Email or username:

Password:

Forgot your password?
sam henri gold

I've been thinking a lot about alt text recently. People don't write alt text as often as they should. And when they do, most people are awful at it.

I think this is one of the single best applications for some multi-modal LLM to step in and write alt text for you.

10 comments
sam henri gold

I think there was a part in Google IO that demoed something like this. "Coming later". But the state of affairs *today* is tragic.

Open the instagram app, turn on a screen reader, and listen to the dog shit generated alt text. It just lists out the number of people and objects in the photo. "May be a meme of 2 people. Bath robe, laundry basket, and text."

Ed

@samhenrigold i kinda feel like education is a big thing here… people are awful at it because very few people know what good alt text reads like. feels like an opportunity to guide / teach people β€” in context β€” vs glossing over it and making it happen in the background without people thinking about it…

(and then fall back to machine generated / community suggested alt text if the author intentionally skips it)

sam henri gold

@edwellbrook Agreed. For an upcoming Mastodon app update, I've added a dialog that sheds some additional info and tips for alt text. Nothing groundbreaking but a worthwhile first step imo

Kevin T

@samhenrigold Yeah, it def feels like a task that could really get the benefits of this kind of tech. At least for providing some kind of basic template framework for a user to start from.

I had to take a class on website accessibility compliance, and part of it focused on writing good image descriptor text; there are a lot of things that people don't even consider which are actually pretty important, and also a ton of stuff that people think is important to include in an image descriptor which is actually really unnecessary/unhelpful. At the very least, a tool providing some kind of heuristic GOOD/OK/BAD could be super useful and shouldn't be toooooo crazy hard to train. :thaenkin:​

@samhenrigold Yeah, it def feels like a task that could really get the benefits of this kind of tech. At least for providing some kind of basic template framework for a user to start from.

I had to take a class on website accessibility compliance, and part of it focused on writing good image descriptor text; there are a lot of things that people don't even consider which are actually pretty important, and also a ton of stuff that people think is important to include in an image descriptor which is...

sam henri gold

@ovmoro And, ya know, give that model a little nudge in the right direction.

```
if (altText.startsWith("An image of")) {
return Score.Awful
}
```

Kevin T

@samhenrigold lol, there are def some worse ways to start a descriptor. But yeah, at the very least "a photo of", "a drawing of", and "a screenshot of" need to rank higher than just "it's a picture" XP

Thomas

@samhenrigold my suggestion would be that you could propose alt texts for others as a Mastodon feature. Either for images with no text or bad text. As an author, you could accept the proposal and it would edit the post.

Go Up