Since LLMs spit out LOTS of text, a person may waste lots of valuable time simply meandering through responses, and trying to keep it all in their head.
Contrast this to talking to a person who might just say "umm.. no no that's not how it works - because blah".
I had a side (still work) project I wanted to complete that was extremely simple. In my AWS account create an EC2 image builder pipeline that set up a specific component. Just one, very uncomplicated and documented component.
I figured I’d just have Codex do it. I realized after about 3 days of various back and forth in between other work things that we were somehow stuck.
I figured this must be more complicated than I had imagined so I take a look. No, apparently Codex had failed, made an incorrect assumption as to why it failed (again, documents were provided and dead simple) and then continued trying to fix the issue based on the incorrect assumption.
That is called "doing your job correctly". If you aren't actually understanding what the LLM puts out and reviewing it, you are going to get burned sooner or later.