Admitting that they don’t know or they can’t do something. Instead, AI will bullshit its way to an answer - any answer.

I use a local LLM - sparingly - at work to code boilerplate and simple stuff I can review, and give me technical advice I can quickly check the validity of (you know, stuff like "which command line argument to do this-or-that). It’s okay most of the time.

But it never says “I don’t know”: instead, it invents shit. And that truly wastes my time on a regular basis.