It's exactly the same, with the same problems.
Human can only create something by compiling previous knowledge and experiences, no creation comes from a thin air.
And humans, and it was proven numerous times, are bound by the same mistake as full glass - if all their knowledge is that glass always half full, they will stick to it even if asked about full glass. Even if explained what full glass and even if pushed hardly for it, even ignoring reasonable explanation (like half of a forum believes that gaas is oversaturated to the point of ignoring real data). And maybe one in a thousand may try something out of the box and most these results will be wicked creation on par what AI do. Because basically our mind is a neural network that work on pretty much the same basis as artifical ones, with the same problems underneath (like it's super hard for extremely low-weighted path to activate even if it's a correct path - NN wil stick to "what it knows and what it believes in")
Take a look at Balatro - no matter how original it is, it's really nothing new, it's blending (and AI for sure has no problem with blending) of two well-known formulas - poker and rogue-like. If you look at how progress are done - it's just repeated checking of probable combinations, and when good combinations run out - just check out every possible combination, maybe something will stick and move progress.
The main problems that AI had is that it hardly pass down knowledge, the pivotal in human progress, and there are no self-critic part of the process so someone else should do the filtering and find a gem (Balatro) in a vast sea of failed attempts (sea of trash 1-man indies filled Steam to the brim). Both points now in process of incorporation to AI, DeepSeek for example has both distillation ("stealing" other AI experience, reducing work) and "internal discussion" for better results.