The bottom rung is gone

For a while the effect of AI on jobs was a guessing game. It isn't anymore. A Harvard study tracked résumé data for 62 million workers across 285,000 U.S. firms from 2015 to 2025 and found a clean, uncomfortable pattern: at companies that actively adopted generative AI, junior employment fell relative to firms that didn't, while senior employment in those same companies kept rising. The researchers gave it a name — seniority-biased technological change — which is an academic way of saying AI ate the bottom of the ladder and left the top alone.

And it wasn't layoffs. Companies didn't fire their juniors; they quietly stopped posting the jobs. People are calling it the broken rung.

The logic is seductive

It's easy to see why a manager makes this call. AI is best at exactly the work a junior used to be hired for: the boilerplate, the routine bug fixes, the scripted tests, the first draft of the obvious thing. So why pay a junior to do what the agent does in seconds? Hire seniors to steer the agents, automate everything underneath, ship faster with a smaller team. On this quarter's spreadsheet, it's obviously correct.

It's also how you quietly dismantle your own future, and almost nobody is pricing that in.

Seniors aren't hired. They're grown.

Here's the thing the spreadsheet leaves out. A senior engineer is a junior who survived a few thousand bugs, outages, and bad decisions. Seniority isn't a credential you buy on the open market; it's the accumulation of thousands of solved problems, fixed bugs, and averted crises — tacit judgment that only forms by doing the unglamorous work and living with the consequences. You cannot download it, and you cannot hire your way around it, because the supply of seniors is just the juniors of ten years ago who got the reps.

The junior years are the factory that produces seniors. When you shut the factory to save money this year, the seniors don't stop being necessary — they just stop being made. The shortage shows up later, on someone else's watch.

We're eating the seed corn

This is the same trap I wrote about with cheap code, scaled up to a whole profession: the saving is immediate and visible, the cost is deferred and invisible, and it lands on a different person — here, a different decade. Cut junior hiring across 2024–2026 and the analysts who model this expect the bill to arrive as a seniority cliff around 2031–2036, when the mid-level pipeline thins and then the senior pipeline behind it. Even Microsoft's own engineering leaders have started warning that AI is hollowing out the junior developer pipeline. We're trading a line item this quarter for a structural shortage next decade, and calling it efficiency.

The cruel part: we closed the school that teaches the scarce skill

Here's the vise. In an AI world, the most valuable thing a person brings is exactly what the model lacks: judgment — the ability to steer the agent, catch when it's confidently wrong, and integrate its output into something that actually works. Studies describe an "AI drag" on early-career workers who don't yet have that judgment and so can't supervise AI well. But the only way anyone ever developed that judgment was by being a junior, doing the work, and getting it wrong enough times to learn. So AI made judgment more valuable than ever and removed the on-ramp that produced it. We made the skill scarce and then closed the school. It's the same point I keep coming back to: knowing what to build is the whole game — and "what" is learned the expensive way, by doing the "how" until it becomes intuition.

The honest hedge

It's not uniform doom, and I won't pretend it is. Some of the early-career dip is remote-work and a soft market, not AI — an LSE analysis argues the AI share is smaller than the headlines. The Harvard figure is a relative decline, not a collapse. And some companies are betting the other way entirely — IBM is tripling its U.S. entry-level hiring, on the theory that an AI-augmented junior is now worth more, not less. "Junior" is being redefined more than erased: the bar is rising to "productive with AI from day one." But a higher bar is its own problem — how do you clear it without the years that used to build the muscles?

What to actually do

If you run a team, the cheap move is to automate the juniors and hire only seniors. The move that's expensive now and far cheaper over a decade is to keep hiring juniors and change what they do — not the boilerplate the AI now writes, but learning to direct it, review it, and own the outcome, on real problems, fast. Use AI as their tutor, not their crutch. The juniors who win this era won't be the ones who type code; they'll be the ones who build judgment early, with the agent as a sparring partner.

The pipeline is a choice, not a weather event. We can keep sawing off the bottom rung because it's cheaper this quarter, or we can build a new one that gets people to judgment faster. But seniors don't fall from the sky, and the decade that learns this the hard way is already on the calendar.