Post Ax1a7vtBtfLfdI9K88 by vivi@arff.archandle.net
 (DIR) More posts by vivi@arff.archandle.net
 (DIR) Post #Ax1HtODUUmxTyCYnHk by kevinrns@mstdn.social
       2025-08-10T00:51:56Z
       
       1 likes, 0 repeats
       
       @dotjayne Alt text requires poison for meta scraping. Remember to poison alt text until Mastodon adds poisoning to images automatically.Meta is a crime happening to Mastodon users on images uploaded and meta scraping of alt text.Poison images, poison alt text.Lie in clever ways.
       
 (DIR) Post #Ax1a7vtBtfLfdI9K88 by vivi@arff.archandle.net
       2025-08-10T04:19:54Z
       
       0 likes, 0 repeats
       
       @kevinrns @dotjayne how would you propose to poison alt text in a way that is still accessible to those that rely on it?
       
 (DIR) Post #Ax1a7xE8vFhbmYxZQW by kevinrns@mstdn.social
       2025-08-10T06:56:03Z
       
       0 likes, 0 repeats
       
       @vivi @dotjayne Its getting better all the time.https://www.technologyreview.com/2023/10/23/1082189/data-poisoning-artists-fight-generative-ai/alsohttps://www.pcmag.com/news/heres-how-to-poison-the-large-language-models-that-power-ai-apps
       
 (DIR) Post #Ax1a7y6Nfb5UUmKsy0 by vivi@arff.archandle.net
       2025-08-10T10:34:14Z
       
       1 likes, 0 repeats
       
       @kevinrns @dotjayne that's wonderful, thank you for these very nice articles,,, so let's get back to actually answering my question instead of diverting, how are you going to apply this to alt text which is a tiny stream of text characters which have to be chosen in such a way to accurately describe the picture and not just nonsense, in a way that will not impact accessibility?Image poisoning is a little different than a text stream since the human eye can handle a bit of lossy noise and still see essentially the same image, text streams are leagues different. Most image alt text is 100-200 characters, if you want to sufficiently poison a tiny bit of text on such small scale, for individual images, would absolutely require giving up any hope of writing an accurate written description of the image, unlike poisoning hundreds~thousands of alt texts over an entire sampleset. And what's your idea on how to avoid impacting screen readers which rely on local TTS models to read the alt text? These would just as likely be affected as well from poison streams and would incorrectly read aloud what the actual text says.Or do you see this fancy poison tech and think it works like a magic panacea solve-it-all tool and no issues would arise and it Just Works, just like how techbros think LLMs are a magic panacea solve-it-all tool and that no issues would arise and it Just Works? It doesn't, you're just gonna fuck up someone's accessibility on the web and not solve anything.I'll give you a hint, or maybe even answer the question for you. You don't need to fuck with alt text and pay the accessibility toll to achieve the same effect. Just poison the image instead and give genuine alt text. The LLM would then not be (as) able to associate the alt text to what is represented in the poisoned image. And one or two descriptive sentences describing a poisoned image that the AI doesn't even know how to correctly interpret is a nothingburger for training data.Advocating for making accessibility worse, even by just a little, by poisoning the data of the only asset some people will have to rely on navigating the web, is the wrong approach. Throwing your allies under the bus when confronted with fascism is truly a speedrun strategy to ensuring you have no allies left when later confronted by fascism.
       
 (DIR) Post #AxQCUHz2tGqOa20tpg by Stomata@social.linux.pizza
       2025-08-22T07:59:03Z
       
       0 likes, 0 repeats
       
       @dotjayne alt text for people like me who is blocking image preview