We need to treat the screen like a PNG which an AI can parse to determine what the text is, what text is relevant to read to the user, where and what the buttons are, how to emphasize reading the text based on the visuals, etc. It's the only thing that will scale and while difficult it seems that creating something like this is within reach.