Ask a local model a question, and it can answer with total confidence while being completely wrong. It isn't lying to you, it just can't feel the difference between what is true and what merely sounds true.
Under the hood, it isn't looking anything up in a database. It reads your words and predicts the next likely word, then the next, one plausible step at a time.
So when the real answer is rare, or just missing from what it learned, it doesn't stop and admit that. It smoothly fills the gap with whatever fits the pattern, and that guess can be pure fiction.
Here is the sneaky part. It sounds exactly as sure when it is guessing as when it is right, like a shimmering mirage that looks solid until you step closer and it dissolves into nothing.
You can shrink this by handing the model real sources to read, and by checking anything that actually matters. Fluent never means correct. That's why AI makes things up. Off the Cloud.