If you graduated this year into a market that felt shut, I understand the conclusion you probably drew. Listings thin out. Replies stop coming. In the quiet a story fills the gap: AI took the jobs, and everyone told to learn to code, or to model, or to analyze arrived a few years too late.

I do not read it that way. AI is not taking work away from people so much as it is changing which work is scarce, and it is doing that faster than any curriculum can follow.

Here is the mechanism as I understand it. In the short run AI has driven down the cost of a large block of routine cognitive work, and when the cost of something falls you need fewer hands for it, so firms hire more slowly into those roles and retire the older versions. That part is real. If your training pointed at that work, so is the pain. But the same capability is quietly creating a job that barely existed a few years ago and that almost nobody has been trained for. Demand climbs. Supply is close to nothing. And from the outside, a shortage of qualified people is hard to tell apart from a shortage of jobs. The slow hiring is the short-run story. The shortage is the long-run one, and it is the one you are graduating into.

In a million years

In a million years, would anyone have guessed that an econometrician’s training would have anything to do with a computer scientist building AI?

For most of the history of both fields it did not. The economist had the identification tools: how to tell a real effect from a confounded one, how to design a comparison that settles a question instead of decorating it. What she lacked was a way into the systems worth asking about. The computer scientist had the opposite problem: the systems, and a benchmark culture that treats a higher number as a finding.

AI knocked that wall down from the economist’s side first. Coding is easier now than it has ever been. Someone who can state what she wants measured can stand up a working pipeline in days rather than quarters. The barrier that kept her out of the machine room is the one that fell fastest.

That matters because of where AI is going next: fields where being right is not a leaderboard question. Radiology, lending, child welfare, drug safety. There the question stops being “does the model score higher” and becomes “does this system, on these patients, improve the decision, and how would we know if it quietly did not.” That is an experiment, and designing one is econometrics.

The hard parts of my own project were econometric problems in ML clothing

I can say that concretely because I spent this year inside it. I am a graduate student at Seattle University, and my research is a controlled experiment on a retrieval-augmented mammography report-generation pipeline: a system that pulls up similar prior studies and hands them to a vision-language model before it writes its report. It crosses four retrieval encoders, each evaluated frozen and again after contrastive fine-tuning, against three frozen decoders, and it was accepted at the Pacific Symposium on Biocomputing 2027, for the proceedings and an oral presentation, with my advisor, Dr. Wenjing Yang.

Almost none of the difficulty lived in the model code. It lived in the parts an economist would recognize on sight.

Start with the null. If most studies in a screening population fall into one category, a system that ignores the query and returns that category every time will look accurate. So the honest benchmark is the do-nothing one, built from nothing but how common the labels are. Clear that first, then talk. Naming the counterfactual is roughly the first thing an economics course teaches, and still not standard in machine learning benchmarking.

Then identification. My encoders differ in the image domain they were pretrained on, the thing I set out to study. They also differ in input resolution. The two are entangled, so a difference between encoders cannot be pinned on pretraining domain alone. The right move was not to bury that but to state it in the paper and mark the claim as provisional. Reporting the limit of what your design can identify is not a weakness. It is most of the discipline.

Then heterogeneity, which mattered most. Aggregate accuracy went up. But when I broke the gain out by category, nearly all of it landed on routine screening cases, not on the band where a radiologist decides whether to biopsy. An average that improves while the decisive subgroup does not is a familiar object in program evaluation: a headline effect carried by the group that needed it least. Reporting the aggregate alone would have been true and misleading at once, in a domain where that is how someone gets hurt.

None of that is exotic. It is the standard equipment of a field that spent decades being told its effects were not real.

Why the pair is scarce

Economists have a word for what is happening: complementarity. A technology that substitutes for one skill raises the value of the skills it cannot perform. AI writes the code. AI does not decide what the comparison should be, which subgroup the effect has to show up in, or which confound sinks the result. Those judgments are the complement. The market is short of people who can build a system and interrogate it with equal seriousness, because the two halves have never been taught in the same building.

I came at this from the other side too. Alongside the clinical work I built a nested-CES general-equilibrium model of AI and cognitive labor and turned it into an interactive data story, “Will AI Make Human Work Worthless — or Priceless?”, which won the Graduate Division of the CAUSE Student Data Scrollytelling Contest. The two projects look unrelated. They ask one question twice: what happens to human judgment when the mechanical part of thinking gets cheap.

What I would actually do, starting now

If you want something more useful than encouragement, here is the plan I would hand myself twelve months back. It is a year of work, not a life sentence.

1. Pick one high-stakes domain and learn its stakes. One, and stay in it. Learn what a missed case costs a real person, in that person’s terms. Stakes are what turn a metric into a decision, and you cannot borrow them from someone else’s paper.
2. Take the causal sequence seriously, and take it for the grade. Econometrics, causal inference, design of experiments, applied regression, panel data and difference-in-differences, instrumental variables, program evaluation. Pair it with the machine learning and statistical computing sequence. The pairing is the scarce good. Either half alone is common.
3. Let AI close the coding gap instead of fearing it. This hybrid is newly possible because the implementation barrier collapsed. Hand the model your boilerplate and your plotting, and spend the hours you get back on the design, which is the part it cannot do for you.
4. Build one rigorous project end to end, and publish how you did it. Not five demos. One study, with the question stated up front, the analysis committed to before you look, an honest null model, a subgroup breakdown, and a limitations paragraph you did not enjoy writing. Then write it up in public, including what failed. One project defended honestly says more than a portfolio of things that always worked.
5. If you cannot be the hybrid, go find the other half. A statistician who walks into the ML lab, or an engineer who brings an economist into the design meeting before the experiment runs instead of after it fails, produces the same scarce good. Collaboration is the fast path, and it is open to you now.

The part I want you to keep

I am not going to promise you a market. I cannot see anyone’s hiring plans, and I will not pretend the next year is easy. Hiring has slowed, and saying otherwise would be a comfortable lie.

Here is what I keep running into. Every time this research got hard, the hard part was not the model. It was a question about design, about what we are entitled to claim. Those questions multiply as AI moves into places where being wrong is expensive, and nobody is producing the people who can answer them fast enough.

So the silence you are hearing is not the market telling you it does not need people. It is the market not yet knowing how to ask for what it needs. Go learn to be the answer to a question no one has finished writing down.