Mastodawn

Even with giving “AI” the benefit of the doubt and say 80% of their alternative texts are okay to use, that means the 20% that are not good enough take more time to correct (because you first have to analyze that the text is not good enough, and then you need to write a new one. And for the rest the “analyze/determine/ok” loop is probably about the same length as the “analyze/write accurate alternative text” loop.

https://yatil.social/@yatil/116234999890719208 (1/4)

Eric Eggert (@[email protected])

"AI is the only way to make accessibility at scale work (and please don't show me different examples; of course, with a human in the loop, it's all perfect)." 💤💤💤

yatil.social

Show thread

Jens Oliver Meiert Mar 16

@yatil, just out of curiosity, are you aware of data on what percentage of human-written alt text is sufficient?

Show thread

Jens Grochtdreis Mar 16

@j9t @yatil

How would we know which is which? And: on larger sites the alt-text may come from another company that provides the licensed images.

Show thread

Eric Eggert Mar 16

@jensgro @j9t If it is generated with the content by trained individuals: 100%. If it is dropped in from somewhere else, but human written, I would think 95+%. Text alternatives are trivial in probably 98+% of the cases, and the rest are complex or special situations where you need specific domain knowledge (or good instructions, which are comparitively easy to provide). So I think if an effort is made at all, it can be basically flawless.

Show thread

Jens Oliver Meiert

@yatil, I love the trained-individual sentiment, but we know alt text isn’t and cannot always be provided by trained individuals.

And from what we know from “pre-AI” times, human-written alt text doesn’t seem to be consistently good—if it’s at all present.

Hence the question if we have more data.

Show thread

Eric Eggert Mar 16

@j9t I mean by trained individuals a 15–30 minute instruction. Everyone can do it. You don’t need to study for it. This is not a problem. Yes, ableism is carried out by humans as well as embedded in the training of models. It’s not a technical problem.

I have not done any studies on this and am unaware of any.