In 1865, the English economist William Stanley Jevons observed something that violated common intuition. As steam engines became dramatically more efficient at converting coal into useful work, Britain did not burn less coal. It burned far more. Cheaper energy unlocked new factories, railroads, ships, and industrial processes that had previously been uneconomical. Efficiency did not conserve the resource; it detonated demand for it. This is Jevons’ Paradox, or the rebound effect: when a technology lowers the effective cost of a valuable input, total consumption of that input often rises because new uses proliferate faster than the efficiency savings.
The same logic is now playing out with artificial intelligence, and it undercuts the most common form of AI hysteria. The dominant narrative claims that large language models and agentic systems will automate knowledge work so thoroughly that mass unemployment follows. Lawyers, coders, writers, analysts, designers, and white-collar professionals of every stripe will be rendered superfluous. This view is wrong—not merely uncertain or incomplete, but contradicted by economic history, by the structure of demand for cognitive work, and by the early empirical record of AI adoption. It is also arrogant: it treats complex, adaptive systems of human wants, entrepreneurship, and institutional response as if they were static and fully knowable by the observer.
Efficiency Creates Demand, Not Scarcity of Work
Jevons’ insight is not limited to 19th-century coal. Every major general-purpose technology that dramatically lowered the cost of a core economic input has expanded rather than contracted the activities that used that input. Cheaper energy produced more industry. Cheaper computation produced more software, more data analysis, more digital services, and more devices. Spreadsheets did not shrink the number of people doing financial analysis; they made quantitative modeling accessible to far more firms and use cases, expanding the profession even as routine bookkeeping contracted. ATMs reduced the need for human tellers to handle cash transactions, yet the number of bank branches and customer-facing roles grew for decades because banking became cheaper and more convenient, increasing the volume of financial activity.
AI applies the same principle to cognition itself. Generative models and agents sharply reduce the marginal cost of drafting text, writing code, summarizing documents, generating images, analyzing data, and exploring design options. Tasks that once required expensive specialist hours now cost a fraction of that time and money. When the price of an input falls, two things happen. First, existing users do more of the activity. Second, new users and new applications appear that were previously priced out of existence. The addressable market expands.
Aaron Levie of Box and economists such as Torsten Sløk of Apollo Global Management have made this point explicitly for knowledge work. Lower the cost of reviewing contracts, building financial models, writing marketing copy, or prototyping software, and organizations do not simply perform the same volume of work with fewer people. They launch projects that were previously too expensive or too slow to justify. They experiment more, personalize more, iterate faster, and serve customers or internal stakeholders who previously received nothing. The volume of cognitive output rises. Because real-world value still requires human judgment, accountability, institutional knowledge, taste, risk-bearing, and integration of context that models do not fully possess, the demand for people who can direct, verify, refine, and take responsibility for that expanded output also rises.
This is not theoretical. Studies of AI exposure have found that higher exposure correlates with higher output, higher employment, and higher wage bills in affected industries, even if not every productivity gain flows through one-for-one. Firms that adopt AI intensively often grow headcount faster than peers. World Economic Forum projections continue to show large net job creation globally by 2030 even after accounting for displacement. AI-skill premiums remain substantial. The early data does not show an economy-wide collapse in labor demand; it shows churn, reallocation, and, in many AI-adopting organizations, expansion.
The Historical Record of Failed Apocalypse Predictions
The claim that “this time is different” has been made about nearly every major technology. Luddites smashed power looms. Keynes worried about technological unemployment in the 1930s. Wassily Leontief suggested human labor might go the way of the horse. In the 1960s and again with the computer and internet revolutions, confident forecasts of mass joblessness appeared. The Oxford study that claimed 47 percent of U.S. jobs were at high risk of automation became a media staple. Warehouse robots were supposed to eliminate half the jobs in logistics. None of these predictions produced the predicted secular rise in unemployment.
What actually happened is creative destruction accompanied by net expansion. Roughly 60 percent of employment in recent decades has been in job categories that did not exist in 1940. The economy creates tens of millions of gross new jobs in a typical multi-year period. Agricultural employment collapsed as a share of the workforce while absolute living standards and total employment soared. Manufacturing employment shares fell in advanced economies even as output rose. White-collar work expanded enormously. The lump-of-labor fallacy—the idea that there is a fixed quantity of work to be done—has been repeatedly falsified. Human wants are not fixed. New goods, services, and experiences continually appear. Productivity growth raises real incomes, which are spent on additional activities that employ people.
AI is different in that it targets non-routine cognitive tasks more directly than previous waves. That is precisely why the rebound effect is so relevant. The tasks being cheapened are exactly the ones that have historically had elastic demand: analysis, communication, design, coordination, and problem-solving. When these become cheaper, the economy does more of them. Radiologists have not disappeared despite diagnostic AI; demand for imaging and interpretation has grown with capability and lower effective costs in some settings. Software developers using AI tools are not uniformly being laid off en masse; the volume of software being attempted is rising. Marketing, legal services, consulting, and research face similar dynamics.
Why the Apocalyptic View Is Arrogant
The confidence with which some commentators assert that AI will “take everyone’s jobs” reveals a particular intellectual vice: the belief that one can survey the current state of technology, map it onto existing occupational categories, and declare the future of the labor market settled. This is arrogant for several reasons.
First, it treats the economy as a static set of tasks rather than a dynamic system of discovery. Entrepreneurs and organizations continually invent new ways to combine human and machine capabilities. Roles that do not yet have names—AI workflow architects, agent orchestrators, verification specialists for high-stakes domains, hybrid judgment roles that sit at the boundary of model output and institutional risk—emerge precisely because the technology lowers the cost of experimentation. Predicting the absence of such roles is far harder than observing current substitution in routine subtasks.
Second, it underestimates complementarity. Most high-value work is not pure pattern completion. It involves accountability, ethical and legal responsibility, deep domain context, relationship management, novel problem framing, and the ability to act under genuine uncertainty. Current systems remain brittle outside their training distributions and require human oversight for consequential decisions. Even where models handle 80 percent of a workflow, the remaining 20 percent—the edge cases, the stakeholder negotiation, the final sign-off—often determines whether the work creates real value. Firms that treat AI as pure replacement frequently discover this the hard way.
Third, it ignores the adaptive capacity of people and institutions. Labor markets reallocate. Workers retrain, switch occupations, and move into growing sectors. Educational systems, however imperfect, respond. Policy can ease transitions through portable benefits, skills funding, and wage insurance without attempting to freeze technological change. The historical pattern is temporary dislocation concentrated on specific groups, followed by absorption into a larger and richer set of activities. Claiming permanent, economy-wide technological unemployment requires assuming that adaptation mechanisms that have operated for two centuries suddenly fail—an extraordinary claim that demands extraordinary evidence, not extrapolation from task-level automation potential.
Early evidence already shows nuance rather than catastrophe. There is a measurable squeeze on some entry-level and young workers in highly AI-exposed occupations: firms hire fewer juniors for routine tasks that AI now handles, raising the skill bar for remaining roles. Certain administrative and content-production functions face pressure. Layoffs sometimes cite AI, though attribution is noisy and often mixed with broader restructuring. Yet aggregate employment has not collapsed, AI-exposed companies frequently expand headcount, and the overall unemployment differential between high- and low-exposure workers remains modest. The canary is singing about transition frictions and distributional effects, not about the end of work.
The Real Issues Versus the Hysteria
None of this implies that AI is costless or that every individual will smoothly transition. Displacement is real for specific people and communities. Inequality can widen if the gains concentrate among those who own the complementary capital or possess the scarce complementary skills. The quality of remaining work can change—more supervision of systems, less pure craft in some domains. Young workers may face steeper paths into expertise if the apprenticeship tasks that once built judgment are automated away. These are serious questions of distribution, training pipelines, and institutional design.
They are not the same as the claim that AI is “taking everyone’s jobs.” The latter is a media-friendly apocalypse that collapses complexity into a single terrifying image. It flatters the speaker with the appearance of prophetic clarity while ignoring the most robust regularity in economic history: productivity-enhancing technologies expand the economic pie and the set of valued human activities. It also risks bad policy. Panic encourages either Luddite restrictions that slow beneficial adoption or over-centralized control that concentrates power rather than dispersing capability.
A more accurate framing is that AI is a powerful general-purpose technology that will transform the composition of work, raise the premium on uniquely human capacities (judgment under uncertainty, creativity in open-ended domains, trust, and physical and social presence), and generate large productivity gains if societies invest in complementary human capital and flexible institutions. The volume of useful cognitive and creative activity will rise because its effective price is falling. That is the Jevons effect applied to intelligence. The people who insist the opposite—that the pie is fixed and machines simply seize slices permanently—are not hard-headed realists. They are repeating a 200-year-old error with greater rhetorical volume, and they are remarkably sure of themselves about a system whose adaptive capacity has repeatedly surprised its most confident critics.
Work will change. Some roles will shrink or vanish. New ones will appear, many of them currently unimaginable. Total employment and living standards have far more often risen than fallen under such transformations. Treating the present wave as the exception that finally proves the Luddite case requires ignoring both the paradox that bears Jevons’ name and the long record of human ingenuity in inventing new things worth doing. That is not prudence. It is arrogance dressed as concern.