The problem with the tractor analogy is that there is not much that can go wrong in digging with a machine versus by hand. However, AI, like people, makes non-obvious mistakes that require a high level of skill to catch.

For instance, what we're seeing with pure vibe-coding is that even the best models make small mistakes, usually in the architecture but often in even lower-level abstractions, that compound over time resulting in unmanageable complexity. Senior engineers are able to spot and fix these issues early based on intuition, but that intuition has come from a lot of hands-on experience.

Using AI for a given skill is becoming a separate skill in itself, for sure. But this still requires having an understanding of the original skill itself.

I think this requirement won't change until we have super-intelligence, where AI is consistently much better than humans, i.e. beyond AGI. Because even at human-level performance, the only human that can verify an expert's work is another expert.