Maybe consumer AI image/video generation sucks so that the real deepfakes are never suspected.
Discussion
I have little doubt this is true to some extent at least currently and probably the gap was more pronounced years back.
Considering how it feels like the text generation from llms is worse than it started, this seems plausible.
I think they were hiding it as long as they could, and the recent seeming acceleration was largely inability to keep car in the bag. Partially anyway