Good point, but at least for now, video deepfakes can still be detected via machine learning.

Also, deepfakes work well for replacing faces, but not yet for synthesizing actual videos of complex scenes (like crowds) :

https://www.youtube.com/watch?v=IMZkLVBhcig

But then you're right, it won't take long before that's possible. Whether we can continue sniffing out deepfakes with ML is unclear.