The plumber and the programmer

The jobs most exposed to AI are among the best paid in America: the typical wage in the government's top tier is $78,105, by our arithmetic, against $49,120 in the lowest.

"Learn to code" has split in two: the government projects computer programmers to shrink 7.3 percent over the decade and software developers to grow 10.2 percent.

The hands-on trades sit near the bottom of every exposure list, and the tier that holds them and care work is projected to grow fastest of all.

For thirty years the advice was: learn to code. Now everyone asks, is my job safe? For two years the answers came from opinion: a chief executive's forecast, a consultant's slide, a viral post.

So we went looking for an answer that was not opinion. On 27 August 2026 the US Bureau of Labor Statistics, the government's jobs statistician, rated every occupation it tracks, 831 of them, by how exposed each one is to AI. Exposed means how much of a job's tasks AI can do, or is already used for. The Bureau combined five measures: two from the companies whose AI people use, Microsoft and Anthropic, and three from academic teams. Then it set those ratings beside its own ten-year job projections.

About 31 percent of American jobs, some 53 million, sit in the Bureau's highest tier, "Very high." Take the typical wage of each job in that tier and find the middle one: in our arithmetic on its table, it is $78,105. In the lowest tier it is $49,120.

This is not the factory story of the last century. It is office work: jobs whose output is text, numbers or code, produced at a desk. Clerks, customer service, data entry, writing and translation, accounting and programming.

The exposed workers are, on average, better paid and better educated. Anthropic found the most exposed quarter of workers earn 47 percent more than the least exposed. They are almost four times as likely to hold a graduate degree.

Programming is the sharpest case, because it splits in two.

"Computer programmers" write code to someone else's design. They sit at or near the top of every exposure list. In Anthropic's data on how people actually use its AI, they are number one: 74.5 percent of their tasks show real use. The Bureau projects their numbers to fall 7.3 percent over the decade.

"Software developers" design the systems. The Bureau rates them just as exposed. But it projects them to grow 10.2 percent, nearly three times the average for all jobs. It gives AI itself as one reason for the demand.

So the same technology means losses for the narrow coding job and growth for the wider design job. The person who was told to learn to code, and learned exactly that, is in the wrong half.

One occupation shows what an exposure score looks like when it comes true, and why the layoff counts miss it. Translators.

An Oxford study by Carl Frey and Pedro Llanos-Paredes looked at US data from 2010 to 2023. In regions where people used Google Translate more, translator jobs grew more slowly. Frey estimates roughly 28,000 more translator jobs would have been added without machine translation. He told CNN it is not, so far, a story of mass displacement, though he thinks that follows.

In Britain, more than a third of translators told the Society of Authors in 2024 that they had lost work to generative AI, the kind that writes text on request. Forty-three percent said their income had dropped.

CNN met one translator, working in a rare language, who had years of steady European Union work. When it dried up, he lost about 70 percent of his income. What is left is polishing machine output. He refuses it, because the corrected text is fed back into the software. You are, he said, "essentially expected to dig your own professional grave."

Nobody was laid off in that story, and the work is gone all the same. No layoff notice, no company filing and no tally of announced cuts will ever count it.

At the other end sit the jobs done with hands and bodies, in a physical place. Plumbers rank between 613th and 748th on every exposure list, near the very bottom. Electricians show no measurable use at all in Anthropic's data.

The Bureau projects electricians to grow 9.2 percent, and names the power demand of data centres as a reason. It projects plumbers to grow 6.8 percent, and heating and cooling mechanics 10.9 percent.

There is one catch. The Bureau does not put the trades in its lowest tier. It rates them "Moderate," because the paperwork, estimating and diagnosis in the job register on the theoretical measures. The hands-on work is safe. The quote and the invoice are not.

The "Moderate" tier, home to the trades and to care work, is projected to grow fastest of all: 6.2 percent, against 2.0 percent for "Very high." Home health and personal care aides alone, rated "Moderate," are projected to grow 18.1 percent.

The public got there first. In one week of posts on X to 29 October, in English and not counting reposts, the trades paired with AI were the most common pairing of all, about 1,900 posts. The electrician, the plumber and the welder were held up as the work a machine cannot take. "Learn to code," the advice a generation was given, appeared about 1,100 times, much of it now said with bitterness. These counts are a signal of the public mood, not a measure of jobs.

Business leaders said it too. At the Aspen Ideas Festival in late June 2025, Ford's chief executive, Jim Farley, called the trades the "essential economy." In the same breath he forecast that AI is going to replace "literally half of all white-collar workers" in America. That is a forecast, not a count.

The trades story has its own rebuttal built in. In 2026 Meta, Google and BlackRock put more than $265 million, as reported, into training electricians and builders for the data-centre boom. The president of North America's Building Trades Unions called Meta's part of it "a brilliant public relations move."

Meta's course is four weeks. The electricians' union apprenticeship is five years and 10,000 hours. Only about 45 percent of registered apprentices finish.

The viral figure that Gen Z trade-school enrolment rose 1,421 percent is an artefact, a quirk of how it was counted. It measures a generation ageing into eligibility from almost zero. Real trade-school enrolment grew about 0.4 percent a year from 2017 to 2023. What does hold up is the apprentices: construction apprentices passed 480,000 in 2025, up 28 percent in five years.

And retraining, the usual answer for displaced office workers, has a weak record. A review of 56 randomised trials in the United States, studies where chance decides who gets the training, found that job training raised employment by 1.7 percentage points, from a base of 63 percent. It raised earnings by about $800 a year. It cost about $13,000 and six months per person. Its authors wrote that "it seems doubtful that we currently have programs capable of meeting the moment." The review was published by Anthropic.

Even the most cheerful voices hedge. Mike Rowe, the trades' best-known champion, said AI is coming for the coders and "not yet" for the welders.

The best direct test of what AI can actually do comes from Epoch AI, a research group that studies AI progress. In October 2026 it gave six leading models real tasks from its own work, and graded them against its own staff's standards.

The two leading models handled the well-defined parts reliably: the coding, the calculation. None could do the work end to end. They missed the judgment, the house style, and the sense of what matters to a reader.

That fits the split in the numbers. Well-defined work at a screen is exposed. The judgment, and the work done with hands, are not. Not yet.

There are two ways to read this, and the evidence supports both.

The first is a squeeze on the desk. Almost a third of American jobs sit in the top tier, and they are the better-paid ones. The narrow coding job is projected to shrink. And the translators show the loss can come without a single layoff, which the usual counts miss.

The second is that exposure is not loss. Software developers are rated just as exposed, and they are projected to grow. Frey says the translator story is not, so far, one of mass displacement. And in Epoch's test, no model could do the work end to end. On that reading, AI takes tasks, and the people who hold the judgment keep the job.

Either way, none of this says "panic." If your job is mostly producing text, numbers or code to someone else's specification, it is exposed, and the government says so. Its judgment, design and dealing with people are the parts to grow. If your job is done with your hands, in a place, it is among the safest there is. But the paperwork around it will change, and a four-week course is not a trade.

And the folklore is wrong in both directions. The machines are not coming for everyone. They are not coming for the factory floor first. They are coming for the desk.

Our ask is simple. Look at what your work produces, not what it is called, and put your effort into the part of it a machine cannot yet do.