What it can't do

A network only knows what it was trained on.

Show it something far outside that territory and it will still produce an answer — confidently, fluently, and possibly completely wrong. There's no built-in "I don't know." The network always outputs its best guess, because that's the only thing it knows how to do. It has no way to notice when it's out of its depth.

This isn't a flaw that can be patched. More training data helps, but it doesn't give the model a sense of its own limits. Self-awareness isn't in the architecture. The model can't step back and ask whether a question is within its expertise. It just pattern-matches on what it's seen and produces an answer.

A doctor who has only ever seen patients in one city might be confidently wrong about a disease that's common elsewhere. The knowledge is real, but it has a boundary. The doctor doesn't know where that boundary is.

That's the model. Knowing when to trust it requires knowing what it was trained on.