No shit. This is anecdotal evidence, but I was recently teaching a university CS class as a guest lecturer (at a somewhat below-average university), and almost all the students were basically copy-pasting task descriptions and error messages into ChatGPT in lieu of actually programming. No one seemed to even read the output, let alone be able to explain it. "Foundational skills" were near zero, as a result.
Anyway, I strongly suspect that this report is based on careful whitewashing and would reveal 75% cheating if examined more closely. But maybe there is a bit of sampling bias at play as well -- maybe the laziest students just never bother with anything but ChatGPT and Google Colab, while students using Claude have a little more motivation to learn something.
People who spent the past two years offloading their entry-level work onto LLMs are now taking 400-level systems programming courses and running face-first into a capability wall. I try my best to help, but there's only so much I can do when basic concepts like structs and pointer manipulation get blank stares.
> "Oh, the foo field in that struct should be signed instead of unsigned."
< "Struct?"
> "Yeah, the type definition of Bar? It's right there."
< "Man, I had ChatGPT write this code."
> "..."