Yep. I'd say it's an order of magnitude more effort to read code you haven't written too, compared to reading code you wrote. So there is approximately zero chance the people using AI to generate code are reading it at a level where they actually understand it, or else they would lose all of their supposed productivity gains.
Good news: the evidence points to it being slower than non-ai workflows. So we're destroying our economy, society, and planet to make worse software, more slowly! :)
I'm sorry but I'm not going to take "research" about Claude seriously from Anthropic, the company who makes and sells Claude. I'm also not going to do that for Copilot from Microsoft, the company who makes and sells Copilot.
But useful in the context of these debates isn't that it solves any single problem for someone. Nobody is arguing that LLM's have zero utility. So I don't really see what your point is?
I've seen some interesting work going the other way, having LLMs generate constraint solvers (or whatever the term is) in prolog and then feeding input to that. I can't remember the link but could be worthwhile searching for that.