The leading 3d gen techniques are latent diffusion models that work off voxelized intermediate representations of objects. When you convert these into meshes, very often they will have poor topology. Topology is important if you need to modify these meshes or animate them. Bad topology and excess verts is also bad if you're building assets for a game.
There's still more refinement of existing techniques that need to be done before AI generated 3d models match what humans produce.
The most interesting/surprising development for me has been that LLMs are getting better and better at synthesizing procedural content.
Fable 5 can write procedural texture shaders, CSG, surface meshing, l-systems, even rigging + realtime IK, etc. and do it all in a visual loop within an environment. The model understands the topology, and can tune things with mathematical precision where necessary. Which is much closer to using 3D tools like an artist would.
My experience here has been so bad with the topology that I usually use the created mesh as an example and rebuild the whole thing from scratch to reduce the verts. Not even worth the time.