A study out of Stanford this year put numbers on something many of us have been sensing. Using payroll data from the largest payroll provider in the country, Brynjolfsson, Chandar, and Chen found that employment for early-career workers in the most AI-exposed occupations has fallen 13 percent relative to their peers since late 2022, while experienced workers in the same occupations held steady or grew. The decline concentrates where AI automates the work instead of augmenting the person doing it.
The piece of that finding I keep turning over is not about this year's juniors. It is about 2031. A person who does not get hired in 2025 does not become mid-level in 2028 and does not become the senior you are trying to hire in 2031. The pipeline does not have a fast-forward button, and the industry is quietly deciding, one frozen entry-level requisition at a time, to find out what happens without one.
Engineering is at least having this argument out loud. My corner of the industry is not, so let me make the design version of the argument here.
What junior designers actually did
I spent thirteen years at ThoughtWorks, and a meaningful part of that time was spent turning juniors into leads. Watching that crossing happen dozens of times taught me what the apprentice years are actually made of, and it is not what job descriptions say.
A junior designer's listed output was production work. Spacing fixes, component variants, redlines, the fortieth state of a form nobody senior wanted to draw. The real product was different. Every one of those tasks was a low-stakes mistake waiting to happen, and the mistakes were the education. You watched your tidy form design collapse in a usability session. You shipped a pattern and saw support tickets teach you what you missed. A lead pulled your work apart on a Tuesday and you learned to tell the difference between criticism of the work and criticism of you. Almost none of this is teachable in advance.
It compounds into the thing we call judgment, and judgment is the entire job description of a senior.
Here is the uncomfortable part. That apprentice work, the variants, the redlines, the fortieth state, is a lot of what generates well now. I watch it happen on my own project. Agents do the apprentice-shaped work on my learning engine, and they do it fast, without needing the work to teach them anything. The economic case for hiring a person to do it gets thinner every quarter, and the Stanford data suggests companies have noticed.
The mistake hiding in the efficiency
If the apprentice work goes to machines, the production loss is small. The apprentices were never the engine of output. What disappears is the mistake-making, and with it one of the most dependable ways the industry had to grow judgment.
I do not think most organizations are choosing this consciously. Each individual decision is locally sensible. Why hire a junior to draw states an agent drafts in minutes? But sum five years of locally sensible decisions and you get a cohort that never existed, and seniors are rarely hired into existence, they are mostly grown out of cohorts. The companies skipping juniors today are spending down an inheritance of judgment that earlier, less efficient years paid for, and the bill arrives around 2031 with no sender's address on it.
Can judgment grow without the years of mistakes?
The honest answer is that I do not know, and I distrust both confident camps.
Maybe the apprenticeship just moves. The juniors I would hire today spend their early years reviewing machine output instead of producing drafts, and the mistakes that teach them happen in the review queue. Something like this happens on my own platform, where the review work is where all the learning lives. But reviewing requires the very judgment it is supposed to build, which is the old chicken-and-egg of senior work arriving before senior skill. An apprenticeship of pure review might produce people who can spot wrongness without being able to produce rightness. I genuinely cannot tell yet.
Or maybe the ladder shortens. The optimistic case says tools collapse the years between novice and competent, the way they have for solo builders, and 2031's seniors will simply have grown faster. There is something to this. The person who builds three real products with agents by age 24 has seen more full lifecycles than I had at 30. What that person has not seen is consequences arriving slowly, a design decision rotting over two years, an organization bending a product against its intent. Some lessons are time-locked, and I suspect the time-locked ones are disproportionately what we pay seniors for.
What I would do about it
If I ran a design org today I would keep hiring juniors and change what I hire them into. Pair every junior with the machine output stream, make them the person who must say why something is wrong before a senior confirms it, and protect a slice of real production work for human hands, deliberately, the way teaching hospitals protect procedures for residents even when the attending is faster. It costs efficiency on purpose. Teaching always has.
I might be wrong about the scale of all this. Official projections still expect these professions to grow over the decade, and the industry has absorbed scares before. But growth in headcount does not contradict a hollowing of the entry rung, the Stanford data shows exactly that combination, and only one of those problems fixes itself. If the seniors of 2031 are going to exist, somebody has to be paying for their mistakes in 2026. I am curious, and a little worried, about who that somebody turns out to be.