This Industrial Revolution Is Not Like the Last One
Policymakers’ approach to automation won’t work for AI.
Foreign Policy
75
11 min read
0 views
The space-cum-artificial intelligence company SpaceX just went public at an eye-watering valuation—the largest in history—making its founder, Elon Musk, the world’s first trillionaire. In the same stretch of days, the U.S. Bureau of Labor Statistics reported that the average American worker had lost a year and a half of wage gains as of May. Two more frontier labs, Anthropic and OpenAI, are lined up to raise fortunes of their own, even as trust in the technology craters and anxieties about job losses and inequality climb.
Policymakers feel the pressure to prove they are not mere bystanders while a revolutionary technology threatens to upend the social order. Their instinctive response borrows from an earlier age, when automation came for the factory. That playbook never worked especially well: Since the 1980s, it failed to make displaced workers whole, and across the advanced world, its deployment coincided with rising inequality. And it is even less suited to what comes next.
The space-cum-artificial intelligence company SpaceX just went public at an eye-watering valuation—the largest in history—making its founder, Elon Musk, the world’s first trillionaire. In the same stretch of days, the U.S. Bureau of Labor Statistics reported that the average American worker had lost a year and a half of wage gains as of May. Two more frontier labs, Anthropic and OpenAI, are lined up to raise fortunes of their own, even as trust in the technology craters and anxieties about job losses and inequality climb.
Policymakers feel the pressure to prove they are not mere bystanders while a revolutionary technology threatens to upend the social order. Their instinctive response borrows from an earlier age, when automation came for the factory. That playbook never worked especially well: Since the 1980s, it failed to make displaced workers whole, and across the advanced world, its deployment coincided with rising inequality. And it is even less suited to what comes next.
A black-and-white historical engraving depicting the interior of a large industrial factory. Overhead, giant wooden or metal wheels are connected by long mechanical belts. Several workers in 19th-century attire are scattered throughout the space, some sweeping, holding tools, or carrying sacks.
A depiction of machinery used for crushing silver ore in Nevada in the 19th century.Universal History Archive/via Getty Images
The core ingredients of governments’ approach to automation, pioneered in the United States and adapted to varying degrees elsewhere, were to retrain workers for new jobs, cushion the gap with unemployment insurance, compensate the losers through trade adjustment assistance, and nudge employers to hold on to jobs as long as possible. The premise is that, after a painful adjustment, the economy resettles at a higher level of productivity, demands more workers in new roles, and leaves everyone better off.
Governments are once again placing retraining and upskilling at the center of their adaptation strategies. India’s government touts a “national reskilling engine to upskill and reskill millions.” Canada promises AI literacy and placement for younger workers. South Korea offers vocational retraining and conversion programs. The United States still lacks a federal framework for workers exposed to AI, but bipartisan bills with upskilling at their core are under discussion in Congress. States have moved on their own: Gov. Gavin Newsom’s executive order proposes retraining at-risk white-collar workers in California, as does Premier Jacinta Allan’s career-rescue scheme in Victoria, Australia. Others have gone further still: Beijing not only encourages employer-provided training but, backed by multiple court rulings, has warned companies against cutting jobs at all as they adopt AI.
These measures are well intentioned. They are also likely to fall short because to meet AI with the manufacturing-era toolkit is to fight the last war. Several differences stand out.
For one, the scale and shape of the AI displacement are different. Research that my team at Digital Planet has conducted finds that 9.3 million U.S. jobs are at risk within five years, a figure that climbs to 19.5 million if adoption accelerates, putting between $757 billion and $1.5 trillion of income at risk as well. That is a big hole for retraining programs or unemployment funds to plausibly fill. Outplacement firm Challenger, Gray & Christmas tallied some 1.2 million job cuts across 2025, and the pace has only quickened since. Moreover, the cuts are structural rather than cyclical; firms are not trimming for a downturn but permanently rebuilding their workforce, using AI as the primary reason.
The central pillar of policy intervention—worker retraining—rests on a premise that might not hold in the AI era. Historically, machines, from the spinning jenny to the welding robot, took over the physical, repetitive tasks and opened room to move displaced workers up into cognitive, better-paid roles. AI runs that escalator in reverse. Its impact falls hardest on precisely that cognitive, analytical knowledge work and on the people who already invested in a college degree and the expectations that come with it. The reassuring counter that AI will augment knowledge workers rather than replace them offers less comfort than it appears: Our research finds that the jobs AI is best positioned to augment are also the very ones it is best positioned to automate. Augmentation and displacement are not opposites. They are two readings of the same capability. And the displacement can happen quickly, unlike earlier waves of displacement that took decades. This makes retraining programs harder to organize and implement.
A close-up outdoor group photo of smiling people, many wearing black T-shirts that read "SPCX LIFTOFF." On the right, a person wearing a white futuristic spacesuit and helmet points directly toward the camera. Modern city buildings rise in the background.
SpaceX employees celebrate the market close of the company’s initial public offering in New York City on June 12.Spencer Platt/Getty Images
The least appreciated and existential dilemma is that the fiscal model itself could break. The safety nets that cushion displacement are funded disproportionately by taxes on labor, which supply roughly half of all revenue in rich countries; corporate taxes contribute barely an eighth. Labor income is taxed at an average of 35 percent across advanced economies, while capital is taxed closer to 20 percent—and often less, once loopholes are counted. AI’s promised productivity gains shift income from a broad base of relatively well-paid workers toward the owners of capital. In doing so, the transition hollows out the very tax base that funds the programs meant to ease it. The result is a vicious cycle in which workers absorb the loss twice: first as the earnings that vanish and then as the share of the gains that flows to the owners of the algorithms. Inequality widens from both ends at once.
What, then, should replace the factory-era playbook? Five principles can guide the redesign.
First, treat skilling as permanent infrastructure rather than a one-off rescue, with AI fluency at its core: AI-skilled workers earn 62 percent more than their peers worldwide. Better still, deliver that fluency the way public health delivers vaccinations—as anticipatory immunity, distributed before exposure rather than after a worker has already been “displaced.” Several governments offer models worth adapting. Singapore treats lifelong learning for every adult as a public good; Estonia is putting frontier AI tools into every upper-secondary classroom; and France gives each worker a portable, individually owned training account that follows them across a working life.
Second, because such programs are costly and resources are likely to be limited, prioritize them by adaptive capacity—a worker’s savings, age, skill transferability, and local labor market resilience. Well-paid software developers, financial managers, and lawyers, with financial buffers and transferable cognitive skills, will largely pivot on their own. The priority should be those in routine clerical, sales, and administrative roles, who bear both the highest exposure and the least capacity to adapt.
Third, design for an uncertain tempo. AI will arrive abruptly in some sectors and grind slowly in others as supporting systems catch up, so policy will need both flexibility and buffers. Denmark’s “flexicurity” pairs easy hiring and firing with generous benefits and retraining; Sweden’s job-security councils move people into their next role before unemployment bites; and Germany subsidizes shortened hours to avert mass layoffs, a tool it leaned on during the 2008 downturn.
Fourth, insist on data before the displacement, not after. Governments are among the largest buyers of enterprise AI, and they can make any major contract conditional on an audited estimate of the jobs the tools are expected to eliminate. More broadly, employers could be required to disclose the labor impact of their AI deployments, much as they already report their environmental impact.
Fifth, and hardest, find a way to share the productivity gains more widely, whether through tax reform or an “AI dividend.” This is where the economics and the politics turn thorny and where inaction is itself a choice whose political and economic bill lands on leaders down the road.
The deepest distortion is the U.S. tax code’s standing bias against labor, which rewards companies for automating tasks to arbitrage the rate differential rather than to capture genuine productivity. Equalizing the treatment of labor and capital would tilt adoption toward labor-augmenting tools and help fund the safety net; many economists argue the asymmetry should be redressed. The objections are real: Tax capital too heavily, and you risk dampening investment or chasing operations to friendlier jurisdictions. But workable models exist. Sweden, Denmark, and the Netherlands raised energy, transport, and pollution taxes and recycled the revenue into labor tax cuts; others have taxed passive wealth and closed corporate loopholes. South Korea was the first to levy a “robot tax,” paring back deductions for firms that invested in automation—though the idea has yet to cross into the AI age, in part because no one has settled what an AI taxable base would even be.
Beyond the tax code, governments hold other levers to give workers a stake in AI’s gains. They can reward firms that fund support for displaced workers; levy surcharges on AI consumption or on the power drawn by large-scale compute and data centers; treat human knowledge as the capital that AI is trained on and require royalties for its use; or take equity in AI companies and pay citizens an AI dividend. The dividend idea echoes the natural resource funds of Alaska and Alberta, where the public shares in the profits extracted from public land. In the United States, it has produced an unlikely convergence among figures as far apart as Donald Trump, Steve Bannon, and Bernie Sanders, all open to the government taking equity stakes and routing the proceeds back to citizens. What no one agrees on is how, whether it ought to be voluntary or mandatory, and on whose terms.
A high-angle shot of a formal hearing room. Four individuals in business attire sit at a long, black-clothed table in the foreground with nameplates before them. Behind them, several rows of spectators in suits sit in tiered seating, watching the proceedings.
From left to right: OpenAI CEO Sam Altman, AMD CEO Lisa Su, CoreWeave CEO Michael Intrator, and Microsoft President Brad Smith prepare to testify during a U.S. Senate commerce committee hearing on artificial intelligence in Washington on May 8, 2025.Brendan Smialowski/AFP via Getty Images
The architects of the AI economy are minting trillion-dollar fortunes from a technology that, for now, is subtracting more from the average paycheck than it adds. The reflex to meet that disruption with the manufacturing-era toolkit is understandable; it is the only playbook governments have ever run. But that toolkit was built for a different kind of shock, one that climbed the ladder from the bottom, advanced at the pace of decades, concentrated in a few factory towns, and did not disrupt the tax base that pays for the programs to the same degree as it will do now. AI inverts all four.
Applying the remedies of the factory floor to the upheaval of the corner office will not merely fall short; it will fail in compounding ways: workers stranded, the safety net starved, inequality widening at both ends. Whether the technology pays a productivity dividend to everyone else is not a forecast to wait for; it is a choice to make. Policymakers ought to consider it for their own survival as workers displaced by AI are not on the inside when the IPO or productivity gains are distributed. But they can design websites, write code and op-eds, are skilled at marketing and influence campaigns, are on social media, and have their political representative’s number.
