When we talk about artificial intelligence and work, we almost always ask the same question: how many jobs will it replace?
It is a reasonable question, but perhaps we are looking at the change too late. Before a technology eliminates an entire profession, it can do something much quieter: reduce the number of people who manage to enter that profession.
We do not see an office emptying out. There is no headline announcing thousands of layoffs. One fewer junior role simply opens. Then another. And another.
For a while, nothing seems to happen. Teams keep functioning because they still have experienced professionals. Until an uncomfortable question appears:
If we stop hiring juniors, where will tomorrow's seniors come from?
A signal worth paying attention to
In August 2026, the Stanford Digital Economy Lab published an update to its research on employment and artificial intelligence, based on U.S. payroll data from ADP.
The first result helps lower the anxiety: the researchers do not find widespread, economy-wide job displacement associated with AI.
But separating the data by age and occupational exposure produces a different picture. Among workers aged 22 to 25 in highly AI-exposed jobs, employment is approximately 19% below where it would be if it had followed the path of less-exposed occupations. There is no comparable gap among more experienced workers.
One detail matters even more: the adjustment appears to be happening mainly through reduced hiring, not increased layoffs.
That changes the interpretation. Companies may not be visibly replacing people who already know the work. They may be choosing not to bring in people who need an opportunity to learn it.
The Stanford research is careful: it describes patterns and does not prove that AI is the sole cause. Some of the gap shrinks after controlling for education, there were earlier trends, and a U.S. sample cannot automatically be generalized to every country.
We should not turn one study into a prophecy. But we should not ignore a signal merely because it is not yet a verdict either.
Junior work was never just production
Someone can study programming, law, marketing, finance, or design. They can learn concepts, procedures, and tools. But when they reach their first job, they discover that the real world is far less orderly than any course.
The client explains the need poorly. The data is incomplete. Two documents contradict each other. The solution that looked perfect depends on an old system nobody wants to touch. A technically correct decision can be entirely wrong for the context.
That is where judgment begins to form.
In programming, it might start with fixing a simple bug, reading someone else's code, or preparing a test. In a law firm, with reviewing documents and researching precedents. In marketing, with drafting a campaign and studying what happened. In customer service, with resolving routine situations before receiving a delicate case.
Many of those tasks produced something useful for the company. At the same time, they were developing the person doing them.
That second result rarely appeared in a spreadsheet. We did not measure the intuition gained from reviewing a change with someone more experienced. We did not calculate the value of making a mistake in an environment where the consequences were small and repairable. Nor did we record those brief conversations in which someone explained why an apparently good solution was not appropriate.
We saw the deliverable. We did not see the education happening around it.
Knowing something is not the same as having lived it
The Stanford research distinguishes between codified and tacit knowledge.
Codified knowledge can be written in a manual, documentation, or a procedure. It is information that can be stored, searched, and explained—exactly the kind of material AI models process relatively well.
Tacit knowledge develops through practice, observation, mistakes, mentorship, and repeated exposure to real situations. It is recognizing that something is wrong before you can fully explain why. It is knowing which question is missing. It is noticing that this case resembles the previous one except for the detail that changes everything.
A model can help a junior produce a result that would once have taken hours. It can generate the code, analysis, draft, or initial response. That is valuable.
What it does not automatically generate is the experience that person would have accumulated during the process.
We can obtain an answer without developing the ability to recognize whether it is a good answer. That difference starts to matter when the problem stops being standard, the instructions are ambiguous, or somebody has to take responsibility for the decision.
A rational decision that can create a collective problem
From a company's perspective, the temptation is perfectly understandable.
If a senior professional equipped with AI can perform work that once required several people, hiring fewer juniors looks efficient. There is less training, less review, and faster output.
The problem appears when many companies reach the same conclusion.
Each can think: training someone costs too much; I would rather hire a person who already has experience. But if almost nobody offers the place where that experience is acquired, a few years from now they will all compete for an increasingly small pool of trained professionals.
It is like eating the seed because doing so is more profitable today than planting it.
And there is an even larger paradox. Companies could massively increase their productive capacity with AI while weakening the system that generates the human judgment needed to supervise, guide, and take responsibility for that production.
Not because the technology is bad. Because we optimize the short-term result and leave tomorrow's capability out of the calculation.
We should not romanticize old junior work either
It would be easy to argue that new professionals must suffer through the same repetitive tasks as those who came before them. I do not think that is the answer.
Not every junior task was educational. Many involved copying information, waiting for approvals, adjusting formats, or repeating processes that taught very little. If AI can remove that part, all the better.
Moreover, the same tool that threatens one route into a profession can create another. Someone can request explanations, practice with simulated cases, review a draft, translate documentation, or build a prototype without needing a large team and budget from day one.
I have learned a great deal myself by using models as partners for research, programming, and turning ideas into projects. The opportunity is real.
So the question is not how we stop AI from performing entry-level tasks. It is this:
Which parts of those tasks developed judgment, and how can we replace them with a better experience?
AI can close the door if we use it only to avoid hiring. It can also open the door if we use it to build new paths for learning. Technology does not choose between those options by itself. The way we design work does.
A junior is not a slower senior
I think this is one of the fundamental mistakes. If we measure a junior only by what they deliver today, they will almost always lose against a combination of experience and artificial intelligence.
But an entry-level hire is also an investment in future capability.
That requires changing the experience. Assigning a simple task and waiting for the result is no longer enough. If AI can solve it in seconds, training must move toward the reasoning around that result.
Someone can use AI to write code, but they should then be able to read it, test it, find its assumptions, and defend the change to another person. They can generate an analysis, but should explain what evidence could disprove it. They can prepare a customer response, but need to recognize when the case is no longer routine and must be escalated.
This is not about banning the tool to prove that somebody can work “the old way.” It is about using it without allowing it to hide the thinking process we want to develop.
If work no longer provides practice, we will have to design it
For a long time, training happened alongside production. A company paid for a task and, almost without intending to, also created a more experienced professional.
If AI absorbs many of those early repetitions, practice will have to become explicit.
We can use historical cases, anonymized incidents, simulations, test environments, and projects with controlled consequences. We can ask someone to compare several model responses, identify the errors, and explain what information is missing. We can evaluate decisions rather than deliverables alone.
Real human mentorship will also be necessary.
Placing someone in front of a chatbot does not replace a conversation with a person who knows the craft. An experienced professional can reveal what is not written in any documentation: why this client needs a different conversation, why a correct solution is not appropriate, or why one apparently minor detail defines the entire problem.
Of course, that takes time. Training has always taken time. The difference is that the cost used to be hidden inside tasks that also produced value. Now we will have to make it visible and consciously decide whether to invest in it.
The challenge does not belong only to beginners
This problem will not be solved by telling young people to learn more tools or write better prompts.
Companies will have to revisit their entry programs and stop measuring immediate productivity alone. Educational institutions will have to expose students earlier to ambiguous situations, incomplete decisions, and work with real consequences. Those of us who already have experience will have to learn to transmit some of the knowledge we often keep without even noticing that we possess it.
And people who are starting out face a challenge of their own: using AI to expand their practice rather than skip every uncomfortable part of learning.
There is an enormous difference between reaching an answer quickly and developing the ability to judge it.
The first can already be bought as a service. The second is still built through time, attention, conversation, and responsibility.
The entry door can change, but it must continue to exist
We do not yet know whether the gap Stanford observes will grow, stabilize, or eventually reverse. We do not know how much is due to AI and how much to other economic movements. Presenting a definitive conclusion would be irresponsible.
But the mechanism is credible enough that we should act before we can see a crisis.
The real question is not how to preserve every junior task unchanged. It is how to avoid removing the learning that happened around those tasks when we automate them.
If we do this well, we can remove repetitive work and build a better entry path: deliberate practice, mentorship, real problems, simulations, and AI that helps people understand.
If we do not, we may find ourselves a few years from now with extraordinary tools, enormous productive capacity, and a very strange shortage: people who had the opportunity to learn how to take responsibility for what was produced.
Seniors do not appear through spontaneous generation. They were once juniors whom somebody gave time, context, and real problems.
The entry door can be different. What we cannot afford is for it to disappear.
In this video, I develop the idea and ask how we can redesign that first professional stage without giving up everything AI already enables us to do.