A Stanford analysis of payroll records covering 25 million workers found that employment for people aged 22 to 27 in jobs most exposed to AI has been shrinking since late 2022, while employment for older workers in the same jobs kept rising. Recent college graduates in the United States now face an unemployment rate above 5 percent, higher than the rate for the workforce overall. That gap has held steady for months, pointing to something structural rather than a temporary blip.
The jobs squeeze, though, is the easier problem to fix. Roughly a third of companies that laid off workers because of AI have already brought some back, workforce research firms tracking the trend say. Hiring freezes thaw quickly once a CEO changes their mind. A generation that never built the underlying skills cannot be fixed as easily.
A study of nearly 1,000 high school students in Turkey shows exactly how this happens. Students given open access to a GPT-4 chat tool did far better on practice problems while using it. But once the tool was taken away and they sat a real exam, they scored worse than students who never had access to AI at all. The tool let them produce correct answers without learning the material. A second version, which gave hints instead of direct answers, produced the opposite result: students kept their strong scores even after the tool was removed.
That difference, between a tool that does the thinking for you and one that makes you do the thinking, matters everywhere. Researchers studying robotic surgery have found that when hospitals adopt surgical robots, the lead surgeon operates the controls directly and the trainee, who used to assist with hands-on work, mostly watches. The result is fewer hands-on repetitions for trainees, and the most motivated young surgeons end up practicing on their own time just to keep pace.
Office work has a quieter version of the same problem. A manager at a professional services firm described facing the same choice almost every day: slow down to teach a junior team member properly, or fix the work herself because a client is waiting. AI has made this harder by raising what clients expect in less time, pushing managers toward the fast option by default. Multiply that across thousands of managers and you get junior staff spending less time struggling through problems, the struggle that used to build real skill.
None of this shows up in a quarterly report. Job losses get tracked by economists and make headlines. A slow decline in how well junior employees learn their craft does not show up anywhere until years later, when a company finds it has no one ready to step into senior roles. Executives surveyed by Atlassian are far more likely to spend money on AI technology than on developing their teams' AI skills, suggesting most companies have not connected these dots yet.
The fix is not complicated, even if it takes discipline. Before automating a task with AI, ask what a new employee actually learns by doing it manually, and protect that task even when the software could do it faster. Use AI to give employees more frequent feedback rather than replacing the judgment calls a manager should make. Build guardrails into AI tools so they act like a tutor asking hard questions rather than an answer machine handing out finished work. Companies that get this right will have a real advantage in a few years, when competitors let a whole cohort skip the learning curve and run out of people who actually know how to do the job.