Who trains the experts?

Here is a question the big forecasts rarely ask. How does anyone become an expert?

Mostly by doing beginner work. Junior lawyers review documents. Junior coders fix small bugs. New analysts build first drafts. That work teaches them what good looks like. Those beginner tasks are also the kind AI handles well. We will call this the apprenticeship problem.

There are early signs of it. Stanford researchers used payroll data covering millions of US workers. By mid-2026, employment for 22- to 25-year-olds in AI-exposed jobs was 19% below where it would have been. Experienced workers in the same jobs showed no comparable gap. The drop came mostly from less hiring, not layoffs. It showed up where AI replaces work, not where it helps with it. Losing a job or hours to new technology is called displacement (show definition).

Learning is at risk too. In a 2026 Anthropic experiment, people learned a new coding tool, some with an AI assistant and some without. On a later quiz, the AI group averaged 50% and the others 67%. The biggest gap was in debugging, the skill that catches mistakes.

But AI can also lift beginners. When a company gave 5,179 support agents an AI assistant, newer agents resolved 34% more issues per hour. So AI might help beginners perform better while they learn less. Which effect wins is an open question.

History offers one hopeful pattern. As ATMs spread from the mid-1990s, banks needed fewer tellers per branch. Branches got cheaper, so banks opened more of them, and teller jobs held up while shifting toward advising customers. When work shifts like this, people need reskilling (show definition): learning new skills to move into different work.