AI Is Breaking the First Rung of the Career Ladder

Employment of Americans aged 22 to 25 in the occupations most exposed to AI now sits 19% below where it would be if it had kept pace with their peers in less exposed jobs. Experienced workers in those same occupations show no comparable gap. That is the headline finding of the August 2026 revision of a Stanford paper that has become the most cited piece of evidence on AI and work, and it says something narrower, and far more useful, than the "AI is taking jobs" story it usually gets attached to.

What the payroll data actually measures

The paper, "Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence," comes from Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab. It reads payroll records from ADP, the largest payroll processor in the country, covering millions of workers through June 2026. Each occupation gets an AI exposure score, and the authors compare employment trends by age and by exposure.

The first of their six facts is the one headlines skip: they find no widespread, economy wide job displacement associated with AI. The pain is concentrated at the bottom of the ladder. When the authors first documented the gap at the July 2025 data vintage, it stood at 15%. A year later it had widened to 19%.

In raw levels the picture is blunt. Between November 2022, the month ChatGPT launched, and June 2026, employment of 22 to 25 year olds in the two most exposed fifths of occupations fell about 11%. Same age group, three least exposed fifths: up about 10%.

A hiring story, not a firing story

The most important mechanism in the paper is also the least dramatic. The adjustment, the authors write, operates primarily through reduced hiring of young workers rather than increased separations. Companies are mostly not pushing juniors out. They are letting fewer new ones in.

That matters for how you read the news. Layoff trackers count announcements, and AI now tops the reasons employers give, yet those announcements remain a sliver of total layoffs, as our look at the layoff numbers showed. A job that is never posted leaves no trace in a layoff count. That is why the effect took two years to become visible, and why it shows up in payrolls rather than in press releases.

Pay tells the same story. So far, the paper finds, adjustment is showing up in employment rather than in base pay. Firms are not cutting the salaries of the juniors they hire. They are hiring fewer of them.

The jobs are not being destroyed in public. They are quietly not being created.

Why the junior goes first

The new revision adds a distinction that deserves more attention than it gets. Employment fell among young workers in occupations that rely heavily on codified knowledge: the formal, standardized, documented kind taught in textbooks and written procedures. It rose among experienced workers in occupations that rely on tacit knowledge, acquired through practice, mentorship and repeated exposure to real situations.

The logic is uncomfortable for anyone fresh out of school. The knowledge a graduate paid tuition for is precisely the knowledge a language model has already read. The knowledge nobody wrote down, which client is bluffing, why a machine makes that noise before it fails, is still scarce, and it sits mostly with people who have been around for a while.

The pattern lines up with how the tools get used. The declines concentrate in occupations where AI usage tends to automate human tasks. Where AI mostly complements workers, employment is flat or rising, particularly for the experienced.

Other data point the same way

The Federal Reserve Bank of New York tracks recent college graduates, defined as ages 22 to 27. In the second quarter of 2026, their unemployment rate stayed elevated at about 5.6%, and their underemployment rate, the share working in jobs that typically do not require a degree, edged up to 42%.

The International Monetary Fund has framed the same worry at a global scale. Its January 2024 staff analysis estimated that about 60% of jobs in advanced economies are exposed to AI, while noting that roughly half of those exposed jobs could benefit from it. Exposure is a forecast of change, not a count of losses, and the IMF number is often quoted as if it were the second.

The caveats the authors put in writing

The Stanford team is unusually candid about the limits of its own work. The gaps shrink when education is taken into account. Some differential trends were visible before generative AI spread. The estimated gaps are larger in the ADP sample than in national survey benchmarks, a discrepancy concentrated in education, health care and public administration. And after improvements to their data pipeline, estimates that account for overall changes in firm hiring stayed directionally consistent but became more sensitive to specification choices.

They also tested the obvious alternative explanations. The divergence survives excluding technology firms and computer occupations, controlling for exposure to interest rate increases and remote work, and swapping in different measures of AI exposure. It kept widening through mid 2026, well after interest rates peaked. U.S. Census Bureau administrative data show broadly similar raw patterns by age and industry. None of that proves causation, and the authors say so: they do not view this paper, or any single study, as definitive.

What no dataset can tell from the outside is whether companies replaced junior work with AI or simply froze junior hiring while waiting to see what AI can do. The two look identical in payroll records.

Climbing a ladder with a missing rung

For someone starting out, the codified versus tacit split is the most practical finding in the paper. Roles with customers in the room, physical systems, messy judgment calls or a senior person willing to let a junior shadow them are building exactly the kind of knowledge the data says is holding its value. A slightly worse title in a job that teaches things nobody wrote down probably beats a cleaner title in a job a model can already do.

For employers, the arithmetic is simple and easy to ignore. Every senior engineer, analyst or lawyer a firm will need in ten years is a junior it would have to hire now. A market where each company waits for the others to train people is a market that eventually runs out of experienced people to hire.

The likeliest reading of the current data is a pause, not a verdict: firms holding back entry level hiring until they understand what the tools can do. If so, the real decision has not been made yet, and the companies that keep training juniors through the wait will own the scarcest asset of the next decade, people who know how things actually break. Somebody still has to pay for the mistakes that create that knowledge. The open question is who.