10 years? Imagine it 2 years from now. Or even tomorrow, because it's unlikely this is the apex of what's possible now.

I imagine in 10 years it'll be the same. Firstly because if you can't get "enough" good content to model, you will have to manually shape the models to make it sound authentic - and that isn't AI, that's editing. But secondly because humans are very good at recognizing authentically human interactions, and algorithms simply aren't good enough to generate all the nuances we identify as authentic.

If you want to fake someone out, use a regular person, and clean it up in post.