The Bottom Rungs Are Going, and the Data Says Which Ones
Entry-level employment in AI-exposed jobs is down 16% for 22–25 year olds. The useful part of the Stanford finding is not the headline — it is that the losses cluster precisely where AI automates rather than augments. That distinction is a career strategy.
- #career
- #research
- #job-market
Brynjolfsson and colleagues at the Stanford Digital Economy Lab ran payroll records from the largest payroll software provider in the United States and found a 16% relative decline in employment for 22–25 year olds in AI-exposed occupations, after controlling for firm-level shocks. Same occupations, more experienced workers: stable or still growing.
That headline has been everywhere. The finding underneath it has not, and it is the one you can actually act on.
The distinction that matters
The declines are concentrated in occupations where AI is more likely to automate rather than augment human labour.
Not "AI-exposed" as a blanket category. Exposure alone does not predict who loses work — the shape of the exposure does. Software development is heavily AI-exposed, and a senior engineer using Claude Code all day is exposed in the augment direction. The junior whose actual job was producing the first draft that someone else reviews is exposed in the automate direction. Same tooling, opposite outcomes, and the payroll data separates them cleanly.
So the question to ask about your own role is not "is AI coming for this?" It is: does AI make me faster, or does AI produce my output and leave someone else to check it?
Three more facts worth carrying
- It adjusts through headcount, not pay. The paper finds adjustment happens primarily through employment rather than compensation. Nobody gets a warning shot in the form of a smaller raise — the role just stops being opened.
- It is not a tech-sector story. The result holds when technology firms are excluded, and when occupations amenable to remote work are excluded. Those were the two obvious alternative explanations, and both were tested.
- It is accelerating. The Canaries Dashboard — 4.6 million workers, 730+ occupations — has employment for 22–25 year olds in the most exposed occupations shrinking at about 3.8% a year, up from 2.8% in April 2024 and now past 4%. Brynjolfsson's read: "Whatever it is, it's not going away."
What this does to the shape of a career
The traditional professional structure is a pyramid: a wide base of juniors doing routine work, narrowing toward senior roles. Take out the base and you get a diamond — most people entering one or two rungs up, managing the AI that does what juniors used to do.
That is efficient this quarter and expensive later, because the pyramid base was never only about output. It was where judgment got built. An organisation that stops hiring juniors has quietly stopped manufacturing its own seniors.
Where this leaves you
The advice being handed to juniors — "learn AI" — is nearly useless, because the people losing these roles are the ones AI is best at replacing, and being fluent in the tool that replaces you does not help.
What actually moves you is changing which side of the automate/augment line your work sits on:
- Take the verification seat. Reviewing, debugging, judging correctness, owning what ships — the model produces candidates, someone accountable decides. That work is augment-shaped by construction.
- Get close to a domain. Work whose difficulty is knowing the business, the constraint, the customer is not the work being automated first.
- Own an outcome, not a task. Tasks get automated. Outcomes need someone to answer for them.
None of that requires seniority. It requires picking work where the AI needs you in the loop, rather than work the AI produces and someone else checks.
Do this today
Write down the three things you did this week that took the most time. For each, answer honestly: did AI make me faster at this, or could AI have produced this and had someone else review it? Anything in the second column is your exposure. Move one of those items into the first column this month — by taking the review seat on it, or by trading it for work that needs the context only you have.
Sources
Keep reading
More guides like this one.
The AI Skill Worth Learning Now: Context Engineering
Prompt engineering was never really a job, and the models got good enough to do it for you. The skill that's actually sticking is context engineering — here's the full playbook, the six levers, a copy-paste scaffold, and three drills to build it this week.
AI Certifications a Hiring Manager Actually Reads
As an EM who screens resumes, most AI certificates tell me nothing. A few tell me something useful. Here is the honest difference, and how to make any course actually count.
The AI-Native International Job Hunt
How a dev or EM outside the US can use AI to land a role abroad — not by faking skills, but by closing the gap between what you have done and how a foreign hiring manager reads it.