Most organisations do not really train people. They put juniors on small, low-stakes tasks and let competence accumulate. Nobody writes that down as a strategy, but it is how most professionals actually learned their job. That way in is now under pressure from both ends, and the evidence arrived quickly.
What the payroll data shows
Stanford's Digital Economy Lab analysed ADP payroll records covering three and a half to five million workers a month from January 2021 to September 2025. Their paper, Canaries in the Coal Mine, found employment for software developers aged 22 to 25 fell by nearly 20% from late 2022 to September 2025, while employment for older workers in the same occupations held steady or grew.
The finding survives the obvious objection. With firm-level controls, so you are comparing within the same employers rather than across a shrinking industry, they still find a 15 log-point decline in relative employment for 22 to 25 year olds in the most AI-exposed occupations, with no statistically significant effect for other age groups. It is not simply that those companies were shrinking.
The most actionable detail: employment fell where AI automates a task and held up where AI augments one. That distinction is a design choice, not a law of nature. How you build the workflow determines which side you land on.
The authors are careful in a way that deserves repeating. They write that their estimates "may be influenced by factors other than generative AI" and that the results are consistent with the hypothesis rather than proof of it. It is a working paper using one payroll provider's client mix, and it is not peer-reviewed.
Stanford's 2026 AI Index reports independent work finding declines of 15% to 16% for early-career workers in exposed fields, concentrated in hiring pipelines and the youngest workers.
Two reasons to be sceptical of your own conclusion
The Budget Lab at Yale looked at Current Population Survey data through December 2025 and reached a flatly different conclusion: "the picture of AI's impact on the labor market that emerges from our data is one that largely reflects stability, not major disruption at an economy-wide level." They also point out that the occupational shifts people attribute to AI began in 2021, before ChatGPT existed.
And employers say they are about to hire more graduates, not fewer. NACE's Job Outlook 2026 Spring Update projects a 5.6% increase in new college graduate hiring, bucking two weak years. That is an intention survey of 185 employers, so treat it lightly, but it does not support a collapse narrative.
Both things can be true. The Bureau of Labor Statistics projects software development growing 15.8% through 2034 while Stanford shows its entry point narrowing. An occupation can grow while the way into it gets harder to find. That reconciliation is more useful than picking a side.
The finding that should change how you design training
This is the part with direct consequences for L&D. Researchers at Anthropic ran a randomised experiment on how AI affects skill formation. 52 developers learned an unfamiliar programming library, half with AI assistance and half without.
The AI-assisted group scored 17% lower on a comprehension quiz, a large effect statistically, with the biggest deficit in debugging. They also showed no statistically significant productivity gain: 19 minutes with AI against 23 without. They did not learn it as well, and on this task they did not finish meaningfully faster either.
Then the useful part. The researchers classified six patterns of how people actually interacted with the AI. Three produced poor comprehension: delegating wholesale, leaning on it progressively, and iterative debugging without understanding. Three preserved learning: generating then working to comprehend, alternating code with explanation, and asking conceptual questions. That last pattern scored highest and was the fastest of the three.
Their conclusion, and it is worth quoting: "AI-enhanced productivity is not a shortcut to competence."
Handle this evidence carefully. 52 people, one narrow task, a preprint that has not been peer reviewed, and the interaction sub-groups are tiny, between two and seven people each. It is a hypothesis worth testing rather than a settled finding. It is also, as far as I can tell, the only real evidence behind the advice everyone gives about asking the model to explain itself. Notably, it is an unflattering result published by an AI company.
The trap this creates
Put the evidence side by side. AI delivers its largest productivity gains to your most junior people, 34% for novices against 14% on average in the customer support study, and 21% to 40% for junior developers in the MIT trials. And the way most people use AI to get those gains appears to reduce what they learn from the work.
So the tool that makes your juniors productive today may be the thing that stops them becoming seniors. Meanwhile the tasks you used to hand them are the ones being automated first. If your capability model quietly depends on apprenticeship, that model is being squeezed from both directions at once.
What to actually do
Decide deliberately which tasks stay human because people learn on them, and accept the efficiency cost as a training investment rather than pretending it is free. Teach the interaction patterns, not just the prompts: ask for explanation, predict before you look, debug without assistance sometimes. Assess comprehension separately from output, because output no longer tells you what someone knows. And when you design a workflow, ask whether AI is automating the step or augmenting the person, because the payroll data suggests that choice shows up in headcount.
None of this requires believing the strongest version of the disruption story. It only requires noticing that the way people build competence has changed, and that nobody has updated the training model to match.
On sourcing. Every study links to its source and every caveat above comes from the researchers themselves. Nothing here establishes that AI caused recent graduate unemployment. Neither the New York Fed nor Yale supports that claim, and Stanford explicitly declines to make it. Two figures circulating widely are absent because I could not trace either to a primary source: a 129% higher unemployment rate for junior developers, and a claim that 30% of entry-level work hours can be automated.