Comment by wraptile
3 years ago
The thing is people want the benefits of having their stuff public but not bear the costs. Scraping has been mostly a solved problem especially when it comes to broad crawling. Put it under a login, there, no more AI "stealing" your work.
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.
Here are my 2 cents, I think we will need some laws specifying two types of AI models, ones trained with full consent (opt-in) for its training material and ones without. The first one would be like Adobe's firefly model where they allegedly own everything they trained it with, or something where you go around asking for consent for each thing in your training corpus (probably unfeasible for large models). Maybe things in the public domain would be ok to train with. In this case there are no restrictions and the output from such models can even be copyrighted.
Now for the second type, representing models such as Stable Difusion and Chat GPT, it would be required to have their trained model freely available to anyone and any resulting output would not be copyrightable. It may be a more fairer way of allowing anyone to harness the power of AI models that contain essentially the knowledge of all man kind, but without giving any party an unfair monopoly on it.
This should be easily enforceable for big corporations, else it would be too obvious if they are trying to pass one type model as another or even keep the truth about their model from leaking. It might not be as easy to keep small groups or individuals from breaking those rules, but hey, at least it evens the playing field.
Is that login statement strictly true? Unless the login is paid, there's no reason we can't get to (if not already there) the point where the AI scraper can just create a login first.
No, eforcing click-wrap legal agreements is actually possible. With basic KYC the scraper would instantly open up itself for litigation and no internet art piece is frankly worth this sort of trouble.
But then you can rate-limit to a point where scraping everything will take a considerable amount of time.
Of course the workaround would be to have multiple accounts, but that in turn can be made unscalable with a "prove you're human" box.
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.