The problem with LLMs is that it is not only the "irrelevant details" that are hallucinated. It is also "very relevant details" which either make the whole system inconsistent or full of security vulnerabilities.
But if it's security critical? You'd better be touching every single line of code and you'd better fully understand what each one does, what could go wrong in the wild, how the approach taken compares to best practices, and how an attacker might go about trying to exploit what you've authored. Anything less is negligence on your part.
Which seems like an apt analogy for software. I see people all the time who build systems and they don't care about the details. The results are always mediocre.
I think this is a major point people do not mention enough during these debates on "AI vs Developers": The business/stakeholder side is completely fine with average and mediocre solutions as long as those solutions are delivered quickly and priced competitively. They will gladly use a vibecoded solution if the solution kinda sorta mostly works. They don't care about security, performance or completeness... such things are to be handled when/if they reach the user/customer in significant numbers. So while we (the devs) are thinking back to all the instances we used gpt/grok/claude/.. and not seeing how the business could possibly arrive to our solutions just with AI and wihout us in the loop... the business doesn't know any of the details nor does it care. When it comes to anything IT related, your typical business doesn't know what it doesn't know, which makes it easy to fire employees/contractors for redundancy first (because we have AI now) and ask questions later (uhh... because we have AI now).