orders of magnitude as the cost of a book collapsed and demand expanded.
- 2GrewATMs against bank tellersBessen's finding: cheaper branches meant more branches. The role changed rather than disappearing.
- 3ShrankAutomatic exchanges against operatorsCall volume rose enormously and employment still fell, because connecting a call was fully automatable and demand for human connection was not elastic in the relevant sense.
- 4ShrankDigital typesetting against compositorsPublishing volume grew while the trade was eliminated. The skill was absorbed into software and into everyone else's job.
- 5EndedElectronic computers against human onesThe category closed entirely. Individuals moved into programming and analysis, which is the clearest case of the people persisting while the occupation did not.
Where data work sits
Data annotation and evaluation is scaffolding labor by the definition above. The category did not exist at scale twenty years ago. It was created by a technology that required enormous volumes of labelled examples, it grew quickly, it is accessible relative to the systems it supports, and it is load-bearing right now.
The structural question is which phase it is in, and there is a reasonable argument that the answer is already changing. The early demand was for volume, because the binding constraint was quantity of general data. As systems become competent on common cases, the value of another common-case example approaches zero and the binding constraint moves to the cases the system still gets wrong. Those cases require someone who genuinely knows the domain, because identifying a subtly wrong answer in medicine, law, or advanced mathematics requires being able to produce the right one.
That is the compression, arriving in the usual shape. Fewer people. Higher bar. Better paid. And a category that was an accessible entry point becoming something that requires credentials in the domain, which is close to the opposite of what it was.
Volume phase
- ConstraintQuantity
- Worker profileGeneral
- Barrier to entryLow
- HeadcountLarge
- Rate per unitLow
- TaskProduce labels
Selectivity phase
- ConstraintDifficulty
- Worker profileDomain expert
- Barrier to entryHigh
- HeadcountSmall
- Rate per unitHigh
- TaskJudge and correct
We do not exempt ourselves from this. Kariaa operates data work, and the analysis above describes our own position as much as anyone's. Treating it as a permanent category would be a forecasting error. Treating it as worthless because it is transitional would be a different error, since the transitional phase of every previous case was where a large number of people entered an industry that would not otherwise have taken them.
What this implies for anyone inside it
The pattern does not predict a date, and anyone claiming otherwise is selling something. It does predict a shape, and the shape has practical consequences.
The peak is not a signal of durability. Every one of these categories was at maximum employment and maximum apparent stability shortly before compression. Volume of demand tells you about the current gap, not about how long it stays open.
The exit is into the system, not away from it. Human computers became programmers and analysts. Compositors who understood layout became designers. Operators moved into the telephone companies' other roles. In each case the people who transitioned well were the ones who had learned how the surrounding system worked while doing the routine part, and the ones who had treated the routine part as the whole job were the ones stranded.
The remnant is worth aiming at. Verification, edge cases, and judgment are where the compressed category concentrates, they pay better than the volume phase, and they are visible in advance because they are the parts of the work that are currently annoying rather than repetitive.
The scaffolding metaphor holds in the way that matters. Scaffolding is not fake work. Nothing gets built without it, it carries real weight, and the people on it are doing something difficult. It also comes down when the structure stands, and the ones who did well out of it were the ones who spent the time learning the building.
Key takeaways
- Scaffolding labor is created by a technology's immaturity and removed by its completion, which is a narrower and more predictable pattern than general automation.
- The Harvard computers and the NASA computing pools produced foundational work from inside a category classified as routine.
- Compression, not collapse, is the usual end state: the common case automates and a smaller, harder, better-paid remnant survives.
- Verification outlasts production, because trust in a system lags its reliability. Katherine Johnson checking the computer's figures is the canonical remnant role.
- Employment can rise after automation when demand for the cheaper output is elastic, as with print and with bank tellers in Bessen's work. It falls when it is not.
- Data work shows the transition already: the constraint is moving from quantity of general data to difficulty, which implies fewer people at a higher bar.
- Peak employment in a scaffolding category has repeatedly preceded its compression, so volume of current demand says nothing about duration.
This report was produced by Kariaa Research. All data, resources, and analysis are proprietary. For questions, contact research@kariaa.com.