Many today are haunted by the question: Will AI take our jobs? From software engineers to truck drivers to call center workers, a wide range of occupations appear to be directly threatened, and no occupation seems entirely safe. Exaggerated warnings of mass unemployment are circulating, and AI is treated as a prime suspect in the unusually poor job market for recent college graduates, amidst a fairly healthy-looking economy.
The first step to appreciating the possibilities unlocked by AI is to reframe these prospective job losses as grounds for optimism, even if they bring discombobulating and sometimes painful career disruptions. Every great technological leap has involved job loss, usually a lot of it. Joseph Schumpeter, in Capitalism, Socialism, and Democracy (1942), called this “creative destruction,” and it is part of progress. Our anxiety may bemitigated somewhat if we deepen our understanding of knowledge as afactor of production, drawing insights from great economists: Nobel laureates Robert Solow (1987), Paul Romer (2018), Friedrich Hayek (1974), and Claudia Goldin (2023), as well as Goldin’s co-author Lawrence Katz. Their concepts can illuminate how AI will enhance the wealth of nations.
The first step in the argument verges on being too obvious, yet it requires explanation: AI contributes to the growth of knowledge.
Philosophers argue endlessly about what counts as knowledge, often dividing into “foundationalists,” who ground knowledge in inference from indubitable beliefs, and “coherentists,” who validate it through large-scale consistency. In practice, it takes a combination: a coherent worldview, continually tested against evidence and experience. Knowledge in this sense is not only a good thing in itself, but a factor of production that can grow the economy. AI expands economically useful knowledge on two fronts: it helps individuals learn more, and it lets them use knowledge they don’t personally hold. At the social level, it accelerates discovery and reduces the costs of storing and applying what humanity already knows.
There are trade-offs: just as literacy weakened memorization, reliance on AI may dull certain skills. But on balance, AI enlarges our intellectual reach. It makes us more capable of testing ideas, comparing them across domains, and applying them effectively. Some mental habits will atrophy, but the net result will be a society functionally more knowledgeable—and therefore more productive.
Knowledge in the Solow Growth Model
Robert Solow’s classic 1956 paper, “A Contribution to the Theory of Economic Growth,” in the Quarterly Journal of Economics, integrated the growth of technology into macroeconomics, while linking it to capital dynamics. Solow’s model relies on the exogenous growth of knowledge to explain why economic growth happens at all, since otherwise it predicts eventual stagnation. But when knowledge does increase, the gift keeps giving for a while, as elevated productivity causes a period of growth in the capital stock.
Solow wasn’t insightful about why knowledge grows or how it contributes to the economy. He simply assumed that. Knowledge grows “exogenously,” outside the system, and increases output for any mix of capital and labor by a fixed proportion. That explains nothing, but it plausibly describes the world. And once knowledge raises productivity, the theory gets interesting.
Higher productivity fuels investment. Solow assumes investment is funded by savings, always a fixed share of output. Since capital continually depreciates, some investment is needed just to maintain the “steady state.” Economists expect growth to stop at equilibrium, but history shows it continues. Solow solved this by introducing a productivity multiplier—knowledge—which grows over time. Extra productivity raises output, savings, and investment, which then outpace depreciation and cause the capital stock to grow. The short-run boost to output may be dwarfed by the long-term effect as new ideas set in motion a cascade of capital accumulation.
This is directly relevant to AI. If AI truly raises productivity, it will not be a one-off gain. It will launch a prolonged phase of faster growth as investment climbs above depreciation, just as Solow described. That was true of electrification and computers; it can be true again now. The irony is that Solow himself gave his name to the “Solow Paradox”—the 1980s quip that “you can see the computer age everywhere but in the productivity statistics.” For a time, computers disappointed the optimists, but in the 1990s, productivity acceleration came.
The world is finally beginning to look more like Romer’s theory—where new ideas fuel compounding growth—because AI helps prevent old ideas from being lost or underused.
We are probably in the Solow Paradox moment of the AI revolution now. AI seems everywhere but in the statistics. Productivity will rise once organizations routinize AI use and compete by best practices—and job losses will follow. The sooner, the better.
Solow’s model was unfinished, since it left unexplained both why knowledge increases and how it contributes to productivity. Invention is partly economically motivated, and later models captured that. But Solow was partly right, too. Intellectual history has its own dynamics, driven by curiosity and the pursuit of knowledge as much as by profit.
For insight into how capitalism itself fuels the growth of knowledge, we turn to another Nobel laureate: Paul Romer.
The Economics of Capitalist Innovation
Paul Romer, writing in the 1980s and 1990s, and especially in his landmark 1990 paper “Endogenous Technological Change” in the Journal of Political Economy, took Solow as a point of departure but wanted to open the black box. He sought to understand knowledge as the product of capitalist society itself—research and development, intellectual property, and the market power of innovators. His model captured a vital causal chain: invention leads to new varieties, greater specialization, and sustained growth.
Romer’s model represented output as depending on an ever-expanding variety of intermediate goods. The math was oversimplified, but it captured an essential truth: growth depends on novelty and division of labor. More of the same yields diminishing returns; variety spurs productivity. Capitalism, with intellectual property protections, monetizes invention ex post and thereby incentivizes it ex ante.
Romer’s conclusion was optimistic. Capitalists would keep inventing, the economy would keep growing, and knowledge would never decay. If you want faster growth, increase research budgets. His model was more optimistic than Solow’s because, unlike physical capital, knowledge did not depreciate. But while ideas don’t decay, society’s stock of knowledge, like society’s stock of structures and machines, requires constant maintenance—libraries, data centers, and human capacity to use them, laboriously inculcated in schools. That may explain why the real world resisted Romer’s optimism. In the late twentieth century, “the Great Stagnation,” as Tyler Cowen called it, and as Robert Gordon documented in The Rise and Fall of American Growth (2016), underscored the difficulty of making the knowledge economy benefit the masses. Maintenance is costly, pulling Romer’s optimism back toward Solow’s stagnation.
The fun thing is that the AI revolution is making Romer’s over-optimistic theory come true. By making the storage, retrieval, and application of knowledge vastly cheaper, AI makes knowledge nearly as undecaying and universally available as Romer’s simplified model assumed. In that sense, AI makes his optimism more realistic than it was when he first wrote. The world is finally beginning to look more like Romer’s theory—where new ideas fuel compounding growth—because AI helps prevent old ideas from being lost or underused. Every individual with a smartphone and enough education has now been empowered to learn and apply a vast range of knowledge. The frontier of knowledge has become more accessible and easier to push forward.
But the economy needs local knowledge as much as global knowledge, and AI can help with that too.
Hayek’s Decentralized Knowledge
Knowledge can mean different things, and Friedrich Hayek’s famous essay “The Use of Knowledge in Society” (1945) is an important counterpoint to Solow and Romer’s conception of knowledge as a quasi-public good. Knowledge in Romer is general: once discovered, it is generally available and remains applicable forever. But the economy also depends on specific knowledge: of unused stocks, underutilized people and machines, price differentials, and so forth.
Hayek lucidly explained why markets succeed where central planning fails: prices aggregate millions of bits of local, tacit information that no central authority could ever collect. The superiority of capitalism lies in relying not on any one mind but on decentralized coordination.
Cloud computing and AI may seem to centralize knowledge, but because AI remains complementary to human users, knowledge stays dispersed in Hayekian fashion. Farmers using AI-driven weather models, small businesses using AI for logistics, and citizens using AI to navigate bureaucracy—each applies it in local ways. Capitalism is diffusing AI tools into countless hands, empowering innovators and facilitating markets.
An underappreciated threat here is noise. For instance, anecdotes suggest labor markets have turned sluggish because AI makes it too easy to apply for jobs. Applications are now less credible signals of genuine interest. Ironically, this may make personal relationships more important as a refuge from noise. But best practices will emerge to harness AI’s verbal powers for better communication, smoothing markets, and deepening the division of labor.
A New Dynamic Between Education and Technology
If Romer and Hayek capture the significance of general versus specific knowledge, the concept of “human capital” highlights another distinction: social versus individual knowledge. Invention expands what society knows. Human capital consists primarily in what individuals know.
While Gary Becker is often credited with coining the phrase “human capital,” Claudia Goldin and Lawrence Katz applied the concept to twentieth-century US history in The Race Between Education and Technology (2008). They argued that two trends shaped inequality: rising education levels and skill-biased technological change. New technologies raised demand for cognitive skills and widened wage gaps, but mass education offset the effect, helping not only graduates but also less educated workers whose labor was complementary. Inequality fell when education “won the race” by outpacing skills demand, and rose when it lagged. Early expansion of high schools and colleges kept inequality in check; later slowdowns let it grow.
A cognitive elite might end up collaborating with AI, presiding over deskilled multitudes micromanaged into high productivity but compensated with good wages and work-life balance.
This story contrasts intriguingly with today’s “toolbelt generation” and fears that AI will disproportionately replace educated white-collar workers. AI seems adept at the very mental tasks college graduates do, while leaving most manual labor untouched. If Goldin and Katz were right about technological progress being skill-biased in the twentieth century, then AI might represent a striking historical turnabout. But the contrast should provoke skepticism both ways.
Was twentieth-century technological change really as skill-biased as Goldin and Katz suppose? Consider the Ford assembly line as a case in point. Before Ford’s Highland Park plant adopted the moving assembly line in 1913, cars were made by skilled craftsmen. The new system deskilled production into dull, specialized tasks, but productivity soared. Ford paid $5 a day—high for unskilled work—to compensate. Assembly line methods spread widely. The example suggests a reason for early-twentieth-century wage compression opposite to that of Goldin and Katz, namely, deskilling technological change. But of course, there are plenty of other examples of cases where technology did complement skills, starting with the highly skilled managers and industrial engineers at the same Ford factories.
AI may already be causing a similar mix of stratification and deskilling. Moravec’s paradox shows machines find reasoning easy but simple physical tasks hard, so many “unskilled” jobs will persist, though they may become duller under tighter control (see Hans Moravec, Mind Children, 1988). By the 1950s, mass production had compressed wages but concentrated power in a managerial elite. A twenty-first-century Fordism is conceivable: a cognitive elite might end up collaborating with AI, presiding over deskilled multitudes micromanaged into high productivity but compensated with good wages and work-life balance.
Goldin and Katz’s “race” assumes technology is skill-biased and underrates frequent deskilling. Still, their concept has merit. College students can take comfort in education’s long history of complementing technology, and can expect AI to reward mental flexibility and ambition. Meanwhile, the working class has reason to hope for higher wages if they adapt to AI-driven rhythms.
Creative Destruction
If pessimists think AI will kill jobs, and optimists think it will raise productivity, both are right. Job loss—or job churn—and productivity growth are two sides of the same coin. Sometimes individuals keep their jobs, but in general, technology reshapes occupations and skills. A lot of career change is the usual price of translating innovation into economic growth. Job loss is merely the thin end of the productivity growth wedge.
Technological unemployment is temporary. Technological change has often eliminated jobs en masse, but has never reduced a substantial fraction of the population to permanent unemployment. Rising productivity fuels rising demand. Career changes can be voluntary or involuntary, but even involuntary ones can lead to better outcomes. Many who lose one job find a better one. The economics of knowledge is no reason for complacency, but it provides plenty of reason for optimism across the education and income spectrum. Knowledge is a factor of production, AI creates more of it, and that rising tide can lift most—if not all—boats.