The loudest predictions about artificial intelligence and work have focused on unemployment. If machines can write code, summarize documents, produce presentations and answer customers, the argument goes, millions of white-collar jobs should eventually disappear.
The labor market has not produced that dramatic result. U.S. unemployment was 4.2 percent in September 2026. An NBER working paper examining recent college graduates found no statistically significant increase in their relative unemployment during the summer of 2026, including when the researchers expanded the measure to include people who said they wanted a job but were not counted as unemployed.
That evidence is a reason to reject claims that an AI-driven jobs collapse has already arrived. It does not settle a quieter question: whether companies are changing who they hire at the bottom of professional career ladders.
Several recent studies suggest that this is where the first effects may be easier to see. A related essay on this site examines the risk that AI will weaken the first rung of the career ladder.
The evidence changes when employment is separated from unemployment
A Stanford Digital Economy Lab working paper using payroll data from millions of U.S. workers finds no evidence of widespread economy-wide job displacement. Within highly AI-exposed occupations, however, employment among workers aged 22 to 25 stood about 19 percent below the level it would have reached if it had kept pace with similarly aged workers in less-exposed occupations. More experienced workers did not show a comparable gap.
The researchers are careful about what this does and does not prove. The measure is relative, not a claim that AI directly eliminated 19 percent of young workers’ jobs. The economy has also experienced a weak hiring environment, changes in technology investment and sector-specific adjustments. The paper nevertheless finds that the divergence appears to operate mainly through lower hiring of young workers rather than unusually high separations.
A second Stanford working paper examines 1.25 billion job postings and 154 million employment records across 41 countries. At foreign affiliates of companies adopting generative AI, the junior share of employment declined relative to comparable firms. The change was driven primarily by faster growth in senior employment rather than an outright collapse in junior employment, with suggestive evidence of modest overall employment growth.
Those findings help explain why aggregate unemployment can remain stable while the structure of opportunity changes. A company can continue growing, continue hiring and still bring fewer inexperienced workers into some occupations relative to the number of senior workers it adds.
The career ladder can narrow without the labor market looking like a crisis.
Entry-level work produces more than entry-level output
Routine junior tasks are easy to dismiss as low-value work. A new analyst cleans data, a young lawyer searches precedents, a junior programmer fixes simple bugs, and an assistant produces first drafts that more experienced colleagues revise.
From an employer’s perspective, many of those tasks are costs. If AI can perform them faster, reducing the amount of junior labor looks economically sensible.
The difficulty is that early-career work has historically served another function. It exposes inexperienced workers to real cases, errors, feedback, client expectations and organizational judgment. The work being produced today is also part of the process that produces a more capable worker tomorrow.
This is the career-ladder problem. It is not enough to ask whether a task can be automated. Employers also have to ask whether removing the task removes a useful stage of learning, and whether some other training process will replace it.
That outcome is not predetermined. New evidence suggests AI can sometimes expand what early-career workers are able to do rather than simply eliminating their work.
AI can also make apprentices more capable
A field experiment released by the Stanford Digital Economy Lab on October 7 studied 673 final-year IT apprentices at twelve German vocational schools. Participants were randomly assigned access to generative AI while completing occupation-specific tasks.
AI improved performance across task types. On assignments deliberately designed to go beyond the apprentices’ formal training, access to AI increased successful performance by 26 to 31 percentage points. The researchers also found no reduction in immediate task comprehension.
The experiment is a working paper, and its limits are important. It concerns IT apprentices in Germany, not the entire entry-level labor market. It measures immediate comprehension rather than whether workers develop deeper expertise years later. It therefore cannot prove that AI strengthens long-run skill formation.
It does demonstrate that automation and learning are not automatically opposites.
A junior employee who forwards an AI-generated answer without understanding it is receiving a different apprenticeship from one who uses AI to attempt harder work, checks the result with an experienced colleague and remains responsible for errors. The same tool can substitute for a learning opportunity or extend it, depending on how the job is organized.
That puts part of the problem back on employers.
Companies may discover the training problem late
The incentive to automate basic work appears immediately. A firm can reduce labor hours, speed up routine production and allow experienced staff to supervise more output.
The cost of training too few beginners would emerge much later.
Senior workers available in 2026 were trained under an earlier system. A firm can reduce graduate intake today while continuing to benefit from accountants, engineers, analysts, managers and lawyers whose experience was accumulated over many years. If too little new expertise is being created, that shortage would not necessarily appear in next quarter’s financial results.
There is not yet solid evidence that an economy-wide shortage of future senior workers will occur. That remains a risk and an inference, not an observed outcome.
The point is narrower. Unemployment statistics are poorly designed to reveal changes in the mechanism through which workers gain experience. A healthy headline employment number can coexist with weaker entry routes inside particular occupations.
The U.S. Department of Labor’s decision in 2026 to integrate AI skills into Registered Apprenticeships reflects one possible response: treat AI use and skill development as something that should be designed together rather than assuming workers will learn whatever is necessary on their own.
The useful question is not whether AI “takes jobs”
The current evidence is too mixed for a simple verdict.
The NBER study of recent graduates finds no statistically significant AI-driven unemployment spike. Stanford payroll data find a widening employment gap for young workers in highly exposed occupations. The 41-country study finds a decline in junior share that is driven mainly by stronger senior growth. The German experiment shows that AI can help early-career workers perform more advanced tasks without reducing the immediate comprehension measured in the study.
These results can all be true at the same time.
They point toward a labor-market change that is more specific than mass technological unemployment. AI may alter which tasks companies assign to beginners, how many beginners they hire and how quickly new workers are expected to become productive.
The statistics worth following therefore extend beyond unemployment. Entry-level vacancies, hiring by experience level, progression rates inside AI-exposed occupations, the amount of structured training employers provide and the share of work that remains available for supervised learning may become more informative.
AI does not need to eliminate an occupation to change a career. It can change the route by which someone becomes experienced enough to stay in it.
Sources
- U.S. Bureau of Labor Statistics — Employment Situation, September 2026
- NBER — The Early Impacts of AI on Employment among Recent College Graduates
- Stanford Digital Economy Lab — Canaries in the Coal Mine?
- Stanford Digital Economy Lab — How Does AI Change Labor Demand? Evidence from 41 Countries
- Stanford Digital Economy Lab — Task Expansion with Generative AI
- U.S. Department of Labor — AI Skills in Registered Apprenticeships