I assume junior devs can at least search. AI often doesn't even do that. That's why there are things like context7, which in a narrow context helps but not perfect.

There are lots of ambiguous situations where a search and human "inference" can solve that AI still can't.

I can tell the AI to do something, it uses the worst approach, I tell it a better way exists, it says it validated it does not, I link to a GitHub issue saying it can be done via workarounds and it still fails. It's worse for longer tasks where it always shortcuts to if it fails pick a "safe" approach (including not doing it).

Funny enough we need the junior to guide the AI.