What you describe might be more robust in some situations but vastly less for software designed to be accessible. Take "alt" attributes in HTML img tags for ins

What you describe might be more robust in some situations but vastly less for software designed to be accessible. Take "alt" attributes in HTML img tags for instance, or various metadata attached to buttons that use an icon instead of text (like a play button, or a X close button and similar things). You can see an example next to this very comment actually: the up and downvote buttons won't be accessible with OCR, but they have "title" attributes describing what they are. And consider that there's more to understanding a given user interface that raw text: radio buttons are tied to certain labels, there's hierarchy, all sorts of layout cues that would be opaque to a screen reader. I suppose that ideally you'd want both: use native accessibility data if available, fallback to OCR when there's no alternative.