Tech companies are firing engineers to teach the survivors to work for less. That is the sentence the industry would prefer you not articulate quite so plainly, but the documentary record does not leave much room for ambiguity. Software developer job openings on Indeed, through late May, sit roughly 70 percent below their 2022 peak. The Labor Department’s information-sector employment count dropped 332,000 — eleven percent — between its November 2022 peak and this May. Undergraduate enrollment in four-year computer and information science degrees fell 8.1 percent last fall, per the National Student Clearinghouse, reversing the 10.4 percent growth recorded in 2022. Graduates at this spring’s commencements were already booing the AI pep talks. The pipeline is reading the signal.
It is true — the phrase is doing real work here, not polite work — that some of this contraction traces to the post-pandemic correction. Tech firms overhired in 2020 and 2021, when every company with a data center believed it was a technology company, and some retrenchment was inevitable. The trouble is that the retrenchment has a specific shape, and that shape is not “we hired too many people.” The shape is: companies citing AI integration as the reason for workforce reductions, while the tools they claim are replacing those workers cannot, on the available evidence, do the jobs they are being credited with doing.
The workers understand what is happening. Christopher Pack, a 27-year-old with a master’s degree in computer science who landed a Bay Area job during the pandemic hiring boom, told the Wall Street Journal he feels like he “got on the last plane out of Vietnam.” He saves 65 percent of his after-tax income — a savings rate that would impress a Prairie farmer in the 1930s — so that if the job disappears, he can buy a house somewhere cheaper and “still be OK to coast into retirement.” He is 27. He is planning for a career that may not exist. The discipline is admirable. The necessity is an indictment.
Michael Waxman, 39, spent 2021 selling his homes and buying a catamaran that became a remote office. He was laid off early this year after months spent teaching AI agents to do much of his own work. “I am coding myself out of a job,” he used to joke. Waxman, still on the boat looking for contract work, compared the current market to “the farm hands and when tractors came into town.” The analogy is better than Waxman may realize, and also worse. The tractor worked. It plowed the field faster and cheaper than the horse, and the horse did not come back. The current generation of large language models writes code that looks correct, passes demo conditions, and then — and here it is worth being precise about what the technology actually does, because the public discourse has the misleading habit of treating it as a thing rather than a sophisticated pattern-matching system trained on the statistical residue of other people’s labor — produces the kind of subtle, structurally embedded errors that a junior developer would be embarrassed by and a senior developer would spend a week hunting.
Cory Doctorow, who has been more precise about the mechanics of platform decay than most commentators working the tech-policy beat, has been making the point for two years now: an AI cannot do your job, but an AI salesman can convince your boss to fire you and replace you with an AI that cannot do your job. The observation is structural, not cynical. It names a mechanism — what Doctorow calls the distinction between the centaur and the reverse-centaur. In the centaur configuration, a worker is assisted by a machine, and the combination functions better. In the reverse-centaur configuration, a human is pressed into service as a peripheral for a machine, running at the machine’s pace, monitoring its output for catastrophic error, and absorbing the liability when the machine gets it wrong. A developer who spends her day reviewing AI-generated code for subtle flaws — flaws a junior engineer would never have introduced but a senior one must now detect at scale — is not a beneficiary of automation. She is the accountability sink. The company gets the machine’s output and the engineer’s signature on it. The engineer, in a darkly comic inversion of the skilled-trades promise, gets to train her replacement before being shown the door.
The playbook is older than the technology. My father watched a version of it at a steel mill in the 1990s, when new owners bought the plant and laid off a significant part of the line. The new owners did not replace the workers with machines; they used the promise of new machinery as the justification for reducing the workforce and extracting more from those who remained. The millwrights who kept their jobs were expected to run faster, maintain more equipment, absorb more risk, and accept that the old contract — skill in exchange for security — had been unilaterally rewritten. The steel market did not demand this. Management chose it, because the balance of power permitted it, and because the language of modernization was available to launder a cost-cutting operation into a story about progress.
The tech industry’s version of the same operation is underway. Between February 2020 and February 2023, employment in computer-systems design and related fields rose by 11 percent. Companies hired aggressively, citing AI as a growth driver. They are now laying off many of those same workers, citing AI as the efficiency that makes them redundant. The AI did not change in the intervening period in a way that explains the pivot. The balance of power did. When engineers were scarce — when the four forces that Doctorow identifies as the constraints on enshittification included a genuinely tight labor market — companies competed for them. When the layoffs began, and the AI narrative offered a respectable cover, the same executives who had fought over every senior hire discovered that the engineers they had acquired could now be treated as a cost center. Competition for their labor has collapsed. Regulation of the tools being used to replace them is almost entirely absent — the tech workforce is overwhelmingly non-union. The scarcity that gave them leverage was never a permanent feature of their position; it was a condition management could revoke as soon as it had enough bodies, and enough of a pretext, to do so.
What remains is the fourth force — worker organization — and the data suggests it is not arriving. Undergraduate enrollment is falling, which means the next generation of engineers is reading the room and choosing other rooms. Noah Neustadt, a 37-year-old UX designer in Montreal earning around $120,000 a year, spends his spare time browsing farmland listings in British Columbia and Washington’s Olympic Peninsula. When he asked his AI chatbot where to live off the land, it suggested the Pacific Northwest. “It has the most fresh water,” he said. The absurdity is of a piece with the moment: the tool that is threatening his livelihood is also his real-estate agent for the escape.
This is the bezzle — J.K. Galbraith’s term for the interval between the commission of a fraud and its discovery, when the perpetrator has the gain and the victim does not yet know the loss. Companies are routing the productivity gains of the engineers they are keeping onto their balance sheets, while the laid-off workers are told they are the victims of a historical force no one could have foreseen. The force was foreseeable. It was being built, in many cases, by the very people it is now displacing. The decision to wield it against them was a management choice, not a law of nature. Galbraith himself observed that the magnitude of the bezzle always rises in a boom. We are, by the available indicators, in a boom. The question is not whether it is one. The question is who is left holding the bag when the engineering reality catches up with the marketing claim.
The response to a labor-discipline operation dressed as a technology transition is the response to any labor-discipline operation: collective bargaining that gives the workforce a seat at the table, labor-law reform that recognizes the specific vulnerability of knowledge workers under algorithmic management, and — this is the engineering-policy point — an AI-governance regime that requires documented proof of capability before a company can attribute a workforce reduction to automation rather than to the ordinary desire for cheaper labor. The European Union’s AI Act gestures in this direction. Nothing in the United States or Canada comes close.
There is a public consultation at no relevant government body about any of this, because the United States does not have a labor-market-disruption policy worthy of the name, and Canada’s is not much better. The workers who built the platforms that are now threatening their livelihoods are being told, in effect, that the market has spoken. The market has not spoken. The market has been spoken for, by firms with a documented interest in cheaper labor and a narrative that makes cheaper labor sound like progress. Pack is saving his money. Neustadt is looking at farmland. Waxman is on his boat. The work they did was real. The threat they face is, in substantial part, a bluff. The question is whether anyone in a position to call it will bother.