@hannah
(Possibly you already thought of this.)

- this has user-submitted content being spread to an amplified audience. It's going to need active moderation (that scales with its use) and can't be just a passive programmatically-automated tool.

- alt-text is contextual, to its use and to its audience. It will probably need to have multiple options present for selection.

... but maybe not if social media use in particular is the goal, like if you're doing a webpage/news story you're putting in enough effort that alt-text-ing it yourself isn't significantly more?
... but maybe still, people use things differently in social media, there isn't always one single use of an image.
... but maybe there's a notably Most Common use.
(... but should the fact that an image can get used in multiple languages be considered?)

@hannah Like, you can ban slurs/censored versions of slurs ("k*ke", e.g.) algorithmically, sure -- which might limit the category of images you can provide the service for, like some images do legitimate contain slurs and are shared and described in detail for e.g. documenting hate-speech, but it's reasonable to just say "that's out of scope because it's too sensitive, people will need to alt-text this themselves" --

but more to the point, you can neither ban (e.g., these are examples and not trying to be a full or equal list) pictures of black people or pictures of Jews, nor ban the words "gorilla" or "vermain".

There will need to be at some level manual moderation against that.