Hacker News Reader

Comment by ajithshan001 | Hacker News Reader

The era of AI shouldn’t require users to enter alt text. AI should be able to provide one.

AI has no idea what you're trying to convey through a picture. Imagine you're shitposting a "this is fine" meme with some bad collage on it, asking for a LLM to properly convey your take is just a fool's errand.

Moreover some images SHOULD NOT have alt text or at least shouldn't have their associated alt text displayed in all contexts. I put this in all caps because this is a pervasive and common myth. Granted, all images on Bluesky probably should have alt text because ostensibly an image as part of a post is probably content. However, in cases where an image is purely decorative and not meaningfully relevant to the page, a blind or low-vision person doesn't need to, nor do they want to, hear your weird interpretation of some abstract art. If you disagree, take it up with W3C. This isn't just my opinion.

https://www.w3.org/WAI/tutorials/images/decision-tree/

(creator here) I thought it would at least be easy to transcribe screenshots of just text, which were a common part of the dataset. These are harder to misinterpret so I figured automated alt text would be a win.

But I found that even that was not easy to do with "traditional" OCR, notes here: https://digitalseams.com/blog/image-transcription-humbled-me

> asking for a LLM to properly convey

Alt text is not there to explain the joke, that would imply cripples are stupid.

AI is fine, if you want to go against the trillions $ come back with a proof.

LLMs fail, wow easy to find, I'll bet alt text on random internet images fails at over 10 times the rate, most are null.

I've seen what happens when people use AI to fill in alt text and it's not pretty.

I agree. But ~99% of Bluesky users deeply, deeply, deeply hate AI and anything associated with it, so don't expect any movement on that front for at least a few decades.

What makes for good alt text depends on the author's intent by including the image, an LLM can't know that.

Authors can often use an LLM to save time by outputting an initial draft an author can fix.

Readers can also use their own LLM as a supplement, to ask it questions about details in an image that wouldn't be in the alt text. LLMs will get things wrong but are generally good enough for such supplementary use.

The AI automod can't even tag images correctly. I've seen perfectly innocent drawings get marked as adult or graphic. If it can't figure those out, there's no way it understands enough about what's going on in the images to automatically provide alt-text. At least not well enough that anyone should rely on it when they can just enter proper alt-text themselves.

It feels like little needs to change. If you want to provide alt text to offer high quality specifics, awesome. If you leave it blank, those who need/want alt text should cause it to be generated on-demand. Either client side or by a server which then caches it for the next user.

Same with closed captioning!

And then the alt text Nazis can be quiet, too!

Reply on news.ycombinator.com