I believe accessibility is the part that makes LLMs helpful, when they are given an easy enough task to verify. Being able to ask a thing that resembles a human what you need instead of reading through possibly a textbook worth of documentation to figure out what is available and making it fit what you need is fairly powerful.
If it were actually capable of reasoning, I'd compare it to asking a linguist the origin of a word vs looking it up in a dictionary. I don't think anyone disagrees that the dictionary would be more likely to be fully accurate, and also I personally would just prefer to ask the person who seemingly knows and, if I have reason to doubt, then go back and double-check.
Here's the manpage for bash's statistics from wordcounter.net:
I do try to keep the "unknown unknowns" problem in mind when I use it, and I've been using it far less as I latched on to how OOP actually works and built up the lexicon and my own preferences. I try to only ask it for high-level stuff that I can then use to search the wider (hopefully more human) internet more traditionally with. I fully appreciate that it's nothing more than a very incredibly fancy auto-completion engine and the basic task of auto-complete just so happens to appear intelligent as it gets better and more complex but continues to lack any form of real logical thoughts.