The Corridor Digital podcast was just discussing replacing core actors with AI recently. At least one of them said "never" even though they mentioned there is already heavy use of various forms of "digital doubles" for background actors.

It looks like we are at the stage where text-to-performance of a fully text-to-actor actor is very feasible and it will shortly be a matter of refining it to be more lifelike.

Maybe the most speculative part of that is creating realistic facial expressions and movements that are truly appropriate to a script. But I think we have things like facial pose extraction, videos, and video transcripts in addition to actual scripts. So there is more than enough data to make a robust model.