White Collars Turn Gray: AI and the Repricing of Work

What follows is a riff on Paul Krugman’s 1996 essay “White Collars Turn Blue,” updated for the AI age and written, suitably enough, with a little help from AI.

Looking back from 2126, what is most striking about the early age of artificial intelligence is not that people expected too much. It is that they expected the wrong things.

The mistake was not a failure to foresee machine intelligence. It was the assumption that it would arrive evenly. As AI moved into documents, screens, forms, code repositories, financial models, and slide decks, commentators and startup pitch decks kept imagining humanoid workers, autonomous delivery fleets, and robot service staff.

They had reasons. The technological dream of the industrial age had long been the conquest of manual toil. Steam replaced muscle. Electricity reorganized factories. Engines transformed transport. Machines steadily took over lifting, hauling, digging, moving, sorting, and making. So when artificial intelligence began to accelerate, many assumed it would complete that project.

By the late 2020s, however, an awkward fact had become impossible to ignore. The robots were still fumbling with door handles, while software was already absorbing a disconcerting share of the tasks on which professional prestige depended. The earliest and deepest effects of AI were felt wherever work took place inside screens, documents, forms, and systems that could be transmitted, recombined, and checked without ever touching the stubborn physical world.

It turned out that reality has friction, while language does not.

In retrospect, this should have been obvious. The physical world is full of edge cases. Floors are slippery. Pipes are hidden behind walls in unpredictable places. Human bodies are fragile, heavy, uncooperative, and legally protected. Customers are rude. Dogs bite. Weather changes. Battery life runs out. Stairs remain stairs. Every cheap household task that techno-optimists in 2026 imagined would soon be done by robots turned out to require a formidable combination of balance, dexterity, perception, judgment, social tolerance, and mechanical robustness. A machine that can write a competent legal memo in three seconds is not therefore able to carry a hot water heater down a narrow staircase without denting the wall, injuring someone, or getting sued.

Meanwhile, the white-collar world had quietly made itself legible to machines. Its work product already existed in digital form. Its intermediate steps were recorded in text. Its standards were increasingly measurable. Because so much of it lived inside software, it could be reproduced, tested, and iterated at extraordinary speed. Digital environments could be spun up endlessly. Tasks could be replayed. Performance could be scored. Iteration was cheap. Reality, by contrast, had to be encountered one stubborn instance at a time, or else simulated at great expense and with imperfect fidelity. Long before a robot could reliably clean a cluttered teenager’s bedroom, an AI system could summarize ten research reports, draft a board memo, produce six passable marketing concepts, compare contract versions, and write software that once would have occupied a senior engineer for days.

The surprise was not that AI could do these things. The surprise was how many career ladders had been built around doing them.

Junior analysts became investors by building spreadsheets and writing research notes. Associates became senior counsel by drafting, reviewing, and revising contracts. First-year engineers became senior ones by writing code, testing prototypes, producing technical drawings, and troubleshooting minor problems.

This was more than a labor-market disruption, because these were not merely useful tasks. They were also bound up with status. Advanced societies had come to treat certain kinds of cognitive performance as evidence of merit: verbal fluency, abstract reasoning, polished presentation, comfort with symbols. Parents urged children toward the professions, toward coding, finance, consulting, law, design, management, media, and administration: toward clean jobs in climate-controlled rooms, jobs in which one manipulated representations rather than objects.

And then, with cruel speed, these became among the first qualities to be cheaply replicated.

This did not mean that elite work disappeared. Far from it. Many of the winners of the AI transition were highly capable people. But their advantage was no longer just intelligence. It was leverage: control of systems, judgment, responsibility, taste, clients, capital, distribution, and trust at the point where machine output met reality. The analyst who merely produced a report was in trouble. The investor who knew which report mattered, and what to do about it, was not. The junior lawyer who assembled a brief lost bargaining power. The senior rainmaker who could calm a client, frame a negotiation, and stake a reputation did not. The coder who translated tickets into syntax saw that labor commoditized. The engineer who could define a product, orchestrate machines, and own a business problem prospered.

The premium shifted, then, not exactly from brains to brawn, but from performing cognitive labor to directing it.

Unfortunately, many people imagined a simple blue-collar renaissance. As white-collar labor cheapened, they assumed the economy would smoothly rebalance into trades and hands-on service work. But embodied occupations were constrained not only by demand, but by supply.

To say that a job cannot be automated is not the same as saying that anyone can do it.

A society that had spent decades preparing millions of people for sedentary, screen-based employment had quietly produced a workforce that was often neither physically prepared nor temperamentally suited for strenuous embodied work. The jobs that remained stubbornly human often required not merely a body, but a certain kind of body: strength, coordination, stamina, pain tolerance, spatial sense, willingness to wake early, willingness to work in heat or cold, willingness to confront dirt, disorder, risk, and demanding customers. They also required a certain kind of mind: patience with repetition, calm amid minor chaos, practical judgment under uncertainty, comfort with direct consequences. Many educated professionals who airily supposed they could “always become a plumber” were, in fact, no more suited to plumbing than plumbers were to writing strategy decks.

This is one reason the revaluation of embodied work was uneven. Not every manual job became lucrative. Many remained badly paid, because institutions are perfectly capable of undervaluing socially necessary work for long periods. But the subset of physical occupations that combined skill, certification, trust, local scarcity, and unpleasant realities did better than the old prestige hierarchy would have predicted. Good electricians, mechanics, lineworkers, heavy-equipment operators, carpenters, welders, plumbers, equipment installers, field-service technicians, and innumerable others who could do useful things in the real world discovered that they possessed something rare: capabilities that software could not instantly flood.

In the early AI age, embodied occupations offered something increasingly scarce: a tolerable bargain between effort and scarcity. A young person could acquire a skill that machines could not yet cheaply imitate, earn a living not immediately exposed to global digital competition, and accumulate practical knowledge, reputation, repeat customers, and savings. The most successful then converted that protection into a business, a client base, property, or some other form of ownership. The blue-collar revival was not about romanticism. It was about one of the last remaining ladders from labor into ownership.

This was especially true because globalization, which had exposed so much white-collar work to international competition, had less purchase here. You can outsource a spreadsheet model to another continent. You cannot outsource a burst pipe in your basement, the rewiring of a school, the installation of a heat pump that arrived at the loading dock this morning, or the care of a panicked elderly patient.

So the labor market of the late 2020s and 2030s began to display a pattern that would have seemed perverse to the age that preceded it. Some of the brightest graduates from prestigious institutions found themselves in crowded tournaments for a shrinking number of elite cognitive roles, while reliable, sober, physically capable people with unglamorous practical skills often enjoyed more immediate bargaining power. The educational ladder did not disappear, but it no longer carried the same universal promise. A higher degree remained useful as a positional signal and as preparation for certain high-end roles. It was simply no longer a secure claim on scarcity.

There were, of course, many denunciations of this state of affairs. Some insisted that the machines were not really intelligent, as if that settled the matter. But markets do not pay for metaphysics; they pay for substitution. If a machine can produce a first draft that is better than yours in one second, the philosophical question of whether it “understands” matters less than people hoped. Others predicted that robotics would soon equalize matters by displacing physical work as well. In the very long run this proved correct, as such predictions often do. But “soon” did a great deal of work. Decades passed in which software intelligence advanced much faster than practical machine embodiment. During those decades, the asymmetry reshaped education, class identity, migration, politics, and family strategy.

The real cultural shock of the AI age was that it unsettled a comforting belief. Advanced societies had come to assume that progress naturally favored abstract, intellectual, and physically clean forms of work. Instead, the first mature wave of AI made something awkwardly clear: some of the tasks most closely associated with education and status were also among the easiest to reproduce. What remained scarce was often messier, more local, more physical, more interpersonal, and less prestigious.

By 2126 this no longer seems shocking. We have had a century to absorb the lesson. The economy does not necessarily reward what a civilization finds most flattering about itself. It rewards what remains scarce under the technologies of the age.

And for a surprisingly long time after artificial intelligence became cheap and ubiquitous, machines found it easier to mimic the mental work we most admired than to manage the small physical competencies of everyday life.