The reason is that alt text is scraped in order to train multimodal text + image models.
But for anyone who actually listens to the blind community, it doesn't take long to learn that these models are pretty important to a lot of people for accessibility. And regardless, alt text is directly helpful. I'm happy to put my displeasure with big tech on hold for things that genuinely improve people's lives.
Maybe, IDK. It depends on what they're training the model for, and how they're using it. But do we really want to make it unusable for training the model, when those models are then used for accessibility as well?
@dilmandila
I think alt text should focus on what you want to communicate with the image, not some "objective" description of what's in it.
I like it when people do that even though I'm seeing, sometimes it helps me when I find it cryptic why people posted the image.
@hosford42 @ChristineMalec @ClimateJenny
@dilmandila @hosford42 @ChristineMalec @ClimateJenny
Another approach would be to write clear, descriptive alt text but use something like Nightshade to poison your image so that "AI" can't make sense of it:
I still think that's working against the public good. If the models being trained on this data are used for accessibility by the blind community, and we are intentionally working to lower their accuracy, we are hurting the blind community. We should fight the AI vendors on a front where there aren't innocent bystanders who will be hit.
@hosford42 @CppGuy @dilmandila @ChristineMalec @ClimateJenny
You've hit the nail on the head here.
When we fight big tech, which we should, we shouldn't be fighting to destroy accessibility. like you said, it hurts innocent bystanders.
People often don't realize what IS accessibility, and so I wrote a zine to break it down and try to teach folks, so they don't do harmful stuff like poisoning datasets used specifically for improving accessibility. https://reshapingreality.org/2025/06/17/accessibility-zine-completed-and-ready-for-download/
Sure, that's poisoning the image itself, not for the purpose of ruining models that could generate alt text, but ruining models that could generate othe images of the same style. I say go for it.
@hosford42 Is there a way to do one but not the other, though? Do alt text-generating models and generative diffusion models use different datasets or get their data differently? 🤔 Ideally there would be NO non-consensual scraping of any websites, and models that purport to have pro-social uses shouldn't be resorting to it.
I think the core problem here is that the laws need to catch up to the reality -- and be enforced, too. Someday we will live in a world where we each have irrevocable ownership of our own data, and can choose to donate it all or in part to those causes we deem just. I have no idea how long it will take us to get there, though. Probably not our lifetimes, from the look of it.
This is how far you'd have to go to actually stop it from happening. Or at least you can't post *publicly* without it happening.
This is not to say that it's right for that to be the case, or that it's the individual's responsibility to curb bad behavior from these corporations. We need laws for that. But if people are going to take it upon themselves to disrupt this behavior, picking on alt text in particular is severely misguided.