This FAQ addresses the potential effect of artificial intelligence (AI) on the U.S. economy, particularly on labor markets. Many recent concerns about advances in AI—for example, its implications for national security, cybersecurity, or media disinformation—are largely outside our areas of expertise. But we do have considerable expertise regarding AI’s potential effect on labor markets and the broader impacts of technological change on the economy.
Further, many of AI’s potential effects on the economy are largely distinct from its consequences in these other areas, making the economic effects amenable to separate, granular analysis. For example, if poorly regulated and recklessly deployed AI systems lead to disruptive cybersecurity incidents, this could obviously have some economic effects (imagine a large bank suffering a cyberattack and depositors’ funds being frozen or inaccessible for a time). But regulatory failures and irresponsible management of AI systems that lead to cyberattacks do not really signal anything about the economic potential for AI tools to replace human labor at large scale or about how a rapid evaporation of AI investment or stock market gains would ripple through the economy. Nor does the right policy to mitigate possible cybersecurity or disinformation problems exacerbated by AI have much at all to do with the optimal policy to ensure working families’ prosperity and economic security in an age of AI.
Across all realms of AI policy, much attention has been paid to regulations that are specifically tailored to AI. This might be appropriate in some fields. However, we suggest that not enough attention has been paid to simply applying protections, norms, and institutions that predate AI to its uses. For example, tort law already makes it possible to hold companies liable for harms their products create. While the application of these existing legal remedies to AI hacking incidents could potentially be clarified, it doesn’t seem to require a huge new field of law or legislation to make this possible. Further, we should significantly strengthen the application and enforcement of existing protections for workers. For example, employers should be held liable when they discriminate against workers, rather than being able to evade accountability because the discriminatory impact stemmed from their use of an algorithm or other automated system.
This notion that existing best practices, policies, and ideas are likely to be the best solutions to the problems AI may cause holds particularly true when it comes to its economic effects. The single largest concern expressed in this realm is AI’s potential to permanently displace very large amounts of human labor in a short period of time, causing economy-wide joblessness and wage declines. This FAQ will examine the economic evidence surrounding many of these concerns, but one key conclusion to state here is that even if such a scenario seemed plausible, it is highly unlikely that policies tailored specifically to AI tools’ potential uses in workplaces could ever keep up with, let alone proactively block, the ways that millions of private-sector firms are adopting these tools. What would reduce the most harm in such scenarios is implementing ideas already firmly on the policy agenda for many: keep the macroeconomy well managed and unemployment rates very low; provide effective protection for workers’ rights to form unions and bargain in the workplace over issues like how technology is implemented; offer a more protective social safety net that makes spells of joblessness less devastating for workers experiencing them; and implement progressive taxes—particularly higher tax rates on income derived from capital.
It is understandable that people are anxious about what AI could do to workers’ jobs and wages in coming years; the U.S. labor market has been rigged by policy for decades to make most workers insecure and keep their wages suppressed. AI deployed into this rigged labor market, as is happening currently, is unlikely to provide much benefit to most workers and could inject more churn and uncertainty. But making the policy changes noted above and deploying AI in a high-road economy where policy forces labor markets to perform better could allay many of these worries. In fact, if power were rebalanced in our economy to ensure workers are protected and empowered, AI might deliver real tangible benefits to most workers in the form of faster wage growth.
To be clear, arguing that broad, pre-existing ideas for how to unrig labor markets and generate shared prosperity are the best response to the economic concerns of AI does not mean we’re not taking AI “seriously” or recognizing its potential harms across many other realms. Instead, our recommendation stems from granular analysis of how AI might affect the economy and labor markets.
How should we think about AI’s likely effects on jobs and wages?
The effect of AI on the wages and employment of typical U.S. workers will, like most technological advances, hinge almost entirely on the broader economic, social, and political context. Crucially, the U.S. labor market has been largely rigged by policy to suppress wages and opportunities for most workers. In this context, there is little upside for U.S. workers from widespread deployment of AI tools. There are also widespread fears that AI might threaten jobs and wages on a large scale. But were the labor market substantially unrigged, AI tools could well raise the living standards of working people. Nothing inherent to AI lets us forecast its potential effects on U.S. workers, for good or bad. This will be determined by policy.
Being relatively agnostic about the direct effect of technology and AI on labor markets does not imply that we think labor markets have been working well for U.S. workers. On the contrary, unemployment has been too high and wage growth too low for most years in recent decades. But the roots of labor market dysfunction—both past and future—have very little to do with technological changes. Instead, this dysfunction is driven by the concerted policy push to exacerbate the extreme imbalance of power between typical workers and the corporate managers and capital owners who hire and supervise them.
It is important to get the facts and analysis right on the questions of why labor markets have not delivered enough jobs or acceptable wage growth, and what the real threats are to decent jobs in the future. The current fixation on AI—and the resulting pressures to respond with policies that wrap tightly around AI-specific parameters—distracts from the more fundamental problem of imbalanced power in labor markets, pulling policy in less useful directions.
What’s AI’s most immediate threat to the U.S. economy?
AI-related spending accounts for roughly half of U.S. growth in recent quarters—a narrow and fragile base for economic growth. That spending comes from two sources: (1) tangible investment in data centers, hardware, and research, and (2) consumption financed by the AI-driven stock bubble. Both spending flows could reverse quickly if expectations change about the future profitability of AI tools. If this reversal happens, the economy may well enter recession unless policymakers react nimbly and smartly.
There are two main spending flows associated with the race to develop and sell AI tools for mass-market use. The first is tangible business investment—what we’ll call capital expenditures, or capex. Many firms are hoping to be in the first wave of companies whose AI products are adopted, so they are currently investing heavily in data centers, servers, racks, and research. This wave of AI-related spending is driving demand for land, hardware and software components, construction materials, and the labor needed to build data centers, develop the models, and market them.
The other spending flow comes from the sharp rise in the market valuation of AI-related companies, which has driven a stock market boom—or, likely, bubble—in recent years. Part of households’ annual consumption spending comes from wealth (not just income). According to general estimates, $1 in increased stock market wealth leads to around 3 cents in additional consumption spending. To the degree AI is driving above-normal valuations of the stock market relative to current fundamentals, the extra consumption spending out of this “excess” wealth can be attributed to AI.
To measure the AI effect on capex, macroeconomic observers have looked at the categories of business investment where the AI build-out is concentrated—data centers, information processing equipment, software, and R&D—and calculated how much these categories are contributing to GDP growth over and above historical norms. This excess contribution has recently reached roughly 0.7% of GDP. In an economy where growth runs at roughly 2%, this would constitute roughly a third of all increases in GDP. That said, some portion of this capex spending—likely a third to a half—probably stems from increased imports, which are then subtracted out of the contribution to domestic growth. Very roughly, we estimate that the AI capex build-out has likely contributed 0.5% to GDP growth in the past year.
For the consumption effect associated with AI, we look at today’s ratio of price to earnings (P/E ratio) in the overall stock market (the S&P 500 specifically). This measures how expensive it is to buy a dollar’s worth of earnings. Today, it is more expensive than in any period in history except the very late 1990s–2000, an unambiguous bubble that was also driven by wild enthusiasm about tech and saw a very large correction. Today’s P/E ratio is close to 40. A more normal ratio—though still high in historical terms—would be 30. If the stock market fell from a P/E ratio of 40 to one of 30 in a relatively short period of time, that would constitute a 25% decline in household wealth in corporate equities. That decline in wealth would imply a consumption reduction of about $200 billion, or roughly 0.7% of GDP.
Combined, the contributions of these two AI-related spending flows to GDP growth likely exceed 1% of GDP. Since the first quarter of 2025, growth has averaged roughly 2%, which implies that over half of all growth in the past year was being propelled by AI-related spending.
But this AI-related spending is a narrow and fragile base for overall growth. A genuinely healthy economy should be able to generate enough demand to keep unemployment low and stable without a stock market bubble. And both the stock market valuations and the determination to keep spending heavily on data centers and other capex hinge on the belief among companies and investors that AI companies’ eventual profits will justify them.
If the potential profitability of this investment in AI ever appears lower than expected, this could lead to a sharp deflation of the current stock market bubble. A drop in the market value of AI companies would almost surely also convince at least a few of the big capex spenders that their investments will not pay off, and they will likely cut back. All this means that the AI contribution to GDP growth could evaporate quickly if just one or two dominoes fall.
This would leave a large hole in aggregate demand (spending by households, businesses, and governments) that would cause a recession unless policymakers reacted nimbly to fill it. This could largely be accomplished with standard recession-fighting monetary and fiscal policy. But there would still be pain. And if political obstacles prevented the quick deployment of standard recession-fighting policy, the recession could linger and cause lasting damage.
Is the current fixation on AI the first time a technology-driven stock market bubble has threatened the broader economy?
Today’s AI boom bears striking parallels to the internet build-out and stock market bubble of the late 1990s. That episode clearly led to a stock market bubble, and the resulting deflation of that bubble led to a very stubborn recession. But the late 1990s boom also gave the U.S. labor market four to five very strong years before the bubble burst—something the current AI frenzy has largely failed to do.
The parallels between today’s AI frenzy and the late 1990s build-out of the internet are striking. The late 1990s boost to GDP from tech-related capex was even larger than today’s, and by a significant amount. The late 1990s stock market bubble (measured by P/E ratios) was the only one in history larger than today’s AI bubble, and it was even more concentrated in tech-related stocks than today’s is.
Taken together, the productivity boost delivered by the internet boom from the mid-1990s through the mid-2000s was very large—roughly in line with the more optimistic projections of how much AI could boost productivity growth in the decades ahead. By 2005, the economy was about 15% larger than projecting forward from the early 1990s would have suggested.
In short, the introduction of the internet was a huge economic story. We don’t think of it that way for the same reason that fish don’t notice water: It’s now so fundamental to daily life that it no longer registers. But it was a mammoth change with measurable, obvious economic effects.
One important difference, however, is that the internet boom lifted the labor market broadly. The late 1990s saw large reductions in unemployment relative to the previous five (or 10) years, and inflation-adjusted wages grew strongly up and down the wage distribution for the first time in a generation.
The AI boom has not benefited the broader labor market in the same way. Since late 2022, when ChatGPT was first introduced, the U.S. labor market has clearly softened, and wage growth has decelerated, most sharply for the bottom half of the wage distribution. The AI boom began with lower unemployment and tighter labor markets than the internet boom did, so it is striking that it has not been associated with a stronger—or even an equally strong—labor market.
Part of the reason is policy—not AI-specific policy but broader policy decisions. Many policy choices over this year slowed growth in the large parts of the economy outside the AI boom. For example, in normal times, much of the Trump administration’s agenda in 2025 would have been seen as very bad for business investment. Its tariff policy was extreme and chaotic; its mass deportation policy threatened to weigh heavily on both the labor supply and the customer base of very specific industries; the large deficit increases locked in by the 2025 Republican tax and spending megabill (even with the large spending cuts) seemed likely to keep upward pressure on interest rates; and the administration’s historically unprecedented meddling with the Fed’s policy independence also put upward pressure on rates and increased uncertainty.
AI-related investment has been strong in recent years even in the face of these poor broader policy choices simply because this happened to be the historical moment when the technology matured enough for firms to race to deploy it.
But non-AI-related investment—over half of all business investment—has clearly been hurt by the policy trajectory of the past year and a half. This non-AI investment decelerated through 2025 and then turned firmly negative by year-end. In short, the policy headwinds for investment really do seem to be binding on all areas of the economy not being swept along by the AI build-out.
Is AI inevitably bad for the jobs and wages of typical workers over the longer run?
There’s no obvious reason to think so. Economists’ best summary measure of how rapidly technological change is moving forward is the pace of productivity growth—growth in the average output and income created in an hour of work. Throughout history, periods of fast productivity growth (that is, periods of rapid technological advance) have been good, not bad, for the wages and jobs of typical workers. What has overwhelmingly determined how workers fare is policy, not the pace or nature of technological change, and there is no sign yet that AI will break that pattern. Current policy is notably bad for workers’ jobs and wages and should be fixed, but that poor policy environment predates AI by decades.
U.S. labor markets have performed poorly for most workers for decades. But the root of this poor performance has largely not been technological change that displaced workers or made their skills obsolete. Instead, the main drivers of dysfunctional labor market performance for most U.S. workers have been intentional policy changes meant to tilt the balance of bargaining power in labor markets away from typical workers and toward the owners and managers of corporate business. These policy changes tore down institutions (like the minimum wage and unions) that had supported workers’ leverage and bargaining power in earlier decades and introduced new institutions (like noncompete agreements) that further eroded workers’ ability to win pay increases that matched overall economic growth.
Too slow wage growth for the bottom 80% of workers—and the rise in inequality that came with it—began in the late 1970s and accelerated through the 1980s and 1990s. The early conventional wisdom about this inequality, especially among economists, was that labor markets are generally competitive and efficient, so inequality must stem from deep market forces shifting the relative demand and supply for different types of workers—things like technology and the pace of educational attainment. According to this view, policy institutions like the minimum wage and unions could affect only the small sliver of workers they touched directly, and any such effect introduced its own costly distortions. If one believed this, then the 30-year period of equal and fast wage growth after World War II could not have been driven by distortionary institutions like unions and a robust minimum wage, and the attack on unions and minimum wages could not be a convincing explanation of why inequality increased after 1979.
A tide of cutting-edge labor market research has overturned most of this conventional wisdom and shown that while technology—including AI—is a prime determinant of how fast the overall economy grows, how that growth is distributed as jobs and wages is mostly a function of policies and institutions. Like any technology, AI will help set the pace of growth, but whether typical workers share in it or whether today’s inequality widens even further will be decided by the policies and institutions that govern workers’ wages and labor market power.
The political and economic implications of this shift in understanding can hardly be overstated. If inequality were simply the product of rapid technological changes interacting with competitive and efficient markets, then it would be regrettable but essentially nobody’s fault, and little could be done about it. That is a very convenient story for the powerful. But it does not fit the historical record or economic evidence. Instead, across history, rapid technological progress has generally been good for workers, not bad: Periods of fast productivity growth have been periods of lower unemployment and slower growing inequality.
Whether U.S. workers thrive in an era of AI remains mostly a policy question. The decisions that will determine it have to do with the level of labor market power available to typical workers through policies like the level of minimum wages, the determination of macroeconomic policymakers to sustain full employment, the effectiveness of labor law in allowing workers to unionize and bargain collectively, and how robust the U.S. welfare state is.
Won’t AI automate many jobs away?
AI will likely destroy some specific jobs through automation—but it is also highly likely to create jobs more broadly over time by boosting incomes for those in the non-automated sectors. How much any particular group of workers suffers when displaced from their jobs will be determined by policy decisions. Throughout U.S. history, we have generally been terrible at managing these transitions in worker-friendly ways. It would obviously be better if we changed that and put in place the policy supports workers need to get through them.
Rapid technological change concentrated in any one sector can reduce employment in that sector. So perhaps AI really will, say, reduce the number of software coding jobs.
But if this technological change allows the same output to be produced with fewer hours of work, it raises incomes across the economy. Sticking with software coders, if the price of software design stays the same as its productivity rises, incomes within the sector will rise, and the workers and capital owners within the sector will have higher incomes they can use to buy other sectors’ goods and services. More likely, competition pushes the price of software design down, and those who purchase software design services (including purchasing indirectly as consumers buying goods and services from other businesses) will see their real (inflation-adjusted) incomes rise.
This increased demand stemming from these productivity gains means that job losses within the AI-using sector can be counterbalanced (or more than counterbalanced) by expanding employment in other sectors as long as aggregate demand (economy-wide spending by households, businesses, and governments) is maintained.
Autor and Salomons (2018) analyzed these cross-cutting effects of technological changes within particular economic sectors. They found that the own-effect of a positive productivity shock—its effect on jobs within the sector where it happens—is indeed negative: Fewer hours of work are needed to produce the sector’s output, and rising demand for that now less expensive output doesn’t fully make up the difference. However, the cross-effect—the boost that sector’s productivity growth gives to employment in other sectors—is strongly positive and generally large enough to outweigh the negative own-effect for total employment.
This offset is not always automatic. Market mechanisms could fail, at least in part, to channel higher incomes driven by technological advance into higher demand elsewhere in the economy. If the initial large income gains in the AI-using sector flow to already wealthy households, for example, these households are unlikely to spend rather than save all of these gains. That is why policymakers should keep aggregate demand high enough to keep unemployment low. They have ample tools to do it, from lower interest rates set by the Federal Reserve to redistributive fiscal policy that funds more generous public benefits out of progressive taxes.
In reality, however, market mechanisms likely will do much of the work in terms of maintaining aggregate demand in the face of productivity increases. Recessions and damagingly high increases in unemployment rates are overwhelmingly not caused by bursts of productivity growth. Instead, as the economy’s productivity potential grows year after year, those gains tend to translate relatively smoothly into growth in demand. When demand does falter, it is usually because of an outside shock, like an asset bubble popping or central bankers deliberately slowing the economy to forestall inflation.
If AI can both create and destroy jobs, does this mean policymakers can totally ignore its effects?
No. If anything, the opposite is true: Policymakers should be doing far more to protect working families from the hardship and uncertainty of earning a living in labor markets that are volatile and offer little protection. If AI-induced technological change turns out to be large and causes significant job displacement, it will hit a U.S. labor market with few worker protections and create real pockets of concentrated harm—as every other economic change of the past several decades has. But the main driver of this harm is the weakness of U.S. labor market protections, not the specifics of the technology.
To understand the scale, start with how much churn the labor market already absorbs. About 1.5 million workers involuntarily lose their jobs every single month in the U.S. economy—about 18 million jobs lost involuntarily each year. If AI raised the growth of productivity by 0.5 percentage points over the next decade and there were no countervailing job creation from increased incomes, this would very roughly translate into another 700,000 jobs each year being shed, with workers in those jobs needing to find alternative employment. Compared with the 18 million separations the economy already churns through, that is a real but manageable increase.
Even so, many individual workers may suffer from sector-specific job displacements due to AI, even in years when overall unemployment is low. The harm these displacements cause is real and significant: The labor market is not frictionless, and it can be a painful stretch of time before a displaced worker finds comparable work elsewhere. Some workers (particularly older workers) may never find a specific job as good as the one they lost.
But, again, most of this suffering is not an AI-specific problem. The vast majority of painful, forced job separations are not about AI currently and likely won’t be in the future. Workers displaced for reasons not due to AI deserve protection too. The solution is not a search for novel AI-specific policies but the broad, labor market-wide policies that help in every case: stronger social insurance, more widespread collective bargaining, and sustained high-pressure labor markets. These policies are all highly desirable regardless of the pace of technological change.
It is the absence of these broad-based labor market and social insurance policies that makes losing a specific job so damaging for some U.S. workers. Because the guaranteed baseline of job quality and social protection is so low for U.S. workers, a job that does provide some of those things is rare, and workers know it will be hard to replace. This imparts a strong sense of rational loss aversion for workers who happen to hold a high-quality job.
One reason specific jobs are occasionally highly valued in the U.S. labor market is that they come bundled with nonwage compensation—like health and retirement benefits, which are not universally available. Making those benefits universal through more protective social insurance would make losing any particular job far less damaging. Another key social insurance system—unemployment insurance (UI)—is too stingy in the U.S., imposing large income losses while workers search for alternative employment after a layoff. Making UI more generous would directly reduce the pain of a layoff.
Another reason some specific jobs can be highly valued in the U.S. economy is that they are unionized. Union jobs should be much more common than they currently are, but decades of employer hostility and policy indifference have made large-scale organizing in newly created jobs incredibly challenging. The sectors that remain unionized do so largely because of a historical legacy that saw their unions formed decades ago, so a worker separated from one today has little chance of finding another union job elsewhere. Anything that destroys today’s union jobs, technological progress included, will leave their former holders worse off. If policy changed to make it more likely for newly created jobs to be unionized, much of the damage from loss of union jobs spurred by technological churn could be ameliorated.
Additionally, the U.S. economy has spent much of the past four decades with excessively high unemployment, which radically increases the cost of losing a job. When unemployment is low and vacancies are high, a laid-off worker can find a new job quickly, putting employers under constant pressure to keep job quality high enough to retain and attract new employees.
If U.S. policymakers built a more protective social insurance system, restored the effective right to organize unions, and maintained high-pressure labor markets with low unemployment, then a large part of the damage done by any job displacements—including any driven by AI—would disappear.
Will AI lead to mass unemployment across the U.S. economy?
Almost surely not. The historical evidence is clear that business cycles with faster productivity growth see lower unemployment, not higher. And the ease with which AI or any capital investment can substitute for human labor at scale is often overstated.
Throughout modern history in advanced economies, mass unemployment has never been caused by technological advances that accelerate productivity growth. Unemployment crises are instead caused by a shortfall in aggregate demand (economy-wide spending by households, businesses, and governments). Within a given business cycle, faster productivity growth is accompanied by lower, not higher, unemployment.
It is worth stepping back from this historical evidence to think through why technological advance has never led to mass unemployment. The rest of this answer walks through some of that logic. We offer these illustrations to take the current fears of mass unemployment seriously, not to suggest that any of these scenarios are likely or have characterized the real economic history of the U.S.
One important economic reason why AI or any technological change is unlikely to cause mass unemployment is simply that technology is not free. Firms will replace labor with capital (including AI tools) only when the costs of doing so are less than the savings. But while shedding labor saves firms money, investing in new capital costs money.
And if the demand for AI tools rises rapidly as businesses across the economy try to replace large swathes of their workforce, the price of these tools will rise, and eventually (or even quickly) the price of investing in AI tools will exceed the benefits to firms.
Because so much of what is called “technological change” in recent years has occurred in online spaces that seem unconstrained by capital costs, it is easy to underestimate how costly AI tools will likely be if a broad swathe of U.S. employers try to buy them at scale to economize on labor costs. But AI tools require vast amounts of capital, labor, and energy to build and maintain (see Figure A about the scale of investments being made to build up AI capacity and enter the race to try to sell AI tools across the U.S. economy).
Further, if AI opened up historically large-scale possibilities for the replacement of human labor in the future, this would radically raise economy-wide productivity and reduce the relative cost of labor, as unemployment reduced workers’ leverage to demand wage growth that keeps pace with productivity. This reduction in the relative cost of labor would provide some businesses a large incentive to figure out how to profitably employ some of the now-jobless workers—who, again, are now quite a bit cheaper relative to other inputs into production (i.e., AI tools).
In short, mass unemployment due to technological change is highly unlikely to be an equilibrium in the economy—a place the economy remains at for long periods—unless policymakers completely fail at keeping aggregate demand (economy-wide spending by households, businesses, and governments) stable. Have policymakers failed in the past at this fundamental part of their jobs, allowing excess unemployment to linger far longer than necessary? Definitely. But that had nothing to do with AI, and the toleration of mass unemployment that might occur due to an AI-driven acceleration of productivity would be a failure of these policymakers, not an inevitability of the technology.
Also, this scenario—where AI generates extremely rapid productivity growth and the relative price of labor must fall to keep unemployment in check in the face of this AI-biased technological change—does not imply absolute reductions in wages. Instead, it only implies wages fall further behind productivity growth. But if productivity growth is supercharged by AI (and that’s the entire root of this scenario), workers’ wages and living standards could still grow in absolute terms, even more rapidly than they have in the recent past. Only wages relative to capital costs have to fall. To be clear, this isn’t a particularly happy scenario: Rising inequality is objectively bad, even if all incomes see some absolute gains. But it’s not a purely pessimistic scenario either.
Finally, much of the discourse that seems to assume a very easy replacement of human labor with AI tools underappreciates the complexity of most jobs in the economy. Jobs are a bundle of specific tasks that are often hard to clearly describe (let alone fully replace). AI will replace many tasks, but this will often just allow workers to focus more time and attention on the other tasks their jobs entail. Many of the tasks of modern jobs are extremely hard to define in binary terms that would allow easy automation. This difficulty in crisply describing many job tasks often leads to sneering about the alleged worthlessness of “middle managers” or “bullshit jobs” (to borrow anthropologist David Graeber’s term), but it stems from the fact that we have an amazingly complex economy built on ever-finer divisions of labor and information across workers and functions. Just because any single person has almost no idea what the key tasks are for most of the workforce doesn’t mean these tasks are unimportant or easy to replace. It just means the economy is complex.
How much might AI boost productivity growth—and has the U.S. labor market ever absorbed that fast a productivity acceleration before?
Most credible estimates project that AI could boost productivity growth over the next decade by an average of 0.5–1.5 percentage points per year. Even at the high end of 1.5 percentage points per year, that implies an acceleration of productivity growth that the U.S. economy has easily absorbed in the past with no labor market distress. In fact, periods of productivity acceleration have been good for U.S. workers; it is productivity slowdowns that have coincided with bad outcomes.
The effect AI will have in boosting productivity growth is still genuinely unclear. There are clear reasons to think it could run faster for a stretch as AI is incorporated into firms’ workflow, but as with all technological innovations, the pace and the ultimate depth of that penetration remain highly uncertain.
As of the first quarter of 2026, there is very little reason to think AI has contributed anything measurable to U.S. productivity growth. Since the last business cycle peak (the last quarter of 2019), productivity growth has averaged 1.96%. Looking only at the period since the third quarter of 2022 (when ChatGPT was introduced to the public), productivity growth has averaged 2.08%—a small acceleration in a very noisy data series. The pre-2022 numbers also include the large productivity declines stemming from the supply chain bottlenecks caused by the pandemic and the Russian invasion of Ukraine in the early 2020s. Figure F shows the simple level of productivity over time, along with an extrapolated trend for recent years that shows what the level of productivity would be today if the 2017–2019 productivity trend held after the rollout of ChatGPT. Two things stand out: the large drop in the level of productivity as the early 2020s supply chain snarls formed, and the fact that as of the first quarter of 2026, the level of productivity is only about 0.2% higher than what would have been extrapolated from it following its immediate pre-Covid trend.
Further, even this small edge over projection is entirely driven by the first quarters of 2023—the gap between actual productivity and trend productivity has actually declined in the past two years. None of this looks like evidence of a large technological innovation that is diffusing rapidly across ever-larger shares of the U.S. economy.
Finally, even the quite optimistic projections of how much AI might add to U.S. productivity growth in the coming decade do not imply a pace of productivity growth the U.S. labor market has never successfully absorbed before. Figure G shows a range of projections of how much AI could raise productivity growth over the next decade and compares the implied acceleration with past episodes of rising and falling productivity growth in the U.S. economy. We intentionally choose historical periods that have seen sharp changes in productivity growth to compare with projections of AI-driven productivity growth.
U.S. history has clearly seen sharp changes in productivity growth roughly as large as the more optimistic projections of AI’s effect on productivity. Importantly, the historical period that saw a rapid acceleration of productivity growth—the late 1990s and early 2000s—was an unusually strong period in the U.S. labor market, with lower unemployment and faster, broader wage growth than had been seen in a generation. The beginning of wage suppression, in contrast, followed the sharp productivity slowdown of 1973–1979. The other sharp productivity slowdown, from 2007 to 2019, was also a long period of extreme dysfunction in U.S. labor markets, with extended periods of excess unemployment and many years of near-stagnant wage growth.
In short, the actual history of the U.S. economy and credible projections of AI’s effect on productivity do not warrant extreme concern about AI-caused labor displacement.