The backlog is the opportunity
Thousands of images without alt text, videos without captions, and documents written in dense prose. This work has always been deferred because it was manual, and it is now tractable — which makes it one of the more genuinely useful applications available.
The stakes are higher than they look
Someone using a screen reader cannot see that the alt text is wrong. A fabricated description is not a minor error for them; it is the only information they get, which puts it squarely among confident failures that are invisible to the person affected.
Wrong alt text is worse than none. The reader has no way to know it’s wrong.
Alt text needs context the image doesn’t have
Good alt text depends on why the image is on the page, not just what is in it. Supply the surrounding content, and treat generated text as a draft — decorative images should be marked as such rather than described.
Captions need review for names and terms
Transcription handles ordinary speech well and mangles product names, jargon and unusual spellings. A glossary of your terms plus a human pass over the specifics is the workable process — the same draft-and-edit shape as everywhere else.
Plain-language versions are not summaries
Simplifying must preserve meaning, especially where the text carries obligations or instructions. Anything with legal or safety content needs the same review as the original.
Do the structural work too
Alt text does not fix contrast, keyboard traps or unlabelled fields. Those are design decisions, and they remain the ones that decide whether people can complete your form.