I don't know that it makes sense to isolate just the narrow concept of deep fake creation. It seems like fundamentally a lot (most?) of the breakthroughs that make creation of deep fakes possible are the same ideas that make possible the current state of the art for classification, decision problems, advanced NLP, and other things we call ML or AI.

So to do this pro/con analysis you probably need to include the pros of these related technologies as well, which are certainly more than trivial amusement.