Gregory's Server@RoyReed @ChanceyFleet FYI, I work for the Flickr Foundation and we've been looking into AI for writing "added text descriptions" for images in Flickr Commons that may not have any description at all. It's surprising how good some of the descriptions are. We still haven't decided what to actually do with it (Flickr doesn't have an actual alt text capability) but I went in skeptical and now have an "It's a lot better than nothing" feeling.
|
6 comments
 @ChanceyFleet @RoyReed I, too, am concerned about this iterative effect (i.e. models getting trained on themselves) so we're definitely in favor (I am anyway) of making sure there are human checks. I mean I don't get to build any of this, but I have some oversight into how things might be built. I think our concern is: given that we have some institutions which have no image descriptions, what is the best way to get more description in there? I do take all of your points, they are good ones. @RoyReed @jessamyn It’s a conundrum. Most institutions and platforms have so much legacy content that needs alt text, and accessibility teams aren’t resourced enough to do it in house. I’ve heard of crowd-volunteering efforts, which seems like a good thing, and there are also companies like Scribely that will write alt text as a B2B service. At a minimum, a human in the loop is necessary to make sure misleading or plain wrong description doesn’t get published. @ChanceyFleet @jessamyn Yes, I agree about the sentimentality of some of the suggestions. I do edit what AI proposes, but I do find it a useful starting point. @jessamyn @RoyReed I use AI description a lot most recently to learn that a discarded mattress on the sidewalk was what it was, without having to touch it. I think that generally it’s best if descriptive AI is used at an individual’s discretion, not at scale. This way, individual users can choose a model, refine their prompts and have their preferences reflected in custom description, and meanwhile nobody is invited to outsource institutional descriptions to a bot. |
@RoyReed @jessamyn Simultaneously far better than nothing, and a threat to human-created alt text. AI descriptions tend to use very normcore, Halmark-esque language. Every photo with a puppy is heartwarming, every photo with a couch is cozy. I fear that institutions will adopt AI descriptions as a perceived time-saver, and the gains we’ve made in human alt text literacy will be wiped out.