A first-year associate at a firm using Harvey, the legal AI platform now deployed at firms including Allen & Overy, opens a contract that the model has already redlined and spends her morning correcting its errors rather than drafting the clauses herself. The work is genuinely useful to the firm. It teaches her something narrower than the apprenticeship it replaced.
The traditional internship was an inefficient but functional apprenticeship: a young person did unglamorous, low-stakes work under supervision and absorbed the informal knowledge of how a profession actually operates. A growing share of entry-level work now looks different from the outside, same title, same reporting structure, same entry-level pay, but the object of the labor has changed. Reviewing a model's draft teaches judgment about what looks wrong. It does not teach the slower skill of generating a first draft from nothing.
This shift is easy to miss in aggregate hiring statistics, because entry-level roles are not disappearing at the rate some predicted. Scale AI's own annotation workforce is the extreme version of the same pattern: an entire labor category built around correcting and rating model output as a business model in itself, rather than around producing anything from scratch.
What breaks is the implicit contract that made internships worthwhile for both sides: the organization got cheap labor, and the intern got a genuine, if inefficient, education. When the labor is data preparation for a system that will eventually reduce demand for the intern's own future role, that exchange looks less like an apprenticeship and more like extraction, even when no one involved intends it that way.
Some firms, aware of the tension, have started assigning junior staff to draft contract language unassisted before they ever see Harvey's version, specifically to preserve the generative muscle that reviewing alone would not build. That is a real cost, paid in billable hours nobody gets to invoice, and most firms are not choosing to pay it.
An internship was always a bet that today's unglamorous work builds tomorrow's expert. The bet still makes sense. It just requires someone to deliberately protect the unglamorous part, now that the easiest version of it has been automated away.