Study Says AI Reshaping Labor Markets

A new Dallas Federal Reserve study suggests generative artificial intelligence adoption is reshaping the Texas labor market, primarily by reducing demand for some types of work, not so much by eliminating existing positions but by reducing demand for recent college graduates. 

“The most exposed occupations are generally in software development, web design and other computer-heavy occupations,” the report says. “Managers, clerical workers, editors and other white-collar occupations are also subject to some of the highest levels of AI task exposure.”

source: Dallas Federal Reserve

“There is strong evidence that GenAI has decreased labor demand for occupations consisting of tasks that can be performed by these new tools,” the researchers say. Still, “the overall effect on aggregate online job posting behavior thus far has been modest.”

“Recent college graduates, whose unemployment rate rose to unusually high levels during this period of rapid GenAI adoption, are where effects of GenAI on employment and earnings are likely to first appear,” the study suggests. 

So far, AI's labor-market effect is that “fewer people get hired into certain jobs,” rather than “large numbers of existing workers fired from those jobs.” 

Study	Data / period	Main finding	Job impact
Brynjolfsson, Chandar & Chen — “Canaries in the Coal Mine” (Stanford, revised 2026)	ADP payroll data, millions of workers, through June 2026	No widespread economy-wide displacement, but employment of 22–25-year-olds in highly AI-exposed occupations is 19% below the counterfactual. Experienced workers show no comparable gap.	Strong evidence for the hiring channel. The authors say the adjustment occurs primarily through reduced hiring of young workers rather than increased separations. (Stanford Digital Economy Lab)
Hosseini Maasoum & Lichtinger — “Generative AI as Seniority-Biased Technological Change” (Harvard, 2026 revision)	65 million résumés, 280,000+ firms	Junior employment falls following GenAI adoption, particularly in highly exposed occupations; senior employment is largely unchanged.	Decline is driven primarily by slower hiring rather than increased separations. (SSRN)
Tucker — “You're (not) Hired” (U.S. Census, 2026)	Matched employer-employee administrative data	Early-career employment in the most AI-exposed industry/state cells fell 12% over 10 quarters after ChatGPT.	The paper finds the decline in employment was primarily caused by a large decrease in hiring. This is perhaps the clearest administrative-data evidence of the mechanism. (Census.gov)
Liu, Wang & Yu — “Labor Demand in the Shadow of Generative AI” (World Bank, 2026 revision)	285 million U.S. online job postings, 2018–2025	Postings for occupations highly vulnerable to AI substitution fell 9% relative to less-vulnerable occupations, with the differential reaching 15% by the third year.	Direct evidence of reduced labor demand, rather than layoffs. Particularly important because it examines vacancies at enormous scale. (SSRN)
Audoly, Guerin & Topa — New York Fed, “Do Job Postings Show Early Labor-Market Effects of AI?” (2026)	U.S. Lightcast postings	Overall hiring has slowed, but they find little evidence of a distinct AI-driven decline in postings for AI-exposed occupations.	Important counterweight: the aggregate slowdown in postings cannot confidently be attributed to AI. (Liberty Street Economics)
Federal Reserve Board — “AI Adoption and Firms' Job-Posting Behavior” (2026)	Firm/industry AI adoption + job postings	No evidence that firms or industries with greater AI adoption have reduced total job postings.	Suggests that if AI is eliminating some positions, firms may be switching hiring toward other jobs, rather than simply reducing total hiring. (Federal Reserve)
Gimbel, Kendall & Nunn — Yale Budget Lab (2026)	Monthly CPS employment/wage data	After controlling for differences between exposed and unexposed occupations, they find no statistically or economically significant aggregate employment or wage effect.	Again, little evidence of broad existing-job destruction. The effects may be concentrated in particular populations. (The Budget Lab)
Humlum & Vestergaard — “Still Waters, Rapid Currents” (NBER, 2025/26)	Danish administrative records + AI adoption surveys	No detectable effect on earnings or hours, even among early adopters and highly exposed workers. But substantial task restructuring and occupational switching occurred.	Little evidence of job elimination so far. Firms appear initially to be reorganizing work rather than cutting employment. (National Bureau of Economic Research)
Chandar — “Tracking Employment Changes in AI-Exposed Jobs” (2025)	U.S. CPS, Q4 2022–Q1 2025	No substantial aggregate employment/earnings difference in highly exposed occupations, although software and customer-service occupations diverge.	Overall employment effects small; evidence of localized employment declines, not economy-wide displacement. (SSRN)
Frank et al. / related job-posting research summarized by Stanford	U.S. job postings	Several studies find greater declines in postings in AI-exposed occupations. But some declines began before ChatGPT and correlate with interest rates/remote work.	Supports reduced postings, but attribution to GenAI is contested. (Brookings)
“Winners and losers of generative AI” (JEBO, 2025)	Online freelance marketplace	About 10% of postings were judged substitutable by GenAI; demand for those skill clusters fell as much as 50% in short-term roles.	Strong evidence of demand substitution in particular tasks, although aggregate freelance demand did not fall. (ScienceDirect)
PwC 2025 AI Jobs Barometer	Lightcast job postings, 2019–2024	U.S. occupations with greater GenAI exposure experienced substantially slower job-posting growth: roughly 2% vs. 20% for less-exposed occupations.	Strong descriptive evidence of slower demand growth, though not necessarily causal evidence of AI. (PwC)

But trends could, or maybe, should, change over time, as actual job displacement or elimination, plus creation of new jobs, happens. 

Effect	Evidence so far
Mass layoffs caused by GenAI	🔴 Little evidence
Economy-wide employment decline	🔴 Little/no evidence
Reduced total hiring because of AI	🟡 Weak/mixed
Reduced hiring in particular AI-exposed occupations	🟢 Increasing evidence
Reduced entry-level hiring	🟢 Significant evidence, but causality disputed
Reduced job postings in AI-substitutable occupations	🟢 Fairly strong evidence
Task substitution/reorganization	🟢 Strong evidence
Productivity increases without employment reductions	🟢 Strong evidence
Wage effects	🟡 Generally small so far
Long-term displacement	❓ Still largely unknown

On balance, it is possible net job creation could happen, as AI creates new jobs and roles, despite AI-induced reductions. 

Study	Period / data	Main finding	Implication
Autor, Chin, Salomons & Seegmiller, “New Frontiers: The Origins and Content of New Work” (QJE, 2024)	U.S., 1940–2018; ~35,000 Census occupations	About 60% of employment in 2018 was in occupations that did not exist in 1940. New work emerged alongside technological change, particularly in professional and service occupations after 1980. (National Bureau of Economic Research)	Perhaps the strongest evidence that technology doesn't simply redistribute a fixed number of jobs. It creates new categories of work.
Autor, “Why Are There Still So Many Jobs?” (JEP, 2015)	Historical review	Automation substitutes for labor in particular tasks but also complements workers, raises productivity and output, and increases demand for labor elsewhere. (TopCat)	Explains why technological progress can eliminate particular jobs without eliminating work generally.
Acemoglu & Restrepo, “Automation and New Tasks” (JEP, 2019)	U.S. historical/employment evidence	Automation creates a displacement effect, but creation of new tasks produces a reinstatement effect that raises labor demand. (AEA Publications)	Provides a theoretical framework for understanding why job creation can offset automation.
Acemoglu & Restrepo, “The Race Between Man and Machine” (AER, 2018)	Long-run economic model	Technology that automates existing tasks reduces labor demand, while creation of new tasks has the opposite effect. (American Economic Association)	Net employment depends on the balance between automation and new-task creation.
Bessen, “Automation and Jobs: When Technology Boosts Employment” (2017)	Historical industries, including textiles, steel and automobiles	Industries experiencing rapid productivity growth sometimes experienced employment growth, because lower costs stimulated demand sufficiently to offset labor-saving technology. The OECD reviews this evidence. (OECD)	Lower prices can create enough additional demand to more than compensate for labor-saving technology.
Mann & Puttmann, “Benign Effects of Automation” (2018)	U.S. counties/industries, patent data	Automation innovations were associated with declining manufacturing employment but increasing employment in services. (OECD)	Job destruction and creation can occur in completely different industries and locations.
Autor & Salomons, “Is Automation Labor-Displacing? Productivity Growth, Employment, and the Labor Share” (2018)	28 industries, 19 advanced economies, 1970–2015	Productivity improvements can reduce employment within an industry, but positive spillovers to other industries more than offset those losses. The OECD summarizes the evidence. (OECD)	This is especially important: you shouldn't look only at the industry where automation occurs.
OECD, Employment Outlook 2019	OECD countries, long historical perspective	Despite substantial technological displacement, overall employment has generally grown. The OECD concludes that historically the net effects of major technological revolutions on employment have been positive. (OECD)	Broad institutional review supporting the historical pattern.
OECD, Technology, Productivity and Job Creation (1998)	OECD historical evidence	Technological change destroys jobs in some industries while creating jobs in others; historically the process produced net job creation as new industries replaced old ones and demand expanded. (OECD)	An earlier, broad cross-country examination reaching the same conclusion.

AI and Jobs: Correlation is not Causation

It always is difficult to separate correlation from causation in any complex endeavor. Consider the impact artificial intelligence might have on employment. 

Big layoffs at enterprise-sized firms, said to be driven by new AI potential, essentially shift spending from people to tokens but without clear direct financial returns. 

So although we are very early in the process of adopting AI, we still know very little about actual AI impact on jobs. 

A new study by Ramp and Revilio Labs that suggests artificial intelligence adoption actually increases the number of jobs at firms using AI, rather than decreasing employment. 

Or does it?

The study itself suggests a possible “correlation” rather than direct causation: “Companies that adopt AI look very different from companies that never adopt,” the report notes. “AI adopters are larger, more engineering-intensive, more likely to be venture-backed, and were already growing at a faster rate before adoption.”

And that might suggest correlation: the AI adopter firms were growing faster even before AI was adopted. 

It might plausibly also be the case that companies best able to make AI investments can do so because they already are growing revenues and headcount. 

source: Revelio Labs 

“Companies making the largest AI investments grow employment by roughly 10 percent on average following adoption, while low-intensity adopters see no statistically significant change,” the report states. 

Again, the point is that fast-growing firms typically are those adding headcount faster. 

And when the report notes that “among companies making the largest AI investments, the share of entry-level workers increased by 1.15 percentage points compared to not-yet adopters, that might also be because such firms are increasing employment virtually across the board. 

That is not to say AI adoption did not aid employment growth, but only to say we cannot really prove AI was the difference maker, as the data shows the firms adding AI services or apps were faster-growing before AI was added. 

That sort of thinking is in line with other studies of technology adoption that tend to show better-managed firms also are better at integrating new technology. 

Study/Paper	Key Findings	Source
Bloom, Sadun & Van Reenen (2016/2017): "Management as a Technology?"	Management practices (WMS) explain ~30% of TFP gaps; treated as technology-like capital; positive interaction with IT; large cross-country/firm variation.	NBER w22327
ONS (2025): Management practices and technology/AI adoption in UK firms	Strong correlation: better management → higher tech adoption; tech adopters have ~19% higher labor productivity after controls; management predicts AI follow-through.	ONS Article
Cirera et al. (various, e.g., 2021): Firm-Level Technology Adoption (FAT) surveys (Vietnam, Brazil, etc.)	Management quality (incentives, monitoring) strongly predicts technology sophistication indices; linked to productivity; firm capabilities key driver.	World Bank
Babina et al. (2024): AI, firm growth, and product innovation	AI-investing firms show higher sales/employment/valuation growth via innovation; selection via instruments (university AI supply).	ScienceDirect
Alfaro-Serrano et al. (2021): Interventions to promote technology adoption	Reviews evidence linking adoption to performance; management/human capital as key enablers.	PMC
World Bank FAT-related (e.g., Ceará, Senegal)	Management practices and skills correlate with tech adoption intensity; implications for productivity gaps.	World Bank

Better-managed firms might have strong practices in monitoring, incentives, target-setting and talent management, for example. In other words, they have intangible assets that help explain why they are better able to take advantage of new technologies. 

Such firms often also have higher productivity, growth rates and profit margins, making it hard to isolate technology's independent contribution to outcomes. That might be the case with the Revelio Labs study. 

Conversely, poorly-managed firms may lack the complementary skills, processes, or culture to adopt effectively, leading to slower or failed implementations, perhaps with near-term productivity dips as organizational effort is shifted to learning how to use the new tools.

Highly-publicized mass layoffs often are said to be about AI displacement, but often are mostly about correcting earlier overstaffing or simple ways of shifting budgets from people to investing in AI. 

The point is that we cannot discern much, yet, about the actual impact of AI on jobs.