It's not adding roles/alt tags/etc. that's the issue. It's finding out that a screen reader doesn't support the ARIA markup (from the ARIA spec) that you are trying to use [1]. -- For example, using aria-describedby on an img doesn't work on most browser + screen reader combinations [2].
If you serve your UI as out-of-order HTML content abusing tables for browser layout, if you merge broken fragments of mismatched XHTML from a content management system to cobble a page together on the fly, if you hide items behind burger menus, if you make interactivity from CSS and minified Javascript to visually look like input boxes and buttons instead of using standard buttons, if you add overlays and blocks to stop people selecting or copying text, if you care nothing about tab order and keyboard shortcuts and image alt text, if you reflow text around images, if you name CSS identifiers with no semantic meaning, if you mix in arbitrary advertising and tracking and framework session state, why is a screenreader "a piece of crap" for not being able to magically infer a human interpretation of the page?
Why is it embarrassing to need to code against standardised accessibility interfaces so that other tools using those interfaces can function? Is it embarrassing to have to code against a database or filesystem interface to persist data, or a graphics interface to show pictures?
The sad reality is that screen readers have a very small market. There isn't a lot of money to be made so not much innovation happens. Maybe with LLMs there will be a new approach. It's something they might actually be good for.
Right, and the insane part is that instead of this small market improving itself the burder is instead shifted to everyone else, even mandated by laws.
