this is not a crazy idea honestly. i mean, using bots and classic game AI for testing etc is also there ofc and this seems similar but i think considering the state of released games, often game mechanical issues could be fixed

can you give some idea on if and how your model reason about visual bugs though? this is where i find human testing to be more logical. what if a buton click flickers the button instead of 1 time lightup and darken? can it see all the frames and catch it?

this seems where i think a lot of effort is likely still going to be spent so if in some what it could get a description from the code or a person and then reason correctly about visual behaviors in the game.

another tricky area is mandatory testers for appstores. those you can trick perhaps using your AI into thinking someone is testing it really but i am not sure if it would be against their terms... (likely? maybe could be material for exploring market opportunities)