How do people feel about #AltText written by AI?

I often end up thinking it is way too many words about things that don’t matter, and I often end up not boosting it.

@Cassandra
Better than nothing, worse than human alt-text.

@TheGreatLlama I guess I'm wondering if it *is* better than nothing. I honestly would not want to listen to so many pointless words even once.

Edit: I feel like I should add more words. It is important to me that alt text gives people who rely on it an experience equivalent to that of people who don't. It feels disrespectful to me to generate text that is largely a waste of someone's time, especially given what we know about the harmful consequences of generative AI.

@Cassandra
Well, by "better than nothing", I mean by a hair. The difference between not caring enough to bother, and only caring enough to bother momentarily.

I do understand it in certain situations: on scene with a rapidly developing event, for example. I really wish people in that situation would go back and add human text later though.

@Cassandra No, don’t do it. Images invoke emotion and a bucket of transistors can’t explain that, hard as it may try.

@Cassandra I'm terrible at writing alt text that is genuinely useful, so I use Gemini to write it. I also note at the end of the alt text that it was generated by Gemini, so I try to be transparent.

But boosts or not, I'm not worried about that.

@srgower If sometime you want help trying to write useful alt text, let me know.
@Cassandra Sure! Admittedly I haven't looked up many resources to that effect.
@Cassandra If it’s simple enough I write my own. When it feels too complex, I use @altbot but I choose how much text should be on it
@farah Do you mean by setting parameters or by editing the output?
@Cassandra Editing the output
@farah Thanks. I think editing the output helps.
@Cassandra The one use case that doesn’t bother me is when blind people (or others who genuinely need help) add the auto generated text, especially because if they don’t add any alt text then someone will complain in the replies. Otherwise I agree that auto generated text is usually a wall of words that don’t help anyone understand the image.

@Cassandra

It's absolutely wretched for memes that rely on cultural touchstones.

"Cartoon man petting animal" is useless when it's a screen grab of The Emperor's New Groove where Pacha is patting Kuzko's llama head while kuzko is mad at being a llama.

@CorvidCrone Yes! I don't know that particular meme but so many times the description leaves out essential info!
@CorvidCrone @Cassandra
Perfect. This is exactly what memes are like for me! If I do know the original reference, I generally have no idea what its memification has turned it into.
@Cassandra yep, and missing important, obvious "details", like the identity of a celebrity or politician.
@Cassandra not sure i get an opinion as i can see the images? but i'd worry that the text would be wrong
@🔥Cassandra🔥 When I describe a meme-based image macro, it's rather concise. But I've got very good reasons to describe my original images at greater detail in alt-text and then once more at vastly greater detail in the post itself. It depends on what people may not know about the image but want to know.

I hope to limit my future alt-texts to a maximum of 512 characters, though. That's Misskey's limit, but Misskey currently has a bug which makes it delete longer alt-texts instead of truncating them.

That said, AI also hallucinates and guesses too much without actually knowing. This occurs the more, the more obscure the topic of an image is and the less the AI knows about it.

So if you want your image description to be accurate, you'll have to write it entirely yourself.

#Long #LongPost #CWLong #CWLongPost #FediMeta #FediverseMeta #CWFediMeta #CWFediverseMeta #AltText #AltTextMeta #CWAltTextMeta #ImageDescription #ImageDescriptions #ImageDescriptionMeta #CWImageDescriptionMeta
@jupiter_rowland I try to think what about the picture adds to what I've said or what the viewer needs to understand from the image. But that does require figuring out how to put that into words…

@Cassandra there was a post by a blind person a while ago saying it was worse than nothing

They have the ability to run the AI on it themselves if they want to

@ghosttie Ah, don't suppose you remember who?

@Cassandra I think I found the conversation (Mastodon needs a favorites search)

https://mastodon.gamedev.place/@ananta[email protected]/115804707250757691

It was a conversation between a blind person and the person who made altbot, who was taken aback that (some) blind people dislike the bot that he made to help them

Gamedev Mastodon

@ghosttie

Edit: Thank you! I boosted the comment.

I can search my favourites! Is that just on Ivory, maybe?

keyword in:library

@Cassandra oh I hadn't noticed that, I was looking for the word "favorites" in the help text, thanks

@Cassandra
It's important to be as brief as possible and only note the relevant details. Like I recently wrote "wearing a yellow jacket, eating blue cotton candy" and I thought "blind people don't GAF about colors.."

Only reason I noted those details was for recently blinded people who may remember the meme.. I guess..

It's also useful to people on text interfaces, even if they aren't blind. So that they don't have to load it.

@Cassandra I don’t see a fundamental problem when:

- System that has as small environmental impact as possible (it’s probably hard to judge but…)
- The human author reviews and corrects as needed, so it follows the normal guidelines for alt text written by a human.
- The human author cannot easily write the text themselves.

I haven’t actually done this yet as I haven’t cared enough to spend the spoons finding a system I can live with using. The reason I would consider it, is that for me, anything but a very simple photo uses a lot of spoons to write alt text for, so that means there are many photos I just don’t post as I don’t have the spoons.

I find it hard to articulate exactly why I find this kind of AI use different, but core is solving real accessibility problem and sender review for correctness.

@samantha Yeah, I think it's where there isn't review that it's a problem, where the person posting doesn't read the generated alt text to cut extraneous information, correct errors, or add missing details that are essential.