Because you seem to be giving measured responses, even though I and other vehemently disagree with you, here's where I think you're missing things:
computers, the internet, and calculators make many tasks significantly easier. That's the point of technology. However, if you don't understand the how and why of math, you'll have no idea how to use a calculator. If you just take whatever the first result of an internet search tells you at face value, you're demonstrating a complete lack of critical thought. You have to be able to analyse sources, or you'll just be mired in mis- and disinformation. If the only thing you know how to do on a smartphone is search the internet or download and use an app, then not only are you inefficiently using the technology, but you are also making yourself incredibly vulnerable to security and privacy threats.
The same is true for LLMs, but doubly so, because everything they output is in natural language (or code output, in the case of coding). If you are an expert coder, then it is possible to use LLMs to significantly increase your efficiency. If not, though, you completely offload the entire problem to the LLM, and can't even check if the solution is valid because, like the internet, you have to have the knowledge necessary to question the source. If you make the assumption that it's responding to you as a person would, then you're screwed, because you're incorrectly ascribing knowledge to a stochastic parrot. LLMs have all the worst cognitive pitfalls of the smartphone, the calculator, and the internet, but only actually have the strengths of the internet, because they're only as good as the training data. Every single thing an LLM outputs must be treated with scepticism because the issues aren't in every output, but could be in any output.
In a world filled with LLMs being used as expert systems, the low-skill users will, inevitably, fail to actually gain skill from using the tool, because the flaw in the tool is inherent (it isn't actually an expert), and you require preexisting knowledge to be able to check its output. In exactly the same way that an LLM cannot reliably check its own output for errors, low-skill users are similarly unable to do so, and will simply offload their own cognition. This leads to a decreasing number of experts who are actually able to use the system efficiently, and an ever-growing layer of untermenschen who must simply trust the mechanical parrot like some oracle, with about the same level of accuracy as your average cold-reading psychic.
This is not an issue of those with knowledge trying to privilege their own skills. It is that experts are warning that this abomination, which the public wants to use like an expert system, is not and can never be an "expert". Non-experts simply think it's entrenched power trying to protect itself, but it's not. The data agrees with my points above. I can show you several particularly damning studies which demonstrate my point.