Nightshade: An offensive tool for artists against AI art generators

>You just pay for the “looks pretty”, frame and paper.

But you pay that for any piece of art though? You appreciate it because you like what it looks like. The utility of it is in how good it looks, it's not how much effort was put into it.

If you need a ditch you're not going to value the ditch more if the worker dug it by hand instead of using an excavator. You value it based on the utility it provides you.

I don't think that's true at all. Images and text get reposted with or without consent, often without attribution. It wouldn't make it right for the AI companies to scrape when the original author doesn't want that but someone else has ignored their wishes and requirements. Basically, what good is putting your stuff behind login or some other restrictive viewing method if someone just saves the image/text? I think it's still a relatively serious problem for people creating things. And without some form of easy access to viewing, the people creating things don't get the visibility and exposure they need to get an audience/clients.

This is one the AI companies should offer the olive branch on IMO, there must be a way to use stenography to transparently embed a "don't process for AI" code into an image or text or music or any other creative work that won't be noticeable by humans, but the AI would see if it tried to process the content for training. I think it would be a very convenient answer and probably not be detrimental to the AI companies, but I also imagine that the AI companies would not be very eager to spend the resources implementing this. I do think they're the best source for such protections for artists though.

Ideally, without a previous written agreement for a dataset from the original creators, the AI companies probably shouldn't be using it for training at all, but I doubt that will happen -- the system I mention above should be _opt-in_, that is, you must tag such content that is free to be AI trained in order for AI to be trained on it, but I have 0 faith that the AI companies would agree to such a self-limitation.

edit: added mention to music and other creative works in second paragraph 1st sentence

edit 2: Added final paragraph as I do think this should be opt-in, but don't believe AI companies would ever accept this, even though they should by all means in my opinion.

you are not incorrect that this would help mitigate, but it still misses a few key points I think regarding why artists are upset about AI generation

- This is still vulnerable to stuff like mturk or even just normal users who did get past the anti-bot things pulling and re-uploading the content elsewhere that is easier for the AI companies to use

- The artists' main contention is that the AI companies shouldn't be allowed to just use whatever they find without confirm they have a license to use the content in this way

- If someone's content _does_ get into an AI model and it's determined somehow (I think there is a case with a news paper and chatGPT over this very issue?), the legal system doesn't really have a good framework for this situation right now -- is it copyright infringement? (arguably not? it's not clear) is it plagiarism? (arguably yes, but plagiarism in US court system is very hard to proof/get action on) is it license violation? (for those who use licenses for their art, probably yes, but it's the same issue as plagiarism -- how to prove it effectively?)

Really what this comes down to is that the AI companies use the premise that they have a right to use someone else's works without consent for the AI training. While your suggestions are technically correct, it puts the impetus on the artists that they must do something different because the AI companies are allowed to train their models as they currently do without recourse for the original artist. Maybe that will be ruled true in the future I don't know, but I can absolutely get why artists are upset about this premise shaping the discussion on AI training, as such a premise negates their rights as an artist and many artists have 0 path for recourse. I'm pretty sure that OpenAI wouldn't think about scraping a Disney movie from a video upload site just because it's open access since Disney likely can fight in a more meaningful way. I would agree with artists who are complaining that they shouldn't need to wait for a big corporation to decide that this behavior is undesirable before real action is taken, but it seems that is going to be what is needed. It might be reality, but it's a very sad reality that people want changed.

In so far as anger goes against AIs being trained on particular intellectual properties.

A made up scenario¹ is that a person who is training an AI, goes to the local library and checks out 600 books on art. The person then lets the AI read all of them. After which they are returned to the library and another 600 books are borrowed

Then we can imagine the AI somehow visiting a lot of museums and galleries.

The AI will now have been trained on the style and looks of a lot of art from different artists

All the material has been obtained in a legal manner.

Is this an acceptable use?

Or can an artist still assert that the AI was trained with their IP without consent?

Clearly this is one of the ways a human would go about learning about styles, techniques etc..

¹ Yes you probably cannot borrow 600 books at a time. How does the AI read the books? I dont know. Simplicity would be that the researcher takes a photo of each page. This would be extremmly slow but for this hypothetical it is acceptable.

I know what you're saying, and for sure existing models can be difficult to force into the really weird corners of the distributions (or go outside the distributions). The text interfaces are partially to blame for this though, you can take the images into Gimp and do some crude naive modifications and bring them back and the model will usually happily complete the "out-of-distribution" ideas. The Stable Diffusion toolboxes have evolved far away from the original simple text2image interfaces that midjourney and dalle use.

The models will generalize (because that's the most efficient way of storing concepts) and you can make an argument that that means they understand a concept. Claiming "it's not learning concepts, only statistical probabilities" trivialises what a modern neural network with billions of parameters and dozens of layers is capable of doing. If a model learns how to put a concept like line width 5, 10 and 15 pixels into a continuous internal latent property, you can probably go outside this at inference at least partially.

I would argue that improving this is at this point more about engineering and less about some underlying unreconcilable differences. At the very least we learn a lot about what exactly generalization and learning means.