For at least three years, technology corporations and professional services organizations have assumed that AI would increase the productivity of junior personnel. The idea was that if a fresh analyst had access to big language models, they could complete tasks that would have taken a week in just a few hours. A junior developer would produce code twice as quickly as the previous year if they were given GitHub Copilot. The productivity argument was straightforward and correct in the limited sense of doing specific jobs fast. The Stanford study, which was released in August 2025, revealed the issue that the productivity argument had been hiding: AI’s current employment for junior employees is more than just work. It was instruction.
Erik Brynjolfsson and his colleagues at Stanford’s Digital Economy Lab examined payroll data from ADP, one of the biggest payroll processing companies in the US, which employs over 25 million people. They followed the employment trends of workers between the ages of 22 and 25 from late 2022, when generative AI started to be used commercially, until July 2025. Employment for that group decreased by 13% compared to prior levels in professions with strong exposure to AI.
The worst off were software developers between the ages of 22 and 25, whose employment dropped by over 20% from its peak in 2022. The employment of older developers, those between the ages of 35 and 49, increased by 6 to 9 percent over that time. The difference is striking, and it began nearly at the same time as ChatGPT’s introduction.
Important Information
| Field | Details |
|---|---|
| The Study | Stanford Digital Economy Lab working paper (August 2025) — led by economist Erik Brynjolfsson; used payroll data from ADP, covering over 25 million workers at tens of thousands of companies from 2021 through July 2025 |
| Core Finding | Early-career workers aged 22–25 in the most AI-exposed occupations experienced a 13% relative decline in employment from late 2022 to mid-2025; older workers in the same roles saw 6–9% growth |
| Software Developers (22–25) | Employment dropped nearly 20% from its peak in late 2022 — the sharpest decline of any measured group; older developers’ employment grew over the same period |
| Why Young Workers Specifically | Researchers’ explanation: AI replaces “codified knowledge” — the book-learning young graduates bring from universities; AI is less capable of replacing “tacit knowledge,” the accumulated experience and judgment older workers carry |
| Brynjolfsson’s Warning | “Young workers who learn how to use AI effectively can be much more productive. But if you are just doing things that AI can already do for you, you won’t have as much value-add” |
| The Skill Development Problem | Junior roles have historically been training grounds — repetitive tasks built foundational competence; when AI takes those tasks, the repetitions that create expertise never happen |
| The Pipeline Risk | Reduced junior hiring in 2024–2026 means fewer candidates for senior roles in 2031–2036; researchers and analysts describe this as companies “eating their seed corn” |
| Klarna as Cautionary Case | CEO Sebastian Siemiatkowski celebrated AI cutting staff from 5,500 to 3,400 and saving $10 million; by 2024–2025 the company began rehiring as the limits of AI-only operations became apparent |
| Class of 2026 Hiring | NACE projects only 1.6% increase in hiring for Class of 2026 vs. Class of 2025 — functionally flat when adjusted for the growing number of graduates entering the labor market |
The mechanism that distinguishes this from a typical job loss scenario is shown by the researchers’ explanation of why young people in particular are being displaced while older workers are not. They contend that AI is especially effective at replacing what they refer to as “codified knowledge”—the structured technical knowledge that a 23-year-old software engineering graduate possesses in large quantities but hasn’t had time to transform into something more difficult to replicate.
Another skill that seasoned engineers possess is tacit knowledge, which is the collected assessment of how unexpected system failures occur, why clients make demands that don’t make sense on paper, and when to push back deadlines. The function can be written by AI. It cannot inform you that a database migration that was not specified in the ticket will cause the function you asked it to develop to fail in three months.
This leads to a systemic issue with skill development that appears gradual but is actually urgent. There has always been more to junior roles than just the tasks they complete. When a first-year associate at a law firm drafts research memos, they are not really adding value to the firm; rather, they are learning how to think legally, differentiate a useful precedent from one that is superficially similar but inapplicable, and comprehend why a partner rewrote a section they thought was fine.
When responding to regular inquiries, a junior customer care representative is learning how to discern tone in written language, de-escalate without using a script, and determine when a problem has to be escalated. The junior employee receives the output but bypasses the process when AI does such jobs. Expertise is not developed by outputs without a process. as one Allwork analysis.To put it simply, space never produces the repetitions that lead to competency.
Klarna’s experience, which is frequently mentioned in this regard, demonstrated the true costs of eliminating entry-level positions in the name of AI efficiency. When AI enabled him to reduce the company’s workforce from 5,500 to 3,400, Sebastian Siemiatkowski openly rejoiced, calling it an example of operational efficiency.
After a few years, the company started employing again after realizing how difficult it would be to manage intricate operations without the institutional knowledge that junior-to-senior pipelines produce. In 2024, Salesforce CEO Marc Benioff declared that the company would not be hiring any new developers in 2025. After observing both cases, the industry is still figuring out which way the lesson should go.
It is worthwhile to explain out the harsh demographic math. By simple math, a large decline in junior hiring between 2024 and 2026 results in a corresponding decline in the pool of candidates for senior positions between 2031 and 2036. The tacit knowledge Brynjolfsson describes as AI-resistant has a minimum cultivation time that cannot be reduced by providing the following generation with better tools.
Businesses that do away with entry-level jobs in 2025 because AI can do the work are doing more than just saving money now. In a market where they will be competing with every other company that made the same calculation, they are cutting back on the number of experienced individuals they will need to hire in ten years.
As the Class of 2026 enters a job market with only 1.6% more entry-level positions than the previous year, it seems as though the industry is conducting an experiment without fully understanding the procedure.
