In new research, Christos Makridis and Andrew Johnston find that industries exposed to generative AI are seeing an increase in production, employment, and wages. However, the majority of AI-driven revenue growth is channelled back to capital as profits, rather than to workers.
In February, Citrini Research published a scenario analysis that went viral in financial and policy circles. Framed as a memo written from 2028, it described an “intelligence displacement spiral” in which layoffs driven by artificial intelligence pushed unemployment in the United States above 10%, triggered a 38% stock market collapse, and unraveled the $13 trillion mortgage market. Shortly before, a widely cited Stanford Digital Economy Lab working paper concluded, among other things, that young workers in software and customer service were already experiencing AI-driven employment declines. The debate over AI and the future of labor has organized itself around a single question: how many jobs will AI destroy?
However, new evidence from our working paper—using administrative data covering over 95% of U.S. employers from all industries across 2017–2024—finds that sectors more exposed to generative AI have experienced large gains in output, employment, and wages, with effects emerging well before ChatGPT entered public awareness beginning in 2022. AI is not destroying jobs. Yet, the distribution of its gains is not flowing primarily to labor. Workers are capturing only about 29 cents of every dollar of AI-driven output growth, and that benefit is directed to jobs integrated with AI, rather than those replaced by it. The rest flows to capital as profits, especially in states where employees cannot increase their wages by freely moving to new occupations, companies, and sectors.
What individual-level studies miss
Two papers have dominated the recent literature: Anders Humlum and Emilie Vestergaard, using Danish administrative data, find that the use of AI chatbots has no significant effect on workers’ wages or hours. Meanwhile, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond find employment declines among young workers in software and customer service following ChatGPT’s release.
Neither finding is wrong. Both studies use different designs that limit how AI’s impact on labor can be extrapolated. Humlum and Vestergaard compare workers within the same occupation at similar firms who self-report chatbot use versus those who do not. That design captures within-occupation adoption effects but ignores how AI may impact the hours or wages of a whole geography and/or sector. For example, if software engineers become much more productive, then that could generate productivity effects for the region, which in turn attracts more firms to it and leads to an expansion in demand for less AI-exposed jobs. Danish labor market institutions also matter: roughly 82% of workers in Denmark use collective bargaining, meaning that wage adjustment is institutionally constrained in ways that do not characterize the U.S., where coverage is closer to 12%. The effects of AI on wages might show up differently in Denmark than in the U.S.
Brynjolfsson, Li, and Raymond find employment declines specifically in software development and customer service—precisely the occupations with high automating exposure in our framework. Their payroll data covers large employers at the occupation level and detects displacement within those roles. Again, they do not look at how AI impacts other roles in the sector.
Our data from the Quarterly Census of Employment and Wages captures the full sector, including firm entry, exit, and worker reallocation across firms of different sizes. Job loss within specific automating occupations at large firms can coexist with net positive employment at the sector level, once new firms enter the market and labor moves into roles where AI augments human activity, rather than reduces it. This is exactly what has happened.
Approaching a distinct AI study
To identify the causal effect of AI exposure, we compare sectors within the same state and year that differ in their pre-pandemic occupational composition. For example, the oil industry in Texas employs a different mix of programmers and bid writers than it does in Oklahoma. This means that companies from different states faced different AI exposure both before and after enterprise AI tools, designed to extend AI technologies across an organization, reached the market. We measure that exposure using occupation-level task scores mapped into states and industry data from the Census Bureau pre-pandemic, which allows us to create a state-by-industry measure of AI exposure. Then, we trace out the evolution of real output, employment, and other outcomes within states and industries that vary in their occupational exposure to AI, controlling for all time-varying state and industry factors, as well as differences across states and industries. This leaves only within industry-state variation over time.
We find a one standard deviation increase in AI exposure raises sector output by roughly 10% by 2024, with employment up 3.9% and the aggregate wage bill—the total a company pays its workers—up 4.8%. These effects began in 2021, before the launch of ChatGPT, because the relevant technology had already entered professional workflows through enterprise tools: GitHub Copilot for software development, Jasper for marketing and content writing, and Microsoft’s GPT-3-powered business applications. Venture capital investment in AI-related startups surged 139% in 2021, reflecting how investors responded to demonstrated enterprise traction, not consumer hype.
Breaking down AI exposure
Following recent studies, we further decompose AI exposure into two types: automating exposure, where AI can perform tasks independently without human collaboration, and augmentative exposure, where AI raises worker productivity but cannot substitute for human judgment and oversight.
Where AI augments workers—requiring human collaboration to realize the productivity gain—employment rises by approximately 4% and wage growth follows. Where AI can operate independently, employment shows no significant effect, and total wages actually fall slightly, suggesting compression of per-worker compensation even absent headcount reductions.
Radiology is a good example. Machine learning models now flag anomalies in high-volume screening scans—tasks with well-defined inputs and verifiable outputs. These qualities facilitate tasks to be automated. Yet final diagnosis, clinical integration, and patient communication remain tasks where AI functions as a second reader rather than a replacement. Consistent with the augmentative channel, radiology residencies reached an all-time high of 1,451 positions in the 2025 National Resident Matching Program—a 5.2% increase over the prior year. Geoffrey Hinton’s 2016 prediction that training new radiologists should immediately stop has proven to be untrue.
Who captures the gains
AI does not affect all jobs equally, or even all AI-exposed jobs equally. The task structure and management practices within an occupation—and whether realizing AI’s productivity benefits requires human collaboration—determines whether workers see more employment and higher wages, or simply watch output grow while their wage share declines.
Still, even where AI raises both output and employment, the surplus flows largely to capital. Total company wages, representing in theory labor’s share of output, account for roughly 60% of sector output. A 10% increase in output paired with a 4.8% increase in the wage bill implies that only about 29% of gains in output induced by AI accrue to labor—far below the labor share. The gap between output and wage-bill growth implies that AI exposure reduces the labor share by five percentage points per standard deviation of exposure.
This is not simply a story of firms automating away workers. Employment is rising. The issue is that firms are capturing the productivity surplus. AI-driven productivity gains are being retained as profits rather than passed through to wages at the rate the existing labor share would predict. Workers drive the expansion but don’t proportionally gain from it, which makes related work on employee ownership in the era of generative AI especially timely.
We validate these results using the same identification strategy and individual-level data from the Census Bureau’s American Community Survey linked with generative AI exposure. Wage effects emerge in 2022, one year after output and employment rose. This is consistent with firms expanding hiring before adjusting compensation, and with wage bargaining cycles that delay productivity pass-through. By 2023–2024, a one standard deviation increase in exposure raises hourly wages by approximately 1.0–1.1%. The gains are concentrated among college graduates (1.4% versus 0.46% by 2023–2024), larger for workers aged 22–30 than for those 31–65 (1.4% versus 0.95%), workers who are Black (1.4% versus 0.5%), and female workers (1.1% versus 0.95%). This suggests AI-driven wage growth has not been exclusively concentrated among traditionally advantaged groups, though the educational gradient is the sharpest.
When we move the study beyond how AI changes wages within specific roles, estimated effects jump substantially. The implication is that a meaningful share of wage gains reflects workers moving into higher-paying occupations within AI-exposed sectors rather than workers seeing pay increase in their current occupations. New firms entering AI-exposed sectors—which cannot be measured by firm-level or worker-level studies—are creating roles that did not previously exist. For example, coders may become AI architects.
Labor market fluidity as the policy lever
One finding carries direct policy relevance. We split states by their pre-AI labor market turnover rates, using Longitudinal Employer Household Dynamics data from 2016–2019. The positive employment effects of AI exposure are concentrated in high-turnover states. In low-turnover states, employment effects are effectively zero, while wage effects are reduced rather than absent. This gives us extra confidence that reallocation is playing a primary role. AI’s aggregate employment gains depend on workers moving across occupations and firms into newly productive roles. Where labor markets are rigid, that reallocation does not happen, and the productivity gains remain stranded in capital rather than accruing to workers.
This has a specific policy implication. Worker movement—across occupations, firms, and sectors—is the channel through which AI’s aggregate benefits reach workers. Policies that impede movement, including occupational licensing restrictions, location-specific benefit structures, or regulatory frameworks that slow firm entry, directly constrain workers’ ability to share in AI-driven productivity growth. The relevant policy intervention is not limiting AI adoption—that forecloses the productivity gains entirely—but ensuring that labor market institutions facilitate rather than block the movement AI is generating.
The low percentage of benefits going into the labor force presents another problem.. If firms are capturing 71 cents of every dollar induced by AI, and that division persists, it implies a structural shift in how the returns to technology are distributed. Prior waves of technological change raised the demand for labor substantially. Whether capital’s heightened share of AI-induced revenue reflects the initial dynamics of a disruptive technology—firms capturing rents before product market competition dissipates them—or something more durable about the economics of AI remains an open question.
What the data says, and what it doesn’t
Because AI capabilities have continued to advance rapidly, and because the boundary between automating and augmentative tasks is shifting, our estimates likely understate the productivity and labor market effects that have accumulated since 2024. The analysis also has its own design limits and does not, for example, study job movement across sectors. Further, what the sector-level design of our study cannot capture is why two firms in the same AI-exposed industry diverge so sharply in outcomes. Related work on organizational transmission using longitudinal worker data suggests the answer lies largely with management: firms where managers actively encourage and model AI use see substantially higher adoption rates and larger productivity gains. The technology’s potential may be uniform across sectors; whether it gets deployed is not.
The displacement narrative that AI will destroy jobs en masse is not supported by comprehensive employment and output data through 2024. But the narrative of broadly shared gains is not supported either. AI is raising both output and employment, and admittedly it is also widening the gap between what firms capture and what workers receive. Understanding why that gap exists, and whether it persists, is the question that should be driving the AI policy debate, not the body count of jobs that may eventually be replaced.
Author Disclosure: The author reports no conflicts of interest. You can read our disclosure policy here.
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.
Subscribe here for ProMarket’s weekly newsletter, Special Interest, to stay up to date on ProMarket’s coverage of the political economy and other content from the Stigler Center.
