Silicon Valley is spending trillions to make your job obsolete and calling it innovation.
It is true, in the narrow and careful way that economists mean it, that technological revolutions have historically made societies richer in the long run — and that the young women who worked on telephone switchboards in the early twentieth century eventually became stenographers and waitresses. The trouble is that the medium run is a long time to be unemployed, the stenographer position is an artifact of a specific office layout that no longer exists, and the sixteen economists the Wall Street Journal convened this week to reassure you about artificial intelligence and the labour market have portfolios heavily exposed to the technology doing the displacing.
The Journal’s survey asked them what AI will mean for workers over the next five years. The consensus is precisely what the companies making the capital expenditures want you to hear: productivity will rise, offshoring to India and the Philippines will collapse first, and if you are currently processing insurance claims or translating standard ad copy you face genuine displacement risk. The advice, offered with the calm authority of people who do not have to apply for unemployment benefits, is that you should stop training to be a prediction machine and instead learn to be a tinkerer. Companies from Cisco to Meta are already citing AI as the reason for layoffs, streamlining operations on the basis of this exact calculus — replacing the prediction machines with a software update that costs a fraction of the severance package. Two-thirds of the economists surveyed expect productivity to rise; fewer than a fifth expect that to lead to more jobs.
Buried at the bottom of the Journal’s feature is the disclosure that matters: some of the economists surveyed disclosed relationships in the AI field, including involvement with startups, consulting work, and serving on company boards. It is not an indictment of their honesty to observe that their institutional incentives are aligned with the capital expenditure, not the displaced worker. The structural point is that the labour-market models they deploy — models in which displaced call-centre workers magically reconstitute themselves as interpersonal coordinators or judgment-heavy decision-makers — are built on the assumption that there is a job waiting at the top of the privilege gradient. But the current generation of large language models is being deployed precisely to eliminate the middle rungs of that ladder. The tinkerer the economist tells you to become is a role for which the firm has no headcount, because the firm has decided that the automated system does the tinkering, and the human absorbs the liability.
Here it is worth being precise about what “the algorithm” actually is, because the public discourse has the misleading habit of treating it as a thing rather than a continually-tuned set of weights serving a continually-revised objective function. The current generation of large language models are, at their core, next-token predictors — systems that have been trained on immense corpora of human-produced text and images to produce outputs that are statistically plausible continuations of a prompt. They do not understand the content they generate; they do not reason; they do not exercise judgment. They are, as Cory Doctorow has been at pains to point out, not a replacement for a human worker but a tool that a manager can use to discipline a human worker. “An AI can’t do your job,” Doctorow wrote, “but an AI salesman can convince your boss to fire you and replace you with an AI that can’t do your job.” The economist who described AI as “a revolution coming squarely at white-collar workers” was, in this sense, exactly right — and the force of the revolution is not in the machine’s capability but in the boss’s incentive. What the survey’s contributors call a productivity boost is the capital-intensification of the cognitive layer: a reverse-centaur, a human pressed into service as a peripheral for the machine, installed to absorb the blame when the automated system fails a basic safety check.
That incentive is the predictable result of removing the constraints that used to force companies to share productivity gains with the workers who produced them. For the four decades after the Second World War, a combination of strong antitrust enforcement, sectoral collective bargaining, and a regulatory state that was willing to punish firms that externalized their costs onto communities meant that when a company automated a production line, the surplus was distributed, however imperfectly, across wages, pensions, and public infrastructure. Those constraints have been systematically dismantled over the last forty years, and the AI boom is arriving in the vacuum they left behind. The hundreds of billions of dollars the hyperscalers are pouring into data centers are not being matched by a comparable investment in the people whose work those data centers are designed to render redundant.
One of the surveyed economists observes, perhaps unintentionally, that AI is doing cognitive work, and that this is a revolution coming squarely at white-collar workers, meaning they now know what blue-collar workers felt in the 1970s. The inheritance of extraction is visible in the sentence. The political economy that hollowed out Canadian heavy industry — the bar mills and auto parts plants that lost their margins to offshoring and private-equity financial engineering — is doing the exact same thing to the cognitive labour force forty years later. The playbook is older than the mechanism. The mechanism is the large language model and the API; the extraction of the surplus that the existing workforce has built, the locking in of the remaining workers by raising the cost of leaving, and the transfer of value to the shareholders is the identical playbook that an older generation of labour economists called asset-stripping.
There is a second layer of bad-faith technique embedded in the telephone-switchboard analogy, which the economists reach for to prove that technology has always been good for workers. The structural difference between the rotary phone and the current wave of AI deployment is not merely the speed of adoption; it is the monopsony structure of the buyers of cognitive labour. In the early twentieth century, a displaced telephone operator could walk down the street and find a dozen different employers who needed a pair of hands. Today, the consolidated platform layer means that the workers displaced from the routine cognitive layer are not walking into a competitive market with a surplus of alternative buyers; they are walking into a labour market where the same capital expenditure that displaced them has also automated the hiring pipeline, compressed the interview process, and raised the experience threshold for the entry-level position they might have taken. A Harvard economist recently warned that the result could be a permanent underclass — except that the underclass here is not a policy failure but the intended yield of a capital structure that no longer requires a mass cognitive workforce.
The most honest sentence in the Journal’s survey is the one that admits the institutional reality: the United States does not have a glorious history of sharing the gains of technological change or redressing the costs, it is currently unprepared, and the signs from Washington and Silicon Valley are to let it rip and damn the consequences. This is the accurate read. The United States has no sectoral bargaining mechanism to negotiate the deployment of artificial intelligence, no statutory co-determination requirement that forces management to consult with workers before automating the claims-adjustment desk, and no active-labour-market policy that retrains the fifty-year-old paralegal into the judgment-coordinator role the economist promises is waiting. The economists the Journal surveyed are emphatic on this point: “Whether those gains are broadly shared, and whether the workers whose careers are destroyed make successful transitions, these things do not depend on the technology. They depend on the societal institutions and policies we build.” And then, with the candor of someone who has seen the balance of power in Washington and Menlo Park up close, the same economist adds: “We are currently unprepared.”
The economist who predicted that the workers who lose out will get “very, very angry and change the politics” is probably right in the medium term. But anger is not a policy, and the politics it changes can as easily produce a xenophobic strongman as a renewal of the regulatory state. The alternative — the affirmative program that would actually make AI a tool for shared prosperity rather than an engine of unilateral capital consolidation — is hiding in plain sight in the Journal’s survey: an education system that values flexibility over credentialing, a labour-law regime that gives workers bargaining power over the deployment of automation, and a tax system that captures the enormous productivity gains the hyperscalers are booking and redistributes them to the communities that will absorb the displacement. The economists offered the map. The hyperscalers have the treasure. The capital does not require consent to reconfigure the floor plan of work, and every signal from Washington and Menlo Park says they intend to keep both.