Also virtual environment creation, agent training, etc. For any given game you can create a small dataset of recordings of both player inputs and gameplay footage, use that to create a model that can derive inputs from looking at footage, and then create input sequences for your huge backlog of gameplay footage. From there you can use the backlog to train AI that either recreates realistic player actions from screen inputs, or AI that recreates the entire game (like the AI minecraft)

Not to mention the huge amount of voice samples and webcam footage you could use for more typical voice cloning, text to speech, human avatar creation, etc