It can only interpolate, not extrapolate. So the 'new' things you're seeing are just rearrangements of the (millions/billions of) things that the DNN was trained on. It has no understanding of what it has 'learned' (or more accurately: lossy memorization a.k.a. compression) and makes all kinds of mistakes (some due to training losses, some due to garbage/conflicting data fed in.) This is probably why the creative applications (i.e. Stable Diffusion etc) seem more impressive than the functional applications (i.e. Galactica) as even 'wrong' output can be creatively interesting.

For example, if a new comic/movie/video game character came out tomorrow that had a very unique physical look, Stable Diffusion would have difficulty even approximating it (i.e. its training data wouldn't have what was needed to reproduce the appearance.) But it can produce a decent Darth Vader because it's probably been fed at least thousands of drawings/photos of this very well known fictional character.