A career always worked like a ladder. The first rung: an internship, a junior role, a year of fetching things and watching how the seniors do it. You stand there a couple of years, you get underpaid, and you get taught. Then you climb.
Well, that first rung has just been sawn off. And the university is still selling four-year tickets to the top at business-class prices.
The number that made me sit down and write
Stanford's Digital Economy Lab (that's Erik Brynjolfsson's lab) took data from ADP, the largest payroll processor in the US. Millions of real people, not survey answers. Here's what came out: workers aged 22 to 25 in the occupations most exposed to AI saw employment fall 16% relative to older colleagues at the very same firms. Experienced workers held steady. It was the young who dropped.
For young software developers the decline by September 2025 was nearly 20% from the late-2022 peak.
And here's the detail that matters. This isn't a story about AI killing work. It's a story about AI killing the entry into work. That distinction is everything.
What a sawn-off rung looks like
The data adds up to one picture, and it's uncomfortably consistent.
- New-grad hiring at big tech is down roughly 65% versus 2019, and at early-stage startups 76% (SignalFire, 2026).
- Graduates of top-20 US computer science programs are 45% less likely to land an engineering role at a major tech company than they were a few years ago. Even the elite signal stopped working.
- From May 2025 to May 2026 senior-level postings rose 14.7% while entry-level postings fell 7.5% (Indeed Hiring Lab).
- In software development only 4.5% of postings are entry-level. For comparison, in personal care it's 91.3%.
- The Big Four are cutting graduate intake: Deloitte down 18%, EY down 11%, KPMG from 1,399 people to 942.
- Unemployment among recent college graduates is 5.6% against 4.2% nationally (NY Fed). Historically it ran the other way around.
The mechanism shows up in the AI companies' own data
Here's what settles the is-it-really-AI question for me. Anthropic publishes an Economic Index, a report on what businesses actually spend their models on. 77% of enterprise usage is automation, including handing a task over entirely. Only 12% is collaboration, where a human and a model work on something together.
Now think about what junior work consists of. Exactly the tasks you hand over whole: pull the summary, write the boilerplate, reconcile the sheet, review the documents. No company ever decided to cut young people. It decided that a model does this work now. Young people just happened to be sitting on it.
The irony is that this wasn't just work. It was the training mechanism. We didn't automate a task. We automated the practice field.
Now the university bill
If the entry point is breaking, the real question becomes what four years of preparing for that entry actually costs.
Americans owe 1.863 trillion dollars in student loans. That's 42.6 million borrowers, averaging 40,467 dollars each. College costs rose 169% between 1980 and 2019 while earnings for workers aged 22 to 27 rose just 19%. Paying for one year of public university on minimum wage now takes 1,345 hours of work, against 387 hours in 1980.
And most importantly, what you get for it. 52% of four-year graduates are working jobs that don't require a degree one year after graduating. Ten years out, 45% are still underemployed. And here's the number that gives me a chill: of those who start out underemployed, 73% are still stuck there a decade later (Strada and the Burning Glass Institute).
So your first job shapes your trajectory more than your degree does. And the first job is exactly what got sawn off.
Then there's the economics of the degree itself. The college wage premium climbed from 39% in 1980 to 79% in 2000 and has been flat ever since, slightly down by some measures (Minneapolis Fed). The reason is simple: supply outran demand. The share of the workforce holding a bachelor's went from 31% to 45%, while degree-requiring postings per non-degree posting fell from 1.2 in 2010 to 0.6 in 2020 (Cleveland Fed).
And the kicker. FREOPP calculates that 31% of American students are enrolled in programs with zero or negative financial return. For 23% of bachelor's programs the return is negative. Meaning the average graduate ends up poorer than if they had never gone.
But I'm not going to tell you to drop out. Here's why
It would be dishonest to sell you a clean conclusion. The data cuts both ways.
Unemployment among all degree holders aged 22 to 65 is 3.1%, among recent graduates it's 5.6%, and among young people without a degree it's 7.8%. So the problem isn't education, it's experience. A degree still beats no degree: median weekly earnings of 1,543 dollars against 930 for a high school graduate.
And the story about corporations dropping degree requirements is largely theater. Harvard Business School and Burning Glass audited 11,300 positions. After all the loud announcements, fewer than one hire in 700 actually benefited. About 97,000 people out of 77 million hires a year. Plenty of press releases, almost no practice.
And the counterargument I won't hide. The promised jobs apocalypse hasn't arrived. US unemployment sits around 4.2%. Dallas Fed research found that AI simultaneously reduces junior hiring and raises wages for experienced workers in the same occupations. There is no apocalypse. There is redistribution, and it lands on the entry point.
Jobs that didn't exist when you filed your application
Now the good part. The rung didn't vanish. It moved, and almost nobody is showing you where.
AI agent manager. Harvard Business Review formally defined the role in February 2026. One of the first people in the world with that title, Zach Stauber at Salesforce, describes his day simply: he starts and ends it in dashboards, scorecards and agent monitoring. He doesn't write code. He runs a fleet.
AI Orchestration Specialist. Average US pay is 148,233 dollars a year, with the top of the range going past 450 thousand. What it requires isn't programming but domain expertise. You have to understand a process well enough to break it down into agents.
Forward Deployed Engineer. The engineer who lives at the client's site and bends the system to their reality. Median 211 thousand at Palantir, above 785 thousand for seniors at OpenAI and Anthropic.
AI governance and ethics. Demand grew 150% year over year, ethics roles 125%. Median around 150 thousand, C-level from 300. The driver is obvious: regulation and the EU AI Act. Meanwhile 98.5% of organizations admit they don't have enough people for it.
Data Annotator. The fourth fastest-growing role on LinkedIn's 2026 list. The profession split in two: commodity labeling collapsed in value, while expert annotation became scarce. That's when a doctor, a lawyer or a trader teaches a model their own domain.
And the most underrated one: the physical world. Datacenter Technician sits at number 17 on that same LinkedIn list, and datacenter Commissioning Manager at number 11. Robot teleoperators earn 28.24 dollars an hour, and Figure AI pays humanoid operators 35 to 40. Siemens committed to training 200,000 electricians by 2030. The AI boom is also concrete, power and cooling, and there the shortage is hands, not diplomas.
There's one common denominator in all of this. The pay premium for AI skills over a comparable role jumped from 25% to 56% in a single year (PwC). Four of the five fastest-growing US jobs of 2026 are AI-related.
What I'd do if I were eighteen right now
I'm not an educator. I'm an investor and a founder, so I look at this as an allocation with a horizon and a risk profile.
One. I'd stop buying the degree as a guarantee. As an option on network, access and licensed professions like medicine, law and engineering, it works. As a promise of employment, it doesn't anymore.
Two. I'd chase the first real job, not the diploma. The data here is blunt: an internship cuts your risk of ending up underemployed by 49%. If your program doesn't put you in real work by year three, it's selling you an expired product.
Three. I'd compute the return on a specific program, not on higher education in general. The spread is enormous. Criminal justice majors run 65.8% underemployment, engineering runs a fraction of that. Higher education isn't one asset. It's an index, and a third of its positions lose money.
Four. I'd build public proof of work. A resume says you were trusted. A portfolio, a repo, a shipped product, a blog say you can do it. When entry slots are scarce, companies take the person whose proof already exists.
Five. I'd aim where AI complements rather than replaces. Stanford showed this in plain terms: employment falls where AI automates and rises where it complements. There's your map for the next decade.
And yes, the Thiel Fellowship, which pays 250,000 dollars to drop out of university, has quadrupled its grant since launch. Its alumni founded Anthropic, Figma, Ethereum and Scale AI. It isn't a prescription for everyone. But it's a very loud signal about where the price has moved.
The bottom line
Four years and a six-figure sum to prepare for the first rung made sense while that rung existed. It got automated first, because it consisted of exactly the tasks a model now does end to end.
This doesn't mean don't study. It means something else: stop paying for the signal and start paying for the capability. A degree used to be proof that you could. Now the proof is what you've already built.
Nobody removed the ladder. They just cut off its bottom rungs. Now you jump onto it instead of walking on.