Samuel
surprisingly controversial, just want to say two things 1. 100% depends on execution. If the model is bad then obviously it won’t be helpful 2. It is obviously no substitute for good human-written alt text. HOWEVER, even though this site has a good culture for writing alt text, only a tiny…
Samuel@samuel.fm· 1 Jun 2024
huge!
Mozilla Developer @mozhacks Firefox 130 is bringing a game-changing feature: automatic alt-text generation for images using a fully private on-device Al model! Initially available in the built-in PDF editor, our aim is to extend this to general browsing for screen reader users.
7 4 68
Samuel
@samuel.fm
…fraction of images posted even has alt text, and I’d bet that only a fraction of that have *good* alt text. While this tool wouldn’t raise the ceiling, it certainly would provide a baseline for that overwhelming majority of images that would otherwise go without. Please don’t stop writing alt text!
1 1 37
Samuel
and like, sure, probably going to be a lot of slop, but is this really *worse* than nothing? bare in mind this is on-device, aka your browser looking at an image without alt text and describing it, rather than the person uploading it adding lazy alt text
Conversation with chatgpt where I asked them to provide alt text for the famous “Washington crossing the Delaware River” painting appropriate for a microblogging site ChatGPT: George Washington and his troops crossing the icy Delaware River during the Revolutionary War, depicted in a historic painting. 2 1 26
Samuel
I am definitely anti-automatic alt text generation at upload time without human review. that seems bad, but also that’s not what Firefox is proposing
2 32
neko
i would really love an upload version but with mandatory human review by the uploader... like, pre-fill alt text for me but have me change and approve it. would make things so much simpler.
3
fellowship of the prim
Agreed. And while this isn't an excuse for most people, this is still a better description than some people with intellectual disabilities could write themselves. If executed well, it could be a good accessibility tool in both directions--which is AI at its best.
2