My experience is that they uncomfortably do both. You can "understand" something conceptually quicker -- like you have a new brain-muscle-thing that lets you cut through the hard difficult tedious corners to get to the meat of the matter.
But then you also can become reliant on it, and have difficulty doing the mechanistic rote work of working through it yourself.
Like the really big powerful calculator that it is, really.
You can use AI or the internet to learn the basics of how a gas engine works in a couple of minutes. But you'd be incapable of actually working on a gas engine or designing one.
Surface level knowledge gets you surface level functionality. You don't become good at something from surface level knowledge, but you might think you're good at it.
I've used my phone taking pictures + Codex + a PDF of my tractor manual to help me effectively diagnose and manage repairs in my tractor. (Though these models remain terrible at the physical world, getting physical orientations wrong, front back etc. Much like myself)
Likewise I had Gemini help me tear down my mower's carburetor and diagnose issues there.
(So much so that I've wondered about building some kind of "shop buddy" -- some kind of durable laptop and set of cameras ... on a cart. Running models that have access to manuals and cameras and TTS and voice input? "Hey, shop buddy, look at this fuse and tell me what is before and after it in the electrical system.")
This is helping me learn and do something I couldn't really effectively do before by walking me through steps.
My youngest has had Gemini write math questions for them, to help study. Not do the math, but write questions.
In the end it comes down to prompting, like everything.
Which makes me wonder if the answer for higher education is just to provide the students with specific coding agents they're specifically allowed to use -- ones that would push the student through problem solving and working on the problem together.