It’s not just about reading the screen, it’s about using it. It’s quite the downgrade going through a slow API hoping to be able to navigate compared to the 10x speed (or more) of screen readers. Then you have the problem of LLMs being stateless and sensitive information etc etc.