AI Is Not Firing Juniors. It Is Just Not Hiring Them.

Open a job board and run the same search twice. First, look for “software engineer.” Then add “0-2 years.” The first list is crowded and competitive. The second is thin enough to make a new graduate wonder whether the page loaded correctly.

That gap explains more than the claim that AI is coming for everyone's job.

US tech employment has been falling for almost four years. Yet the decline began before the current AI boom. Software job postings peaked in February 2022, months before ChatGPT appeared, according to Vithushan Sylvester's analysis of the tech labor market. Interest rates and the end of pandemic-era overhiring caused the initial crash.

AI helps explain the recovery that followed. Companies are posting more software roles without rebuilding the old headcount. They are concentrating the rebound in senior jobs and leaving many entry-level seats empty.

The term de-juniorization describes a career ladder losing its bottom rung one unfilled requisition at a time.

Read the recovery, not the crash

Since January 2025, US software development postings have risen 16 percent, while information-sector employment has fallen 5 percent. Those figures come from Sylvester's analysis. Together, they show why the word “recovery” needs care. Job ads can rebound while actual employment keeps shrinking.

The mix of those ads matters even more. Senior roles received 71 percent of the rebound in US software postings. The same pattern appears outside the US. Sylvester reports shrinking entry-level demand in China and roughly 1 percent year-over-year growth in fresher hiring in India, compared with stronger growth for senior roles.

His sharpest summary is straightforward: “companies are not firing engineers in large numbers. They are not replacing the ones who leave, and when they do post a role it is a different role.”

For workers, that distinction is practical. A mass layoff is visible. Attrition is quiet. A junior role disappears when somebody leaves and the replacement request never opens. A team of eight becomes seven, then six, while the company points to stable retention and open senior positions.

AI assistants fit that operating model. They can draft routine code, tests, documentation, research and first-pass analysis. A senior employee can review those outputs because they already know what good work looks like and where failures hide. The assistant reduces the blank-page labor. It does not supply the judgment that years of work produced.

Entry-level employees were hired partly to do bounded work while learning the system around it. AI now handles some of that bounded work cheaply. The company still needs judgment, but has less reason to fund the apprenticeship that creates it.

Build apprenticeship into the AI workflow

Hiring managers should stop treating junior development as an accidental by-product of ticket queues. If assistants absorb the small tasks that once trained new staff, teams need to design a replacement learning loop.

Start with paired ownership. Give a junior a bounded production task and require three artifacts: their plan before using an assistant, the assistant's proposed output, and a short review that names what they accepted, changed or rejected. A senior then reviews the reasoning as well as the final patch. The goal is to make judgment visible instead of grading only whether the code passed.

Next, reserve some work for unassisted diagnosis. Juniors still need practice tracing unfamiliar code, reading logs and forming a hypothesis before a model offers one. Without that practice, the assistant can hide weak understanding until an ambiguous failure appears. A simple rule works: write down your diagnosis and the evidence for it before asking the assistant for alternatives.

Finally, change the hiring signal. “Used ChatGPT” says almost nothing. Ask candidates to scope a task for an assistant, inspect a plausible but flawed answer, and explain how they would verify it. The useful skill is controlled delegation: choosing the right task, providing context, testing the result and knowing when to take the work back.

This approach costs senior time. So did the old apprenticeship model. The difference is that companies now have to budget for teaching explicitly because routine work no longer does the teaching for them.

The market is already putting a price on this ability. Sylvester's analysis cites a 56 percent wage premium for roles requiring AI skills. That number is not a reward for prompt tricks. It reflects demand for people who can combine domain knowledge with AI-assisted execution.

Some employers are making the test explicit. Proton is re-interviewing staff about their AI capabilities. One company's policy does not establish a universal practice, but it shows how quickly AI fluency can move from optional training to an employment criterion.

Make yourself expensive to leave out

Junior engineers and career-switchers cannot wait for employers to repair the ladder. Build evidence that you can do the new version of the job.

Choose one real workflow and document it end to end. Take a bug, a data-cleaning task or a small feature. Show the initial diagnosis, the instructions given to the assistant, the output it produced, the checks that failed and the final result. A polished demo proves less than a visible correction loop. Employers need to see that you can catch a convincing wrong answer.

Learn the team's verification tools as seriously as the model. For software work, that means tests, type checks, linters, logs, diffs and code review. For research, it means primary sources, dates and claim-by-claim citations. For operations, it means approvals, audit trails and reversible actions. Assistant fluency without verification is fast guessing.

Senior staff have a different assignment. Break work into pieces an assistant can complete, define acceptance checks before generation and keep risky actions behind approval gates. Review the failure modes, not merely the polished answer. The senior advantage is knowing when the task is underspecified and when an answer that looks finished is still wrong.

At the speculative edge, Delaware legislators have considered a framework for AI-managed companies that could transact, sign contracts and file lawsuits through agents. The proposal faces a liability dispute and remains far from everyday employment. It is useful mainly as a boundary marker. Most companies are not replacing the whole firm with agents. They are asking fewer experienced people to cover more work with assistants.

If companies stop hiring beginners, they reduce the supply of future seniors. The practical response is to rebuild apprenticeship around planning, verification and supervised ownership, then hire for evidence of those skills. Otherwise the industry will save money on junior seats now and discover later that it failed to train the people it depends on.