But you would need to collect that volume of footage. Then create a training set. Then pay for training.

Only one of those gets cheaper and faster with time. (And that's just me guessing it will get cheaper. I could be wrong.)

In any case, the other two scale up with the complexity of the game being streamed. The AAA studios can simply overwhelm with complexity and content. Especially since AI will be generating an increasing majority of the game anyway as time goes on. If it's playable for a year's worth of hours, good luck downloading enough video to replicate it.

Maybe Twitch or youtube could if they incentivised enough streamers to stream their play sessions? Of course, why would Twitch or Youtube give that dataset to you?

It's just not as easy as people are making it out to be.