Entry-level jobs were actually secret apprenticeships all along, and AI just cut the subsidy

I bring this up because there is widespread belief and panic that AI is destroying entry-level white-collar work and college degrees are no longer worth it. Some will cite work from Stanford’s Digital Economy Lab, which reports that workers aged 22 to 25 in the most AI-exposed occupations are running roughly 19 percent behind peers in less-exposed fields, a gap that has widened over the past year.

So, let’s consider the evidence and the timing. Zanna Iscenko and Fabien Curto Millet, analyzing 238 million job postings, tied the posting decline to the tightening cycle and noted that AI-exposed occupations tend to cluster in information, finance, and professional services, sectors especially sensitive to interest rates. The Economic Policy Institute adds that young workers without college degrees, whose occupations score negative on AI exposure, also saw their unemployment rate rise at a similar pace over the same period. If AI is the culprit in this employment decline, then why are unexposed workers suffering too?

And in fairness to the Stanford researchers, their own paper says what the typical media won’t: these are “descriptive patterns, not causal estimates,” and they “do not see widespread, economy-wide job displacement associated with AI.” They have also appropriately pushed back on my monetary policy story by pointing out that the most exposed jobs are not, in general, the most rate-sensitive, and that the employment gap for young workers in exposed occupations continues to widen, even as rates have come down.

The truth is that it’s too early to know for sure exactly what will happen to the labor market as AI continues to be deployed across our economy, because it is impossible to separate the impact of overlapping shocks in real time. That said, businesses and other organizations need to make expensive decisions right now about hiring, education, and regulation as if they already know the answer to this question.

So, as organizations struggle with the implications of AI for the workforce and the overall labor market, one thing has become very clear in the data. Firms are not firing their junior employees; they’re hiring fewer of them, and that decline in hiring concentrates where AI is automating work rather than where it complements it. And this fact may be one answer to the Stanford team’s rebuttal to my monetary policy thesis.

Consider why firms are not firing their junior employees, and ask what they were getting all those years when they hired these junior people? They clearly did not hire them for their output productivity. A first-year associate’s document was checked by a partner and frequently had to be redone, and the patient history the resident doctor took down at 2 AM often had to be retaken by the attending physician. By any honest accounting of the work, it was unproductive. But firms and organizations bought it anyway because that’s how you make a junior employee into a senior partner. The output of the process was the byproduct, and the formation of that employee was the whole point of the process, and what they produced helped offset the cost.

Into the picture comes AI that can now do the work of the junior analyst, so firms invest less in these hires, and the pipeline that turns juniors into seniors thins. Matt Beane watched this mechanism in operating rooms years before ChatGPT, as surgical robots quietly cost residents the case time that made them surgeons. Employers still want experienced people, but many have simply stopped funding the process that produces experience.

What should employers do? Stop treating junior hiring as a cost line that automation just erased. It was never an operating expense; it was a capital investment mislabeled, the mechanism by which the firm manufactured its own future partners. While AI may have taken some of the production value created previously by the junior employee, the value of the training remains, and it must now be funded more deliberately than is assumed. This means rethinking the training offered, the job rotations provided, and the mentoring and coaching they receive, all in the service of developing the judgment of tomorrow’s senior professionals.

Universities face the same problem from the other end, and I say this as a former dean and university president. The Ph.D. is an apprenticeship funded by the productive value of our apprentices’ work, and the editors of Nature warned this spring that early-career researchers now face the danger that “tasks that are crucial to their training as scientists are done by a machine.” The classroom evidence points in the same direction: students learn when the tool is constrained so that effort can’t be skipped, and fail to learn when it hands over answers. One of my old bosses, a Texan, had a great saying: “no friction, no traction.” Forming judgment requires friction, dealing with tough problems, failing, and then doing it again. Getting the right answer isn’t the point. As our beloved math teachers used to say to us, “Show your work.”

Now there is a version of this transition in which AI does the routine work, and an entire generation never gets the reps and experiences the friction that turns talent into judgment. Nothing in the technology makes that outcome inevitable. It arrives only if employers keep booking formation as a cost they can finally cut, and universities keep certifying work the machine did.

The postings data will recover when the hiring cycle turns; it always does. What will not recover on its own is the old bargain in which production quietly paid for formation. Rebuilding that bargain, on purpose and on someone’s budget, is the real AI question in front of us.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.