The missing junior developer: what entry-level hiring data shows in 2026

Ask a senior developer how they learned the job and you'll hear the same story in different clothes. Somebody gave them a small bug in a large system. They got it wrong, a patient colleague showed them why, and a year later they were the one explaining. That first rung, the dull ticket handed to the new person, is how the trade has always made its next generation.

It is also exactly the work an AI coding tool does well. So for three years people have argued about whether companies would stop hiring beginners. There is now enough data to stop arguing about whether, and to start asking how much, why, and what a small company should do while the large ones make up their minds.

What the numbers say

Payroll records

The most careful evidence comes from three Stanford economists working with payroll data from ADP, which covers millions of American workers. Their paper, first published in August 2025 and updated in August 2026 with data through June, compares young workers in jobs where AI is heavily used with young workers in jobs where it isn't. Employment of 22 to 25 year olds in the AI-exposed occupations "now stands 19% below where it would be had it kept pace with that of their less-exposed peers". A year earlier, measured the same way, the gap was 15% (Stanford Digital Economy Lab).

In plain counts, young workers in the most exposed jobs are down about 11% since November 2022, the month ChatGPT was released, while the same age group in the least exposed jobs is up about 10%. Experienced workers in the same occupations show no comparable gap.

Two details matter more than the headline. The first is how it happened. Nobody was fired: the paper finds the decline "operates primarily through reduced hiring of young workers rather than increased separations". The door narrowed. The second is where. The drop is concentrated in work where people use AI to do the task for them, and absent, or reversed, where they use it to help them do the task. Base pay shows no matching drop; the adjustment is in who gets hired.

Who gets hired

SignalFire, a venture firm that tracks career data on a very large number of people, published its yearly talent report on 22 June 2026. Against 2019, it puts hiring of new graduates and entry-level staff down roughly 65% at twelve large technology companies and down about 76% at early-stage startups (SignalFire). That second figure deserves a moment. The startup, the traditional home of the keen beginner who'll do anything, has pulled back further than the giants.

The graduates have noticed. In the same report, top computer science graduates of 2025 are twice as likely to call themselves a founder as the class of 2022, and 45% less likely to land at one of the big names.

Job adverts

Software development postings on Indeed are still roughly 30% below where they stood in February 2020. They have been rising again this year, but Indeed's own economists describe the rebound as one "with senior, AI-fluent roles driving nearly all of the recent gains" (Indeed Hiring Lab). The market for developers is recovering. The market for new developers is not recovering with it.

Graduates generally

For context, this is a hard year to leave university with any degree. The New York Fed put unemployment among recent American graduates at about 5.6% in the second quarter of 2026, with 42% of them working in jobs that don't need a degree (Federal Reserve Bank of New York). Software is a sharp case of a wider squeeze, not a world of its own.

Is it really AI?

The fair answer is: partly, and nobody can yet say how much. The years after 2022 also brought high interest rates, the end of a pandemic hiring binge, and a change to American tax rules that made software salaries more expensive to deduct. Any of those would hit junior hiring first.

The Stanford authors take those objections seriously. In a follow-up note in February 2026 they tested the interest-rate explanation and found that the jobs most exposed to AI are, if anything, less sensitive to rates than others (Stanford Digital Economy Lab). Their pattern survives removing technology companies and computer jobs from the data altogether, which a tech-only slump would not. But they also list what cuts the other way: some of the gap predates ChatGPT, it shrinks when they control for education, and it is stronger in their payroll sample than in national surveys. They call their findings "early, descriptive indicators", not proof of cause.

One footnote explains why this is so hard to settle. The main national survey of American households includes, in a typical month, between 23 and 48 software developers aged 22 to 25. You can't see a trend through a sample that small, which is why payroll and career-profile data carry the debate.

A second study looks at companies rather than occupations. Two Harvard researchers followed about 62 million workers at 285,000 American firms and marked the firms that began advertising for people to build generative AI into their operations. After that point, junior employment at those firms fell 7.7% within six quarters compared with firms that hadn't, while senior employment kept growing. Again the cause was slower hiring, not dismissal (Lichtinger and Hosseini).

Put together: the fall in entry-level hiring is real, it is larger where AI does the task outright, and it has widened every time someone has measured it. How much of it AI caused is still open. For someone deciding whether to hire, that distinction matters less than it sounds. The beginners are available either way.

The problem this creates later

A company that stops hiring juniors saves money this year and finds out the price in about five. Seniors are made, not bought: every one of them was somebody's junior. SignalFire's report says it bluntly: "By eliminating its new grad pipeline to optimize current balance sheets, the tech industry could face a severe leadership vacuum over the next decade."

Some large employers have started to act on that. In February 2026 IBM's head of people, Nickle LaMoreaux, said the company was tripling its entry-level hiring, "and yes, that is for software developers and all these jobs we're being told AI can do". Her reasoning: "The companies three to five years from now that are going to be the most successful are those companies that doubled down on entry-level hiring in this environment" (Fortune). IBM didn't keep the old job, though. It rewrote it: less routine coding, more time with customers and more time checking what the machines produce.

There's a nearer problem too, and it's the one small companies feel. AI tools write a great deal of code, and somebody has to read it. In Stack Overflow's 2025 survey, 46% of developers said they don't trust the accuracy of what these tools produce, and the commonest complaint, from 66%, was output that is "almost right, but not quite" (Stack Overflow). Code that is almost right is the expensive kind. It passes a glance and fails on a Tuesday night. The amount of code being written has gone up. The number of people learning to judge it has gone down.

What this means if you run a small company

You can hire people you couldn't reach three years ago. A ten-person firm used to lose every strong graduate to a bigger name with a bigger offer. Many of those offers no longer exist. The talent is the same; the queue at your door is different.

Hire for reading, not typing. The scarce skill is no longer producing code. It is looking at four hundred lines a tool produced in a minute and saying which twenty are wrong. In an interview, hand over a small change written by an AI tool with two real faults in it, one of them about who is allowed to see what, and ask the candidate to review it aloud. You'll learn more than from any puzzle.

Give them something to own. The old junior job was a queue of small tickets. The useful one now is a small area with a name on it: the import job, the billing emails, the admin screens. They use the tools to build it. They are also the person who gets asked when it breaks, with a senior beside them the first few times. Ownership is what turns a tool user into an engineer, and it is the part no tool supplies.

Put them near customers. IBM's rewrite is right about this. A beginner who has listened to three support calls writes better instructions for a coding agent than a brilliant one who hasn't, because they know what the software is for.

Budget the senior's time honestly. A junior with AI tools produces more than a junior without, and also produces more for someone to review. Plan for a few hours a week of a senior's attention in the first six months. If nobody has those hours, you aren't ready to hire a beginner, and you probably aren't reviewing your AI-written code either.

If you have no developer at all, the same question still applies to you. When a contractor or an agent builds your product, ask who reads the code and who will understand it in a year. "Nobody" is a common answer and a risky one. We wrote about the first checks to make in your AI-built prototype works; check these six things before customers log in.

What the data doesn't tell you

Almost all of this is American. The figures on hiring come from payroll, profile and job-advert data, each with its own blind spots; none of them is a census. They describe averages, and inside the averages are firms doing the opposite. And the measurements stop in mid-2026. If the tools improve enough to need less checking, the case for the reviewing junior weakens. If they keep producing more code than anyone can read, it strengthens. We'd bet on the second, but it is a bet.

What we would not do is read a 65% drop at twelve famous companies as an instruction. Large firms cutting their intake are solving a different problem, with a different balance sheet, than a small company deciding who will look after its product in 2030. The firms in Klarna, Duolingo and the AI-first memo learned that the saving on people arrived at once and the cost turned up later. And investors who admire small teams, as we covered in small teams, big revenue per head, are admiring output per person, not the absence of anyone under thirty.

Whoever builds for you, a beginner with good tools, a contractor or an agent, will do better work from a clear page than from a two-line message. We made a small form for writing one.

Sources

Brynjolfsson, Chandar and Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (Stanford Digital Economy Lab, version of August 2026, ADP data through June 2026); Stanford Digital Economy Lab, Canaries, interest rates and timing (9 February 2026); SignalFire, State of Tech Talent Report 2026 (22 June 2026); Indeed Hiring Lab, US labor market snapshot (23 July 2026); Federal Reserve Bank of New York, The Labor Market for Recent College Graduates (2026 Q2); Lichtinger and Hosseini, Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Resume and Job Posting Data (2025); Fortune, IBM is tripling its entry-level hiring (13 February 2026); Stack Overflow, 2025 Developer Survey press release.