For several years, some of the world’s most prominent technology executives predicted that artificial intelligence would eliminate jobs on a scale and at a speed rarely seen in modern economic history. Those forecasts came from people with exceptional knowledge of AI systems. But technical expertise about what a model can do is not the same as economic expertise about how businesses, workers, customers and markets will respond.

The evidence available through mid-2026 does not show the predicted collapse in employment. AI is eliminating certain tasks, reducing demand in some occupations and making entry-level employment more difficult in several exposed fields. Yet it is also improving worker productivity, creating new work, increasing demand for AI infrastructure and allowing companies to produce more or provide better service.

The emerging story is not one of painless transition. Some workers are being displaced, and others are finding that the first rung of their career ladder has weakened. Nevertheless, the scientifically supportable conclusion is that AI is changing the composition and performance of employment—not causing the economy-wide “jobs apocalypse” that was widely predicted.

How the experts misjudged their own technology

In May 2025, Anthropic CEO Dario Amodei warned that AI could eliminate half of all entry-level white-collar jobs within one to five years and push U.S. unemployment to between 10% and 20%. He identified technology, finance, law and consulting as particularly vulnerable and argued that technology companies and government should stop minimizing the danger. His warning became one of the most frequently repeated forecasts of mass AI unemployment. Axios reported the complete prediction⁠.

Elon Musk offered an even more sweeping long-term forecast. At the November 2025 U.S.–Saudi Investment Forum, Musk predicted that advances in AI and robotics could make employment optional within 10 to 20 years. He suggested that working might become comparable to growing vegetables for recreation when food could simply be purchased at a store. He also said money might eventually lose much of its relevance.

OpenAI CEO Sam Altman had similarly expected much faster displacement of entry-level white-collar employees. By May 2026, however, Altman acknowledged that his expectations had been wrong. “I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened,” he said at a Commonwealth Bank of Australia conference. Altman added that he no longer expected the global “jobs apocalypse” discussed by some AI companies and said he was delighted that the feared employment effects had not materialized as rapidly as anticipated. Reuters reported Altman’s revised assessment⁠.

These predictions were not irrational. Frontier models were improving quickly in writing, software development, document analysis, customer service and other economically valuable tasks. But several analytical errors converted technical capability into exaggerated employment forecasts.

First, an occupation is not a single task. Accountants, engineers, lawyers, managers and customer-service employees perform combinations of technical, interpersonal, supervisory and judgment-intensive activities. Automating one component may change a job without eliminating its purpose.

Second, laboratory capability is not equivalent to dependable commercial deployment. Businesses must integrate AI with databases, workflows, security controls, regulatory requirements and accountability systems. Models still make factual errors, mishandle unusual cases and require human review. Organizational adoption therefore proceeds more slowly than benchmark performance.

Third, predictions frequently assume a fixed quantity of economic output. If a company can produce the same output with fewer labor hours, it could reduce employment. But it can also lower prices, increase quality, serve additional customers, develop new products or expand geographically. Increased demand can absorb some or all of the productivity gain.

Fourth, many forecasts count the work AI could theoretically automate without counting the complementary employment needed to deploy it. AI requires software integration, cybersecurity, data management, electrical equipment, data centers, cooling, power generation, construction and advanced manufacturing.

Finally, technology producers have incentives to emphasize the transformative power of their products. Their knowledge of model architecture does not confer special ability to forecast interest rates, consumer demand, business formation, labor mobility or management decisions.

What happened in the 2024 and 2025 labor economy

The U.S. economy did not experience mass AI unemployment in 2024. According to the Bureau of Labor Statistics, nonfarm payroll employment increased by approximately two million during the year, averaging 168,000 additional jobs per month. That was slower than the 2023 average of 216,000, but it was continued employment expansion, not technological collapse. Health care, government and social assistance produced much of the growth. Manufacturing showed little net growth, while information and some professional occupations remained comparatively weak. BLS reviewed the 2024 employment record⁠.

The labor market weakened substantially during 2025. Revised BLS figures show payroll employment increasing by only 584,000, or approximately 49,000 per month, compared with two million in 2024. Health care added an average of 34,000 jobs monthly, down from 56,000 in 2024. Food services continued expanding, while retail, information, finance and portions of professional employment performed poorly. The unemployment rate reached 4.6% in November 2025, and unemployment among people ages 20 to 24 reached 8.3%. BLS annual figures document the slowdown⁠.

The slowing economy was real, but attributing it primarily to AI would be scientifically unjustified. Employment was also affected by elevated interest rates, reduced labor-force growth, changes in immigration, post-pandemic hiring corrections, government policy, demographic aging and weakness in interest-sensitive industries. Technology companies that had expanded rapidly during the pandemic were simultaneously correcting earlier overhiring.

Aggregate employment statistics can also miss AI’s initial effects. Companies do not have to announce an “AI layoff” to reduce labor demand. They can leave vacant positions unfilled, eliminate jobs through attrition, consolidate responsibilities or hire two beginning employees where they once hired five. The first measurable effect may therefore be weaker hiring—particularly of recent graduates—rather than an extraordinary surge in dismissals.

What peer-reviewed research establishes

The strongest peer-reviewed evidence shows a mixture of substitution and augmentation.

Xiang Hui, Oren Reshef and Luofeng Zhou studied the introduction of ChatGPT, DALL-E 2 and Midjourney on a large online freelance platform. Their difference-in-differences analysis compared workers in highly affected occupations with less-exposed workers before and after the systems became available. Freelancers in exposed occupations experienced approximately a 2% decline in completed contracts and a 5.2% reduction in monthly earnings. Experienced and highly rated freelancers were not insulated from the effect. Published in Organization Science, the study provides credible evidence that generative AI can substitute for human labor in narrowly defined digital markets. It does not establish an equivalent effect across the entire economy. Read the peer-reviewed study⁠

Ole Teutloff and five coauthors analyzed more than three million postings on a global freelance platform. They classified 116 skill groups according to whether AI was likely to substitute for the work, complement it or have little direct effect. Demand for substitutable services such as basic writing and translation declined by approximately 20% to 50% relative to the estimated counterfactual trend. Short assignments were affected most severely. In contrast, demand for machine-learning programming increased about 24%, while demand for chatbot development nearly tripled. The results, published in the Journal of Economic Behavior & Organization, show labor being reallocated from readily automated production toward specialized complementary work. Read the peer-reviewed research⁠

The most important evidence of augmentation comes from Erik Brynjolfsson, Danielle Li and Lindsey Raymond. Their Quarterly Journal of Economics study followed the staggered deployment of a generative-AI assistant among 5,179 customer-support agents. Access to AI increased the number of customer problems resolved per hour by approximately 15%. The largest gains went to novice and lower-performing workers because the system helped them reproduce practices previously associated with more experienced employees. Customer sentiment improved, requests for supervisors declined and turnover among newer workers decreased.

The authors explicitly caution that the study cannot determine aggregate employment or wage effects. Nevertheless, it demonstrates why technical automation does not automatically eliminate an occupation. AI made employees more capable, improved service and reduced attrition. Management could use that productivity to reduce staff, but it could instead serve more customers, improve responsiveness or expand the business. Read “Generative AI at Work”⁠.

A 2025 PNAS Nexus study adds an important methodological warning. Morgan Frank and colleagues tested frequently cited AI-exposure scores against occupation-specific unemployment-insurance records. Individual exposure measures were generally poor predictors of actual unemployment risk. Combining multiple measures improved predictive performance, but the study still could not establish simple causation. The finding undermines forecasts that calculate the number of “exposed” jobs and then treat that number as expected job losses. Exposure is neither automation nor unemployment. Read the peer-reviewed study⁠.

Important evidence that is not yet peer-reviewed

The most current U.S. evidence comes from Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab. Their August 2026 working paper uses ADP administrative payroll records covering millions of employees through June 2026.

The researchers find no evidence of widespread economy-wide AI displacement. They do, however, identify a serious divergence among workers ages 22 to 25. Employment in highly AI-exposed occupations was 19% below the level it would have reached if it had kept pace with employment among similarly aged workers in less-exposed occupations. Employment among young workers in the two most-exposed occupational groups fell about 11% between November 2022 and June 2026, while employment among their peers in the three least-exposed groups increased approximately 10%.

The change occurred primarily through reduced hiring rather than increased separations. Experienced employees showed no comparable decline. Losses were concentrated in occupations where AI usage substitutes for human tasks; employment was stable or rising where AI complemented workers.

The study controls for numerous competing explanations and obtains broadly similar patterns after excluding technology companies and computer occupations. But its authors appropriately call the results descriptive indicators, not causal estimates. Some occupational divergence existed before ChatGPT, education explains part of the difference and the effect is more pronounced in the ADP sample than in national surveys. It is important independent research, but it should be identified accurately as a working paper. Read the August 2026 Stanford study⁠.

Another working paper by Anders Humlum and Emilie Vestergaard connects surveys of approximately 25,000 workers and 7,000 Danish workplaces with government administrative records. Despite widespread chatbot adoption, the researchers found average reported time savings of only about 3% and no statistically significant changes in recorded hours or earnings. Their results could rule out effects larger than approximately 1% during the period examined. The findings illustrate the gap between using AI and reorganizing a business sufficiently to produce large economic consequences. Read the NBER working paper⁠.

Bank of America tests the apocalypse theory

The recent TheStreet report on Bank of America’s analysis⁠ brought this evidence to a broader audience. Bank of America examined employment across 206 industries and compared job growth with estimated AI exposure. Highly exposed industries had generated little employment growth since ChatGPT’s release, while low-exposure industries grew by nearly 2%. Across all industries, however, the bank found virtually no correlation between AI exposure and employment growth.

The bank separately compared reported AI adoption with labor demand, defined as employment plus job openings. Information businesses reported 42.1% AI use and a 1.9% decline in labor demand between January and June 2026. Finance and insurance reported 34.8% AI use and a 1.1% decline. Those figures raise legitimate concerns, but the economy-wide relationship between AI use and labor demand remained weak.

At the same time, AI capital investment was creating employment elsewhere. Bank of America estimated that nonresidential construction added 95,000 jobs during the period studied and AI-related manufacturing added approximately 32,000. Together, these activities accounted for roughly one-quarter of new private-sector employment.

The headline that AI is “not stealing jobs” is therefore too absolute. Bank of America did not prove the absence of displacement. It found that AI exposure does not explain overall industry employment patterns and that the predicted national employment collapse has not occurred.

Jensen Huang’s different model of employment

Nvidia CEO Jensen Huang has consistently rejected the assumption that automating tasks means eliminating jobs. “Many tasks will be automated away,” he has said, while arguing that every job will change and new occupations will emerge.

Huang distinguishes a job’s tasks from its purpose. A radiologist’s purpose is not merely to inspect images; it is to help diagnose and manage a patient. A software engineer’s purpose is not simply to type code; it is to design and maintain systems that solve problems. If AI reduces the time required for individual tasks, demand for the broader human responsibility can remain or even increase.

Huang’s position is not that nobody will be displaced. His argument is that firms using AI will become more productive and competitive and may consequently need more employees to pursue expanded opportunities. Companies and workers that use AI may displace those that do not, but total employment need not contract. This task-versus-purpose framework is more consistent with the emerging evidence than either an assertion of zero disruption or a prediction of imminent mass unemployment.

Productivity, performance and economic growth

Productivity is output per unit of labor. AI can also improve performance beyond that narrow measure: decision quality, accuracy, strategy, innovation, customer service, responsiveness, execution and competitiveness.

The economic outcome depends on what companies do with these gains. A business that treats AI only as a cost-reduction instrument may eliminate positions. A company that uses it to improve products, accelerate development, serve customers faster and enter new markets may expand revenue and employment. Competitive markets exert pressure in both directions: reduce unnecessary costs, but also invest and grow when better performance creates new demand.

There will be dislocations. Routine and easily codified work faces genuine substitution. Entry-level pathways must be redesigned so businesses continue developing the experienced employees they will later need. Workers will have to acquire AI-related and complementary skills, while employers must preserve human judgment, accountability and institutional knowledge.

But displacement is not synonymous with permanent unemployment. The present evidence describes an economy in transition: fewer opportunities in some tasks, greater productivity in others, new infrastructure employment and growing demand for specialized skills. If productivity and company performance increase, economic output can expand, prices can fall, incomes can rise and new businesses can form.

The prediction of immediate mass unemployment overlooked the adaptive power of workers, companies and competitive markets. AI job losses are occurring—but not at the scale, speed or uniformity predicted. The more accurate conclusion is that AI is reorganizing work. Whether that transformation ultimately produces broadly shared growth will depend not only on what the technology can automate, but on how effectively businesses use its increased capabilities to create greater value, stronger performance and new opportunities