Stanford Data Finds AI Is Hollowing Out Entry-Level Jobs, Not the Professions Themselves

A July 2026 brief from Stanford's Institute for Economic Policy Research, SIEPR, put a number on something a lot of people in tech have suspected for a while: workers aged 22 to 25 in AI-exposed occupations saw a 16% relative employment drop in under three years. The occupations named specifically are software development, customer service, and accounting.

Entry-level software-developer job postings are down nearly 20% since 2022. That figure alone would be easy to read as "AI is shrinking these professions." The more interesting part of the SIEPR data says otherwise.

The part that doesn't fit the simple story

Older workers in those exact same fields, software development, customer service, accounting, show flat-to-growing employment over the same period. If AI were simply shrinking these professions, senior employment should be sliding too. It isn't. What's shrinking is specifically the entry-level rung, while experienced workers in the identical field are holding steady or doing better.

Why researchers are calling it a paradox

SIEPR's researchers frame this as the junior-gap paradox because of what it implies about how anyone becomes a senior worker in these fields in the first place. The tasks that used to make up a junior developer's or junior accountant's job, the repetitive, lower-stakes work that builds judgment over a few years, are the same tasks AI systems now handle well. Companies still want senior people. They're just increasingly not creating the entry-level positions where senior people used to come from.

The structural question nobody has answered

This isn't a story about AI eliminating jobs in aggregate; the data doesn't support that reading. It's a story about AI removing a specific rung on a specific ladder while leaving the rest of the ladder intact. That raises a question with no obvious answer yet: if the traditional path into software development, customer service leadership, or accounting no longer runs through several years of junior-level task execution, what replaces it. Nobody, including SIEPR, has a settled answer. But the employment data makes clear the old path is already narrowing, whatever replaces it hasn't been built yet.