Raw LLM Responses

hi there! i've worked with AI algorithms (though not genAI) in a professional setting and i can kinda give you a rundown! poisoning works by disrupting the pixel information in the image in a way that is *usually* imperceptible to the human eye. it also depends on the monitor you're using. i'm currently using a retina display and i can tell the difference between the two images shown at 19:15 for example. the pre-poisoned one is more saturated compared to the poisoned version. but what this actually does to the algorithm (meaning the AI scraping the data, like midjourney) is confuse it as to what it's actually looking at. sorry if this explanation is long, but here's the most simple example i could come up with: let's say you drew a picture of a potato. just a regular potato that's brown and all that. the algorithm that scrapes your potato pic will be able to see that it's an oblong object, earthy colors, got some dimples for the potato eyes (the sprouts, in case anyone didn't know what potato eyes were lol), etc. that's a potato, it's labeled as a potato in the database (images are tagged with text in a lot of cases so the AI can "read" what the image is supposed to be associated with in plain english before examining the pixels), so the AI learns that "an oblong, brown object with dimples/sprouts" is a potato. now let's say that you took that potato pic and applied something like nightshade to it. nightshade is, itself, an "AI" algorithm, but it works in a different way. it examines the picture and, using mathematical equations, basically applies subtle, slight differences in the pixel data that are more likely to be overlooked by the average person, especially if the two versions aren't shown side-by-side. you likely won't be able to see the differences between the original image and the poisoned one, but an AI scraping images *can* now, here's the fun part. let's say this AI scrapes your poisoned potato pic instead of the original one. nightshade may obfuscate the pixel data in such a way that the AI may "see" different colors, outlines, shapes, etc., than what we can process with our eyes and brains. our brains are exceptionally good at "filling in blanks" when it comes to image recognition, but an AI algorithm is as literal as it gets. it doesn't have a brain, it doesn't have eyes, so it "sees" the most "literal" form of an image and its data. so instead of seeing "an oblong, brown object with dimples/sprouts," the AI sees "a boxy/triangular, reddish object with holes." and since that image was tagged as "potato" in the database, the AI "associates" the word "potato" with "boxy/triangular, reddish object with holes." the idea is that after ingesting enough poisoned data, the AI will produce images that don't match the requested prompt. let's say a bunch of people draw potatoes and poison them with nightshade. after a while, if an AI-bro prompts the algorithm with "picture of a potato," the AI will predominantly produce images of boxy or triangular shapes that are red with holes in them. it disrupts the algorithm and its ability to "learn," thus hopefully discouraging these companies from scraping images without permission. that's the hope/goal, anyway. NOTE: keep in mind that with my example, i HEAVILY simplified the process since it's a lot and requires what's called iteration. one singular poisoned image is not enough to disrupt the algorithm's learned data set. it requires hundreds upon hundreds (if not thousands) of hours of training on "bad" data to disrupt the output on one of these algorithms. that's also why AI-bros like to say it doesn't work, but that shouldn't discourage anyone from doing it. i hope that helps and i'm sorry for such a huge wall of text! 😭 i'm happy to answer questions if you have them, too. :)