Raw LLM Responses

A tip about this poisonings: be as subtle as possible bc there's a non zero chance that after scraping the image pass through an human to assess if the image is good or garbage. I have worked on a project to snapshot the homepage of hundreds of thousand of sites and segment the parts that blind people might need the most to be read by a screen reader and catalogue them by functionality. Basically I'd take an image of a page, scrap it immediately if it was mostly images or non textual and then segment it by identifying menus, infos, sidebar etc, ignoring elements that a screen reader would identify semantically like buttons (supposedly at least, it's another whole can of worm there). Point is we can try to train an AI to do a task but it usually starts with human supervision. Idk how big company do it but there's a decent chance that a crawler collects and downloads big batches of images and then they are given to a severely underpaid worker acting as supervisor (there are services that are basically using third world people as human automatons for piss wages, Amazon is big on that market so you know they are evil). The job of the supervisor should be to immediately scrap any image that is deemed garbage, like bad sketches, images not corresponding to the keywords or other AI generated images. Basically if a company uses human supervisors you need to also fool them, it shouldn't be THAT hard bc they have to process hundreds upon hundreds of images per day and when you do that your brain kinda becomes mush after a while. If the company doesn't use human supervisors no big deal, a bit of poison goes a long way since they will also inevitably run into inbreeding problem since the model will eat up A LOT of AI generated images, garbage in garbage out, as it were. Oh, and obviously it will be years before we see truly see solid AI powered assistive technology for accessibility bc it's not that profitable so there's no funding. Also also, having supervised an immense sample of home pages I can tell you it isn't true that most internet sites are porn. It's more like 20%. Single page sites for eateries handily beat them.