gotosocial - Link to source

"ai"-generated image captions are fucking shit

masto.ai/@HourlyPornhubbedHeat…

this one hallucinates a cat, says the dog is standing, mistakes the pink blobs for balloons and hallucinates a yellow balloon, makes up trees that aren't there, and doesn't mention the main focus of the comic, which is the two animals flying through the air thanks to the bubble gum bubbles they're making!

but yeah no cool, really really great technology for disabled people: just lie to blind folks about what's in a picture! what could go wrong!

garbage garbage garbage, if you're pushing this shit technology instead of advocating for people to take a minute to write a decent alt text then fuck you

:smh:

Hourly Pornhubbed Heathcliff 🔞 (@HourlyPornhubbedHeathcliff@masto.ai)

Attached: 1 image #Heathcliff #NSFW #bot #lewd
Mastodon
This entry was edited (2 months ago)

Peter Vágner reshared this.

in reply to tobi is writing bugs

gotosocial - Link to source

modulux

It's unreliable, but it's better than the alternative which in most cases is nothing. Descriptions of unusual images are iffy, but for example finding out what a screen says, which lights are on on a router, what sort of liquid is in a jar... Lots of blind people are benefiting from this routinely. And there are no good alternatives. There's no reasonable world where we can count on a sighted person being always there to describe anything.
in reply to Robert Kingett

gotosocial - Link to source

modulux

Oh, that's definitely viable for images posted on the fedi. But image description of things around us for example is a lot harder. If I need to know what's around me, or what things are on a shelf or such, there's not much way around it. Projects with human volunteers are fine, but, for me at least, I'm pretty inhibited about letting some stranger see my living environment and such.
in reply to modulux

hometown - Link to source

bri

as someone who ran a heathcliff edit account on twitter for years, hand-wrote alt text for four comics posts every day, and convinced several of the other folks in that community to do the same, the alternative is this. it is doing the work, writing the text.

and while i might agree with you that it’s better than nothing on a silly social media post, it’s real bad in my professional life where it just lies about financial numbers in a chart &c.

in reply to bri

gotosocial - Link to source

modulux

Right, where alt text can be provided, a human should provide it. But the world doesn't come with alt text, and can't. There are lots of things that can be done in tech to mitigate it, for example a device might beep in order to give feedback instead of just having lights. But there's tons of visual information in the world that's not really subject to this sort of solution.
in reply to modulux

hometown - Link to source

bri

i get it, i just really worry about the ‘better than nothing’ argument when like… even the op we were talking about… sure it’s low-stakes, but that alt text has zero bearing on the actual image. is having a fundamentally different experience really what we want to be pushing? and again, it trickles down and is getting harder and harder to push back on stuff in my professional life where it’s like… actually important. it worries me.
⇧