1. The cheapest work in an office was also the tuition. Before anyone trusted you with something that mattered, they handed you the work no one would miss if you ruined it: a document to read, a model to build, a ticket to answer, a bug to fix. Dull, low, and always checked by someone above you. That was not the waiting room before the job. That was the job, teaching you.
2. The paycheck is the smaller loss. When the first rung thins, everyone counts the missing salaries. Count instead the missing place to be wrong on purpose, where a supervisor catches your mistake before a client, a court, or a customer ever sees it. That place is where judgment got built. That is the thing actually going.
3. The most educated workers are starting out worse than everyone else. In March, unemployment for degree holders aged twenty-two to twenty-seven was 5.8 percent while the country's overall rate was 4. More school is supposed to buy a better start. Right now it is buying a worse one. Every explanation that follows has to fit that inversion.
4. It is not the weather. This is not a passing season. Overall unemployment is low, so this is not best explained as a broad labor-market collapse. The grads are not quitting the search: their labor-force participation has held and their underemployment is flat, so they are still after degree-level work. Some of the rise is just hiring cooling from its pandemic spike. Not enough of it.
5. The machine eats tasks, not jobs. It does not swallow an occupation whole. One AI lab mapped a million of its own chat logs onto the tasks that make up real jobs: about 37 percent were computer and math work, code changes, debugging, troubleshooting, and only about four percent of occupations showed the tool touching three-quarters of what they do. It starts with the bounded, checkable ones.
6. The first rung and the machine's strong suit are the same list. Read the tasks of a first-year anything, then read what these tools are sold to do. Both say: summarize the document, draft the memo, fix the bug, write the test, build the chart, check the contract against the playbook, clear the routine ticket. The bottom rung was built out of exactly the work that is easiest to hand a machine, because "bounded and reviewable" is what "safe for a beginner" always meant.
7. No one chose to end apprenticeship. They decided it looked optional. The tool never had to replace the junior. It only had to make the junior's first pass look skippable to a manager already told to run lean. New graduates are now 7 percent of hires at the largest tech firms, down by more than half from 2019. Listings for the entry-level corporate jobs young grads used to get are down 15 percent, and there are 30 percent more applicants for each one.
8. The junior stopped being the cheap default. At one large commerce company, the chief executive told staff they must first prove a job cannot be done by AI before they may ask for more headcount. Put that question over every empty desk: what would this team look like if the agents were already on it? The junior hire used to be what you reached for without thinking. Now the junior hire is what you justify.
9. The machine may teach faster than the grind ever did. This is the honest countercase, and it is strong. Give five thousand support agents an AI assistant and resolved issues per hour rise about 14 percent, with almost all the gain landing on the newest workers, up about 35 percent. A two-month rookie with the tool handled calls like a six-month veteran without it. The system had learned what the best agents did and fed it to everyone else, one call at a time, faster than any manager could.
10. In code, the beginner gains the most. Hand developers an AI assistant and the least experienced improve fastest: recent hires and juniors lifted their output by 27 to 39 percent, the seniors far less. In a clean lab task, assisted developers finished about 55 percent faster. The first year used to disappear into hunting syntax and boilerplate. The tool hands that time back.
11. It raises the floor faster than school can. Give weak performers a good model and they jump most. In one consulting experiment, on work inside the tool's range, the lowest-ranked people improved about 43 percent and the whole group's output was rated 40 percent better. In another, one worker with the tool matched a two-person team without it. The old way made you wait years for exposure to the expert down the hall. The new way sits a synthetic one at your elbow on the first morning.
12. So maybe the job just starts higher up. Some employers are already redrawing the rung upward. One accounting firm now gives new graduates the higher-level tax work that used to wait for people with two or three years in. One law firm trains its youngest lawyers on hard contract interpretation instead of only document review. In one survey, more than 60 percent of executives said they plan to hand entry-level work to AI. The pitch is a better first job, not a missing one.
13. We have strong proof the tool lifts the work, and none that it lifts the worker. Every one of those wins measures the same thing: output this quarter. Not one measures whether the person built judgment that lasts a career. The coding study that showed juniors gaining most could not even see the quality of the code they shipped, only how much of it there was. The consulting team that raised the floor admitted, in its own writeup, that it still had to find out whether anyone had learned anything. The work got better. Whether the worker did is unmeasured.
14. The machine is most confident exactly where the beginner is most lost. Push that same consulting model onto problems just past its range and the people leaning on it were right 60 to 70 percent of the time; the people working without it were right 84. Ask the leading legal-research tools real questions and the better ones were wrong or misgrounded more than 17 percent of the time, the worst more than a third. The hazard is not a machine that fails loudly. It is one that is fluent, plausible, and wrong, handed to the single person in the building without the experience to feel the difference.
Judgment was always made out of permitted error.
15. Judgment was always made out of permitted error. Look at how anyone actually got good: a small piece of real work, done imperfectly, caught by someone who knew what was missed. A reviewer marking a young engineer's code with a note that is half fix and half lesson. A senior banker changing the slide and, in the change, showing which number the room will attack. The teaching was never in the task. It was in the correction, and the correction needed a first draft bad enough to correct.
16. There was no golden age to mourn. The old rung could grind as hard as it taught. Junior bankers surveyed a few years ago described running on no sleep and treatment that wore them down. Much first-year work was tedious, slow, and unevenly watched. The point is not that it was good. The point is that it was where you were allowed to be bad, and it is being removed before anyone has said what takes its place.
17. The contraction is uneven, and that is the tell. This is not one flood covering every field. Tech new-grad hiring fell hard. Accounting graduate listings dropped 44 percent in a year while the big firms cut their intakes. But hiring of new lawyers did not collapse; last year's class posted the highest employment rate on record, even as starting pay slipped. The rung is not vanishing everywhere. It is being pulled up wherever the first tasks turned cheap to automate, and left alone where they did not. Trace that line field by field and you can see where the machine made the old first tasks look optional, and where it didn't yet.
18. The whole argument fits inside one summer job. A young man is running his hometown pool for the season. He is twenty-two, holds a mechanical engineering degree he finished this year, and has nearly two hundred applications behind him and no offer in his field. On the recording you can hear the ordinary sounds of the place, someone calling about the front gate, someone asking if he is doing all right. "I was told by a lot of people that I was going to get a job right out of college," he says, "and then all of a sudden there's no jobs." No one can prove a machine took his. He walked up to the first rung and found it gone.
19. We removed the safe place to fail before we learned whether we could. Somewhere, soon, a first-year worker may take an answer from a model that is confident and wrong, and ship it, with no one senior near enough to catch it, because the task that used to seat someone senior beside her is the task that got automated. Maybe she learns to check it. Maybe she only learns to trust it. We are about to find out, on a whole generation at once, whether you can grow judgment in a person who was never allowed to be usefully bad at anything first. The pool stays open until Labor Day.