Is AI replacing entry-level jobs? The first rung is getting harder to reach.

A 2026 U.S. Census Bureau working paper used matched employer-employee administrative data rather than a survey of executive intentions. It found that employment among 22- to 24-year-olds in the most AI-exposed industry-state groups fell 12% on a regression-adjusted basis over the ten quarters after ChatGPT arrived. Less-exposed industries were stable. The drop came mainly through fewer hires, not a sudden wave of separations.

That is meaningful evidence. It is not a clean laboratory test of AI. The paper finds signs that hiring had already shifted around the pandemic, and its historical analysis suggests monetary-policy shocks may explain up to a quarter of the relative early-career employment decline through the second quarter of 2025. Remote work, higher interest rates and the unwind from pandemic hiring all happened in the same period.

One detail is especially easy to misread: the hiring rate largely recovered by early 2025, but the author says that recovery came from a smaller employment base. A healthier percentage does not automatically mean the missing jobs came back.

ZipRecruiter’s July report surveyed more than 1,000 U.S. hiring managers and talent-acquisition professionals. Thirty-eight percent said they had shifted basic data entry and processing away from entry-level workers and onto AI. Thirty-one percent said AI had raised experience requirements for entry-level roles.

At the same time, only 22% said their employer provides formal, mandatory AI training to everyone. Another 23% limits that training to certain departments. The message to a new worker is hard to miss: arrive knowing the tools, produce more quickly and somehow bring experience from a job we may no longer offer in the same form.

This is a disclosed-method employer survey, not payroll data. It records what hiring decision-makers say their organizations are doing. It cannot tell us how many jobs disappeared. It does show how the bar is being moved.

A July Stanford policy brief reaches a calmer conclusion at the economy-wide level. Unemployment rose by similar amounts in the most and least AI-exposed occupations between 2022 and 2026. The authors find little evidence that AI is causing large job losses across the labor market right now. They also note that firms adopting enterprise AI have often grown, and that AI can improve novice performance on some tasks.

Revelio Labs’ July tracker contains the same uncomfortable split. Demand for the most AI-exposed roles fell 42% relative to the least-exposed roles from October 2022, with the weakness concentrated at junior levels. Yet firms it classifies as AI adopters grew headcount 27% more than non-adopters over the period. Senior headcount at those firms grew 31%; junior headcount grew 6%. Revelio also notes that the adopting firms were growing faster before adoption, so the comparison is not proof that AI caused the growth.

A company can expand while making its front door narrower. Total headcount can look fine while the apprenticeship layer thins out. That is why the argument keeps sounding contradictory when it is often measuring different people.

A Centre for Economic Performance paper looked at 243 million new hires and 407 million online job postings across the United States, United Kingdom, Canada and Australia. When the authors measured generative-AI exposure by itself, it predicted a smaller junior share of hiring. Working-from-home exposure did too.

When the two were estimated together, the remote-work effect remained while the AI effect shrank sharply and was often statistically indistinguishable from zero. Their argument is blunt: occupations that can be done remotely also tend to contain the writing, analysis and information tasks generative AI can perform, so studies may be giving AI credit for a shift that began with distributed work.

That does not make the entry-level problem imaginary. It changes the repair. If remote work removed the side conversations, observation and low-stakes practice that helped new people learn, buying an AI course will not replace them. If AI removed the routine tasks, preserving the old job title will not replace them either.

For employers, the useful question is not whether an assistant can draft the first memo, clean the spreadsheet or answer the common ticket. It probably can. The question is how a new person learns judgment once those starter tasks are gone.

That may mean giving a junior worker a bounded customer case from intake through review, letting them compare an AI draft with the source material, or asking them to explain which recommendation they rejected and why. The assignment has to reach a real consequence without making the newest person responsible for an unreviewed high-stakes decision.

A company cannot automate the practice work and keep expecting fully formed seniors to appear three years later. If the software takes the first pass, someone still needs to teach the second look. That teaching time belongs in the cost of adoption.

The weak advice is to paste ‘AI proficient’ into a resume. Hiring managers already receive polished applications by the pile. A stronger proof is one small piece of finished work that shows the rough input, what the tool helped with, the mistake or weak assumption you caught, and the final decision you made yourself.

Keep the example close to the job. For support, turn a messy complaint into a reply and an internal note, then explain what you refused to assume. For operations, clean a small public dataset and show the check that caught a bad row. For design or writing, preserve the brief and one revision so the reviewer can see that you can change direction without starting over.

This is not a guarantee. People should not have to build an unpaid side company to qualify for a first paycheck. It is simply more legible than another certificate with no evidence of how you work when the AI answer is plausible and wrong.

Priya Rao would not let a company put ‘AI eliminated junior work’ on a slide after reading one chart. Job postings, hires, layoffs and employment are different measures. So are AI exposure and actual use. Her test is to follow the same role over time: applications, interviews, junior hires, time to independent work, pay and promotion. If a firm hires fewer beginners but cannot show that the work vanished, the missing labor may have moved onto senior employees or contractors.

Ivy Chen is less interested in winning the causality argument inside a team that is already changing jobs. If the assistant now handles the first draft, the manager has to name the new first assignment, who reviews it and what ‘ready for more’ looks like. Otherwise ‘entry level’ quietly becomes ‘experienced, but cheaper.’

Priya is guarding against a scary story outrunning the data. Ivy is guarding against uncertainty becoming an excuse to do nothing. Both point to the same practical standard: do not celebrate saved labor unless you can show where a new person now learns the work.

The current evidence does not prove that AI is causing mass unemployment. It does not clear AI of the entry-level squeeze either. The cleanest finding is narrower: young workers in exposed fields have had a harder time getting hired, and employers report automating some of the tasks that used to sit at the bottom of the ladder.

For a job seeker, that means showing judgment around the tool instead of competing with it on generic output. For a manager, it means designing a paid path from beginner to trusted contributor before the old path disappears by accident.

A career ladder can survive new software. It cannot survive if every employer decides someone else should build the first rung.