Collapse of Junior Hires: What If It's Not the AI's Fault?

For a year now, anything written about youth employment starts from the same assumption, namely that AI is cutting entry-level jobs. It's a clean story, easy to tell, and it has an empirical reference that’s been circulating for months: the study from the Stanford Digital Economy Lab that measures a 16% drop in employment among 22- to 25-year-olds in the most exposed occupations. Then a few days ago, a paper came out that turned the tables.

It’s titled The Broken Ladder, authored by Peter John Lambert and Yannick Schindler, and it claims that this drop has very little to do with AI and a lot to do with another phenomenon that occurred after the pandemic: remote work. This paper should be read by anyone dealing with the labor market because it is seriously documented and because it hits an automatic reflex, attributing any surprising figure to AI without first looking at what’s alongside it. What I want to do here is not debunk it but read it thoroughly, including the two pages that almost no one has cited.

What the Paper Shows and Why It Holds

The reasoning of Lambert and Schindler is a textbook case of what mathematicians call "omitted variable bias", attributing everything to one cause merely because the adjacent one wasn’t considered. The occupations most exposed to AI — developers, accountants, consultants, technical writers — are almost exactly the same ones that went remote during the pandemic. How exact is that? A number tells us: the rank correlation between the two exposure rankings is 0.77, and at that point, a study focused only on AI risks attributing to it an effect that belongs to working from home.

To separate the two, the authors use a difference-in-differences comparison, meaning they observe how junior hiring changes in a highly exposed group compared to a less exposed group, before and after the shock. When measured one at a time, both AI and remote work predict nearly identical drops. Put in the same model, something interesting happens: the effect of remote work remains robust, while the coefficient for AI deflates, from -1.34 to -0.41 points, and in many estimates, it becomes indistinguishable from zero. And here, the authors deserve credit for serious work, as they dedicated about a quarter of the paper to verifying that the result wasn’t an artifact. They tested alternative exposure measures, removed one country at a time, and one occupational group at a time, calculated diagnostics for multicollinearity, and employed Monte Carlo simulations for measurement error. The result holds against nearly everything.

In 22 out of 24 combinations of measures, remote work weighs more than AI. This is not a paper that can be dismissed with a shrug.

The Paper Frees a Method, Not AI

However, here begins the part that journalistic coverage has flattened, and which significantly changes the conclusions. On page 4, Lambert and Schindler write something that should be printed in large letters: they do not interpret this evidence as proof that AI has no strong effects on work, and they openly declare they do not wish to undermine the literature linking AI to the decline in demand for juniors, as many of those studies use measurement strategies not based on occupational exposure. The declared target of the paper is studies that measure AI exposure, meaning with a desk estimate of how many tasks of a profession a model could impact.

The difference is not an academic nuance. The Harvard study by Hosseini Maasoum and Lichtinger, for example, does not measure exposure: it identifies companies that actually adopt AI by looking at those publishing ads for “GenAI integrators” and finds the same collapse of juniors concentrated precisely there. That type of study is one that The Broken Ladder states it does not want to touch. When it’s read that “a paper has shown that it’s not AI,” what the paper is being asked to say something it, with care, refuses to declare. It debunks studies built with a certain tool, not the phenomenon itself.

The Strong Test Only Applies to Half the Suspects

There’s a second point, and again, I’m not bringing this up, the authors admit it themselves. Their most robust experiment doesn't use exposure; it uses the actual adoption of remote work, measured company by company in 2021-22, before ChatGPT. It’s a strong test, and it works for remote work. However, there’s no symmetric test for AI in the paper, based on actual adoption rather than theoretical exposure. On page 11, the authors explain why: a measure of actual AI adoption at the same scale and coverage does not yet exist (“A symmetric measure of actual GenAI adoption would be a natural complement, but is presently not available at the scale and coverage of our WFH measure.”). And in the conclusions, page 22, they write that future work will require firm-level GenAI adoption measures with independent variation from WFH, and a longer observation period to see how the effects evolve over time (“Future work will require firm-level GenAI adoption measures with independent variation from WFH, as well as a longer post-period to track how these effects evolve over time.”).

In other words, the paper compares a cause for which it has an accurate measure with a cause for which it only has an approximate measure, and the better-measured cause wins. It’s an outcome that should be weighed for what it is, a partial result that the authors framework with honesty, and not the closure of the matter.

The Plateau and The Climb

The point I am most interested in lies in a tucked-away figure, D.1, in the appendix. It shows two curves. The first is AI adoption in American companies: in 2019, before ChatGPT, it was around 3%, then from 2023 it takes off and accelerates, with the index constructed on company spending data growing about six times in a couple of years. The second curve is remote work: it exploded in 2020, stabilized at a level many times higher than pre-pandemic, and has remained fairly flat since then.

These are two different forms of phenomena, and this changes everything. Remote work is a shock that arrived en masse and then stabilized on a plateau. AI is a curve that in the period studied by the paper, 2023-2025, is still climbing. And here lies the problem of the time window. When you compare the effect on hiring of a mature and stable cause with that of a young and rapidly accelerating cause, and you calculate until 2025, the mature cause has an advantage because during the observed period it had more time to unfold and a longer history on which statistics can work. That remote work “explains more” the decline of 2023-20225 is partly an effect of where the authors stopped the clock.

What a Clock Measures

Here enters the work I followed most closely, and which gave the title to my book. There exists an independent model evaluation group, METR, that measures a precise thing, the length of the task chain that a model can complete by itself before getting it wrong with a probability of one in two. That length roughly doubles every seven months, and has continued like this for six years. Another study, Crashing Waves vs. Rising Tides, built on actual job tasks rather than lab tests, projects that at current rates, models will perform 80-95% of most text tasks by 2029.

The point is that exposure to AI is not a snapshot; it is a derivative. The measure The Broken Ladder uses for AI is an estimate from 2024, capturing how many tasks the technology could impact then. But the size of the effect depends on how quickly that capability grows, and a snapshot from 2024 does not capture that. This is exactly what I mean when I put two words at the bottom of my book’s title: Humans for Now. The labor structure described by this paper holds as long as the time horizon of the models is short. The “for now” is the variable, and this paper, as solid as it is, is a photograph taken while the horizon is still short.

What to Make of It

I try to pull the threads together without having the numbers say more than they do. The reflex “it’s AI” is a lazy shortcut, and The Broken Ladder does a valuable service by nailing it down. The entry-level scale that broke in 2023-2025, with entry-level hires down between 14 and 29% in the four countries studied while senior hires went up, was probably cracked more by remote work than AI. On this, the data from Lambert and Schindler is serious, and I take it at face value.

What I do not take at face value is the reading that has been made, namely that the AI and work issue is closed. The paper does not say that, and the authors are the first to write that another measure and a different time horizon are needed. The scale for 2030 is a different matter from that for 2025, because in the meantime, the clock of the models continues to turn at a pace that can be measured. And the first step, from which a young person enters the profession and becomes tomorrow's senior, remains the most fragile point of the entire chain, whatever the cause that first cracked it. Serious choices, genuine training, and a school that trains judgment rather than procedure must be made now, while the horizon is still short and there is enough margin for intervention. Afterwards, it's called damage control.