Anthropic ran a study with 52 mainly junior engineers learning a Python library none of them had used before. Half got AI assistance and half did not. The AI group scored 17% lower on the assessment afterward, with the biggest gap on the debugging questions. They were not meaningfully faster either, because the time saved on writing went into composing queries and deciding what to ask.
The researchers' worry is the one I share. As more code is written by AI with people supervising it, those people may not have the skills to validate and debug what the tool wrote, if their own skill formation was shortcut by the tool in the first place.
I think this is the clearest evidence yet for a distinction that matters in how a team brings on new engineers. Using a tool to do the work is not the same as using it to learn the work.
My view, after years of thinking about what a new engineer's first months are for, is that the point of that time is comprehension: of the codebase, the domain, and the ways things fail. If AI is part of it, the question I would ask is whether the person can explain the change the tool made. If they cannot, the task was delegated, not learned.
How does your team decide when a new engineer should reach for the AI tool, and when they should not?
Photo source: https://photos.robertstowe.com/dolomites