Sometimes jobs disappear without anyone getting fired or laid off. A junior employee leaves. The company does not reopen the requisition. The remaining team absorbs the work with help from AI. Headcount falls through attrition.
A new revision of Stanford’s “Canaries in the Coal Mine?” tracks this pattern in employment data. Between November 2022 and June 2026, employment among workers aged 22 to 25 fell about 11% in highly AI-exposed occupations. It grew about 10% in less-exposed work.[1]
The gap came from hiring: exposed young workers were not leaving or being fired at higher rates, but fewer of them were joining firms.
Stanford defines a hire as a new worker–firm match, and it cannot see open requisitions or replacement decisions. The study may also be missing some important factors at play. A generally terrible market for new entrants, pandemic overhiring, education mix and sector effects remain possible explanations.
Related clues can be found in a study of ChatGPT Enterprise covering 1,764 organizations and 17.4 million messages. Active early-career workers and trainees sent about eight to nine more messages per week than the average active user in the same firm. The study measures usage, though—not productivity, output or headcount.
Juniors use AI heavily, AI increases their capacity, and firms need fewer juniors.
California may begin collecting the missing evidence. SB 951 defines a “technological cessation in hiring” as permanently ending hiring for a role because of AI or automation. If passed, covered employers would report positions they have decided not to fill, along with the work and technology involved. That would capture organizational capacity that disappears before anyone occupies it.
The career-system problem is that its ladder is made from work. First drafts teach framing. Basic analysis teaches judgment. Research teaches source evaluation. Debugging teaches how systems fail. Coordination teaches when to escalate.
Removing those tasks, annoying as they may be to senior people,[2] weakens the system that creates expertise, and the productivity everyone assumes in exchange may not be there. A small randomized study of working developers—most with seven or more years of experience—found that AI users learned less from the task, losing conceptual understanding, code reading, and debugging ability, without finishing meaningfully faster.
Organizations may keep today’s experts while quietly reducing the number of people who can become tomorrow’s.
