NBER Study Suggests Limited AI Chatbot Impact on Earnings, Productivity

A study of artificial intelligence chatbot impact on labor markets in Denmark suggests the economic impact is “minimal.” Indeed, the study authors say “AI chatbots have had no significant impact on earnings or recorded hours in any occupation.” 

The study published by the U.S. National Bureau of Economic Research involved two large-scale adoption surveys conducted in late 2023 and 2024 covering 11 occupations; 25,000 workers and 7,000 workplaces.

Productivity gains were said to be modest, with an average time savings of three percent. But the study notes that AI chatbots have created new job tasks for 8.4 percent of workers, including some who do not use the tools themselves.

Nor has there been any impact on worker earnings. “Workers overwhelmingly report no impact on earnings as of November 2024,” the study says. 

Nor do productivity gains seem to have much impact on earnings. “We estimate that only three to seven percent of workers’ productivity gains are passed through to higher earnings,” say authors Anders Humlum and Emilie Vestergaard.

“Comparing workplaces with high versus low rates of chatbot usage, we find no evidence that firms with greater adoption have experienced differential changes in total employment, wage bills, or retention of

incumbent workers,” the authors say. 

The authors also note that Denmark has institutional characteristics similar to those of the United States, with similar uptake of generative AI; how hiring and firing costs; decentralized wage bargaining and annual wage negotiations. 

The 11 occupations studied included accountants, customer support specialists, financial advisors, HR professionals, IT support specialists, journalists, legal professionals, marketing professionals, office clerks, software developers, and teachers.

The findings should not come as a surprise. The “productivity J-curve" suggests that initial investments in new technologies may temporarily suppress productivity before delivering long-term benefits.

Study	Technology Examined	Lag Time Observed	Key Findings
McKinsey Global Institute 1,5,7	Digital technologies, AI	Years to decades	Benefits emerge after business process redesign and "creative destruction." Historical parallels (e.g., electric power) show lags of decades. Generative AI may shorten lags to months or years.
CEPR Study on French Industrialization 3	General-purpose technologies	5–10 years	Firms delayed adoption due to uncertainty, and early adopters operated technologies inefficiently. Aggregate productivity gains materialized slowly as organizational practices evolved.
Stanford CS Analysis 4,5	IT investments	2–5 years	Executives reported 5-year lags for IT payoffs. Complementary investments and learning curves delayed measurable productivity growth.
Productivity Paradox Research 5	IT, automation	2–5 years	"Productivity J-curve" observed: short-term costs offset gains until workflows adapted. Measurable aggregate gains emerged in the 2000s from 1990s IT investments.
Brynjolfsson et al. (McKinsey) 7	Generative AI	Months to a few years	Shorter lag due to existing digital infrastructure, but still requires process redesign. Early adopters see inefficiencies before optimization.

CoreWeave Provides Food for Both Bulls and Bears

As perhaps often is the case with fast-growing young firms in new areas, observers of CoreWeave’s first quarterly report will find both reasons for optimism and concern. Bulls will point to strong revenue growth, backlogs, revenue growth rates, cash flow and profit margins. 

Bears will likely point to the firm’s present unprofitability, cash burn, debt burdens, need for additional significant capital investment, customer concentration, competition from bigger firms and the potential impact of macroeconomic forces (economic slowdown). As a background issue, some might point to stretched valuation ratios in financial markets overall. 

Such hopes and concerns are typical.

New firms in emerging sectors often experience rapid revenue growth but face uncertainty in sustaining it due to unproven business models, evolving markets, or reliance on a few key customers. Broader market acceptance also is an issue, as adoption rates can vary. 

And, by definition, any emerging firm will be unprofitable for some time. Such issues arguably are magnified for firms requiring heavy capital investment, with the resulting cash flow issues. Depreciation of technology assets also is an issue. 

But bulls will look past all that, pointing to eye-popping growth. CoreWeave reported Q1 2025 revenue of $981.6 million, surpassing the consensus estimate of $853 million, representing a 420 percent year-over-year increase.

Management forecasts $1.06 billion to $1.1 billion in Q2 revenue and $4.9 billion to $5.1 billion for full-year 2025, implying a 363 percent growth rate, exceeding analyst expectations of $4.61 billion.

The revenue backlog stood at $25.9 billion as of March 31, 2025, including $14.7 billion in remaining performance obligations (RPO) and $11.2 billion in estimated future revenue from committed contracts.

Adjusted EBITDA reached $606.1 million, up 480 percent YoY, with an adjusted EBITDA margin of 62 percent, improved from 55 percent a year ago.

Adjusted operating income was $162.6 million, reflecting a 550 percent YoY increase, indicating operational efficiency despite heavy investments.

On the other hand, CoreWeave reported a diluted EPS loss of $1.49, significantly worse than the expected loss of $0.12.

GAAP net income for 2024 was negative at -$863 million, and economic earnings were even lower at -$1.4 billion, contrasting with the more favorable adjusted EBITDA of $1.2 billion, raising questions about financial reporting reliability.

Net interest expense surged to $263.8 million, approximately 27% of quarterly revenue, up 5.5x YoY, driven by a $2.3 billion debt facility with a 14% effective interest rate.

Quarterly loan payments, starting in January 2025, are tied to cash flow and GPU depreciation, with $500 million due per quarter by October 2025, posing a significant financial burden.

CoreWeave anticipates capital expenditures of $20 billion to $23 billion for 2025, raising concerns about sustainability and cash flow.

In 2024, 77 percent of CoreWeave’s revenue came from just two customers, and the firm faces competition from hyperscalers (Microsoft, Amazon, Google) and other startups alike.

But that is what makes markets. 

Embodied AI (Robots) Likely Key to Some Onshoring of U.S. Manufacturing

It is hard to avoid the conclusion that artificial intelligence, in the the form of embodied robots, will be an important part of the business case if manufacturing facilities return to the United States as part of restoring policies and firm responses, jobs will indeed be created, but not at the scale of the pre-offshoring era, it is reasonable to predict. 

Automation, particularly robots using artificial intelligence, likely will handle much of the repetitive, labor-intensive work. 

Study/Source	Estimated Job Creation	Role of Automation	Key Insights
Forbes (2025)	Hyundai's $20 billion investment expected to create 1,500 jobs in Louisiana	Automation is a major factor in reducing traditional manufacturing jobs.	Manufacturing production is high, but job creation is modest due to automation replacing manual labor.
Davron (2024)	Reshoring creates advanced manufacturing jobs requiring higher skills and wages	Advanced manufacturing relies heavily on technology, reducing reliance on manual labor.	Reshoring fosters innovation but requires a skilled workforce trained for automated processes.
Business Insider (2024)	Reshoring could add $10 trillion to the economy over the next decade	High-tech sectors benefit most, integrating automation for efficiency.	Automation is central to reshoring efforts, enhancing productivity but limiting traditional job growth.
Christian Science Monitor (2025)	AI-driven factories may add slightly more jobs than they destroy	Factories integrate AI and robotics to increase efficiency and resilience.	Human roles shift toward managing robots and AI rather than performing manual tasks.
Shoplogix (2023)	Robotics create new roles like technicians and engineers but displace manual labor jobs	Robots perform repetitive tasks faster and more accurately than humans.	Upskilling is essential as traditional roles are replaced by tech-focused positions.
LinkedIn (2025)	Manufacturing output increased 15% since 2020, employment rose only 3%	Automation reduces factory jobs while creating opportunities in high-tech roles.	Reshoring with automation boosts productivity but limits job creation in traditional sectors.

Where an existing garment factory presently operating in Southeast Asia employs X workers, a repatriated operation in the United States might require perhaps 10 percent to 20 percent of those workers. Recall that the reason such facilities moved from U.S. domestic locations to off shore is precisely lower labor rates. 

Assuming higher U.S. wage rates, the only way the economics would work is if far fewer workers were required. 

Study/Analysis	Source	Key Findings on Job Increases	Automation Consideration	Date
Reshoring Initiative Report	Reshoring Initiative	Estimated 1.6 million jobs reshored from 2010-2022, with potential for more if trends continue.	Notes automation reduces job counts; 2022 data shows 364,000 jobs added, but robotics limits scale vs. past.	2022
Oxford Economics:  How Robots Change the World	Oxford Economics	Robots could displace 20 million global manufacturing jobs by 2030, but reshoring may add some back.	Predicts automation will cap U.S. job gains; historical 1.6 jobs lost per robot suggests fewer net gains.	2019
MIT: Robots and Jobs	Acemoglu and Restrepo (NBER)	Between 1990-2007, 1 robot per 1,000 workers reduced employment by 6 workers locally.	Automation in reshored facilities could mean 50-70% fewer jobs than offshored totals due to robot density.	2017
Ball State University:  Manufacturing Study	Center for Business and Economic Research	U.S. lost 5.8 million manufacturing jobs (1980-2016), mostly to productivity automation, not trade.	Suggests reshoring creates high-skill jobs (e.g., technicians), but far fewer than original low-skill positions.	2017
McKinsey Global Institute: Future of Work	McKinsey	Automation could displace 16-20 million U.S. jobs by 2030, but reshoring may offset some losses.	Highlights that returning facilities will lean on robots/AI, creating 10-20% of original job numbers.	2017
ITIF: Robotics and Production	Information Technology and Innovation Foundation	Robotics boosts productivity, potentially adding manufacturing jobs in developed nations.	Job growth tied to engineers/tech roles; traditional labor replaced by robots, reducing total job count.	2019
Brookings: Automation and Jobs	Brookings Institution	Automation offsets job losses from trade, but reshoring impact limited by tech adoption.	Estimates 100-150 jobs per large reshored factory, vs. 500-1,000 historically, due to automated processes.	2015, 2022

How Long Until 50% Use of AI?

Most of us are familiar with the notion that newer waves of computing technology get adopted faster than older waves. Where it took almost two decades for personal computers to be adopted by half of households, it took less than a decade for internet use to reach half of U.S. households. Smartphones arguably reached half of users in about six years. 

So many of us would not be surprised if artificial intelligence use reached half of U.S. households in three years. That would be helped along by the fact that AI is expected to be used on smartphones, by consumer applications (social media, search, e-commerce, entertainment content), in autos and other machines and appliances as well. 

Of course, assumptions matter. We have to select particular products or apps to measure adoption. For purposes of the present analysis, we use the IBM PC as the tracked device. Obviously there were many hobbyist PCs available before then, but the IBM PC became the first mass market device “regular people” rather than hobbyists used. 

Likewise, with the internet, we track the multimedia World Wide Web, even if some people used bulletin boards before then. 

Smartphones likewise were available before the Apple  iPhone, perhaps most notably the Research in Motion BlackBerry, a mobile email device. But it was the iPhone that kicked off massive adoption by consumers. 

source: Pew Research Center 

The point is that adoption rates are lengthened if we use the hobbyist or early-adopter phases of each technology, rather than the point at which consumer mass adoption began. 

Technology	Year Introduced	Years to 10% Adoption	Years to 50% Adoption
Personal Computers	1981	5 (1986)	19 (2000)
Internet	1991	4 (1995)	9 (2000)
Smartphones	2007	2 (2009)	6 (2013)

If we consider advanced AI adoption starting around 2020 (with language models like GPT-3), AI  might reach 10-percent adoption in two years and 50 percent in three to four years. And that might be too conservative an assumption, given the fact that AI already has been used in making content recommendations, voice interfaces and search, even before the launch of language models. 

On the other hand, few consumers likely think of their use of search, e-commerce, social media or voice interfaces as “AI use,” whereas they probably do consider use of ChatGPT and other models as AI use. 

However, hardware embodiments (robots, autonomous vehicles) may align more with smartphone adoption timelines, as significant infrastructure, device development and cost reductions have  to occur.

The bigger question is “so what?” What impact will AI have on user experience or behavior? What new use cases will develop? What new markets could be created?