Numerical Control Didn't De-skill Machinists; Average Skill Barely Moved
Since the 1950s, people have feared that numerical control would de-skill machining trades. But the hourly wages of A-grade machinists and NC operators were nearly identical, and over three decades overall skill fell only about 1%.
The argument · tap a timestamp to hear it
A machinist's job isn't just cutting metal
A machinist doesn't just operate a machine tool to cut metal. They work with designers to read two-dimensional drawings, reconstruct the three-dimensional part in their head, and spot what can't be made ahead of time, which is why geometry, trigonometry, and algebra come into it. What's truly hard to replace is feel: the 1887 book The Complete Practical Machinist said lathe work is most demanding in feed and speed, the most varied, complex, and deceptive part of the whole trade, requiring judgment, perception, and alertness—the tool may sound and look fine while actually drifting off, even producing scrap or injuring someone. An experienced hand, using gauges plus feel, can detect dimensional changes as small as one half thousandths of an inch. Skill here equals judgment, and that is the bullseye of the later de-skilling debate.
Four years and eight thousand hours to become a journeyman
In the first half of the twentieth century, machining was the largest single skilled manufacturing job for American men, and the standard entry path was a formal apprenticeship: generally four years, eight thousand hours; tool and die makers five years, ten thousand hours. Someone who apprenticed at GE in the 1940s recalled spending the first half year learning basics in the shop, with classes three evenings a week in shop math, college physics, blueprint reading, and electricity; then three and a half years rotating through various tools, starting at the drill press and switching about every half year. On completion the company certified you as a journeyman, issued your first set of tools or a cash bonus, and raised your pay—in his words, you were finally qualified for marriage. But this was never the only path: some became self-taught through pure shop experience and home study, and there were courses claiming to teach you how to operate a machining shop, which you could call the 1930s coding bootcamp.
Riveting a boiler: nine machinists become one
Industrialization repeatedly does the same thing: put skill into the machine, then break the work into pieces. When Edison started making light bulbs in 1880, the factory was full of skilled craftsmen who understood glass and wire and knew from scratch how to make a bulb; as specialized machines built their skills in, the work was simplified and segmented, and a few all-around craftsmen were replaced by a large group of workers each knowing one operation. Riveting a boiler originally took nine skilled machinists; with a hydraulic riveter it became one skilled worker plus eight unskilled ones, the latter doing tasks like clamping. In 1883 the young machinist John Morrison testified before Congress that the machining trade had been subdivided since the mid-1870s, and subdivided again after that, so that in sewing machine making, for example, the jobs were split so finely that a person hardly counted as a machinist. NC is just one more step on this timeline.
Bosses hoping NC would save labor were disappointed first
NC controls the machine tool and workpiece with pre-programmed instructions, without requiring direct manual intervention by a machinist: a programmer translates the drawing into coordinates, punches them onto paper or plastic tape, and feeds them into the controller. It was invented in the 1950s, but at first served only the Air Force and aerospace, with specifications far beyond the needs of ordinary shops and small workshops. Crucially, early NC was hard-wired, with limited functions, and a major change meant tearing apart the controller and rebuilding it—very inflexible. So shop owners hoping to lower worker skill with it were thoroughly disappointed: a mid-1960s survey by Lawrence Williams and Brian Williams found that shops still needed operators with the same skills and technical knowledge, especially right after introduction; owners thought the risk of NC breaking down and causing expensive downtime was too high, so they assigned their best employees to run the machine. Only when it matured did the advantages show—one NC drill press was estimated to replace three conventional machines.
Training shrank from five years to half a year
In the 1970s computer technology accelerated NC's spread, and by 1980, 25% to 30% of machine tools sold in developed economies had NC. The real de-skilling evidence came from training time and headcount: a 1982 paper studying Swedish machining shops noted that NC operators needed at most 6 to 12 months of training, while lathe operators had to grind through 4 to 5 years in the shop. In the cases the paper cited, one plant needed only 22 NC operators instead of 44 lathe operators; another replaced 63 qualified machine operators with 21 NC operators. In a developed country like Sweden, skilled workers willing to work night shifts were hard to find, and NC operators accepted stranger hours, so the machines didn't have to stop. As for whether this was lower skill, nearly 80% of a 1980s survey of Canadian manual machinists agreed NC required less skill, but only half of those who had actually run NC saw it that way.
NC operators' hourly wage differed by only twenty-one cents
Wages are a crude ruler for skill, but the result is counterintuitive. A 1989 study using 1981 U.S. Bureau of Labor Statistics data compared the average wages of 80,000 A-grade machinists and NC operators in the machine tool industry: A-grade machinists earned $9.72 an hour, NC operators $9.51. For comparison, A-grade machinists still earned more than B- and C-grade machinists doing more repetitive work ($8.54 and $6.41); interestingly, there is evidence that some NC operators were recruited from B- and C-grade machinists, meaning they were actually the ones getting a raise. Of course wages are affected by region, unions, and so on, and don't reflect job quality or the number of positions. Bill Bowman is the other side: as more people learned the skills he had, his CNC premium flattened, and a few years later, combined with layoffs, he was making only $20,000 a year.
Over thirty years overall skill fell only about 1%
Jeffrey Keefe's 1991 study examined three U.S. Bureau of Labor Statistics surveys spanning 30 years and found overall skill levels fell by about 1%—essentially zero. That number directly refutes the claim that numerical control de-skilled machinists. But the author also stressed that the unchanged average masked changes in job composition: the work of skilled setup workers, semi-skilled machine operators, and unskilled workers changed or disappeared, while other tasks migrated to NC operators, and underneath it was very turbulent. So the more accurate statement is not that skill disappeared, but that its form changed—deep mastery of one thing was weakened and replaced by several other kinds of skill, such as understanding how the whole machine runs and reconstructing the logic of a failure, which is why operators say you can't just go tear into a fault, you have to work backward to how the accident happened.
AI coding is replaying this history
What the author sees is a parallel between AI coding and machining back then: early GitHub CoPilot and early Cursor did only simple completion or inserting a line, but now AI coding systems have evolved into agentic—while modern CNC machine tools have not reached that step. He specifically mentions a February 2026 case study by OpenAI's Ryan Lopopolo about a brand-new product built by a team of agents, with humans monitoring and steering. Programming, like machining back then, is one of America's dominant high-paying jobs; some things don't change—programmers still have to read code and judge where the problem is, only when it's time to fix it you tell the coding agent instead of opening a text editor, the same skills expressed differently. Some jobs will be de-skilled and decline, new fields may appear, and those are the NC operators of the AI coding era. What he remembers is Bill Bowman: you have to keep learning.
In their own words · checked verbatim
for there is no part of the turner's art in which so great a variation of practice exists or is possible, no part of his art so intricate and deceptive, and none requiring so much judgment, perception, and watchfulness
Several years and layoffs later, he was pulling down $20,000 a year or about $9 an hour despite doing more CNC programming work.
In our work you mustn't go directly to the breakdown. You have to think about it, retrace the development of the incident, identify the signs of failure. You have to find out what didn't work and not just push any old button. You have to understand and reconstitute the logic
The drill is working a little bit too hard to cut this piece of metal. It's gonna let you know. It's just not gonna sound right. You're not gonna get the correct kind of finish. So, then you make a correction in the program. You either run it slower, or make the spot bigger.
So perhaps those manual machinists are being a bit narrow-sighted in calling NC operators "less skilled". NC and CNC creating skill change doesn't mean operators became any less "skillful", if you get my driftwood.
One thoughtful 1991 study by Jeffrey Keefe studied three surveys done by the Bureau of Labor Statistics spanning 30 years and found that skill levels declined about 1% in aggregate. Essentially zero. But this unchanging average papers over the fact that the composition of the jobs certainly changed.
Early AI coding tools like GitHub CoPilot and early Cursor worked on simple autocomplete or line insertions. Then in recent days, AI coding systems have evolved to being "agentic" - a thing that modern CNC machines have yet to achieve.
For my part, the takeaway that most sticks with me are the experiences of the aforementioned Bill Bowman. You have to continually keep learning.
Figures
| Machining apprenticeship | About 4 years, 8,000 hours | 3:06 |
| Tool and die maker apprenticeship | 5 years, 10,000 hours | 3:06 |
| Manpower to rivet a boiler | 9 skilled machinists → 1 skilled worker + 8 unskilled workers | 6:10 |
| Conventional machines replaced by one NC drill press | 3 | 9:15 |
| Increase in machine production time from NC in batch production | 4-5x | 10:19 |
| Share of machine tools sold in developed economies with NC in 1980 | 25-30% | 11:19 |
| NC operator training time vs lathe operator | At most 6-12 months vs 4-5 years | 11:19 |
| Share of Canadian manual machinists who thought NC required less skill | Nearly 80% | 12:19 |
| 1981 hourly wages of A-grade machinists and NC operators | $9.72 vs $9.51 | 15:21 |
| Overall skill decline measured by Keefe 1991 | About 1% | 18:25 |
Glossary
- Numerical Control (NC)
- Controlling a machine tool and workpiece with pre-programmed instructions, without requiring direct manual operation by a machinist.
- CNC
- The upgraded version of NC after microprocessors appeared in the 1970s, in which instructions can be generated and edited on a computer.
- journeyman
- A person who has completed a formal apprenticeship and been certified by the company as a qualified machinist.
- agentic
- An AI system able to complete multi-step tasks on its own, rather than just completing a line of code.
How to listen
Investors watching automation and employment structure, and manufacturing engineers; also people building AI coding tools who wonder whether programmers will retrace the machinists' path.
The hardware details of early hard-wired NC and paper-tape control can be skipped.