I can spit out copyrighted work verbatim.
“No Lieutenant, your men are already dead”
See?
Comment on OpenAI says it’s “impossible” to create useful AI models without copyrighted material
BraveSirZaphod@kbin.social 9 months agoThere is literally no resemblance between the training works and the model.
This is way too strong a statement when some LLMs can spit out copyrighted works verbatim.
https://www.404media.co/google-researchers-attack-convinces-chatgpt-to-reveal-its-training-data/
A team of researchers primarily from Google’s DeepMind systematically convinced ChatGPT to reveal snippets of the data it was trained on using a new type of attack prompt which asked a production model of the chatbot to repeat specific words forever.
Often, that “random content” is long passages of text scraped directly from the internet. I was able to find verbatim passages the researchers published from ChatGPT on the open internet: Notably, even the number of times it repeats the word “book” shows up in a Google Books search for a children’s book of math problems. Some of the specific content published by these researchers is scraped directly from CNN, Goodreads, WordPress blogs, on fandom wikis, and which contain verbatim passages from Terms of Service agreements, Stack Overflow source code, copyrighted legal disclaimers, Wikipedia pages, a casino wholesaling website, news blogs, and random internet comments.
Beyond that, copyright law was designed under the circumstances where creative works are only ever produced by humans, with all the inherent limitations of time, scale, and ability that come with that. Those circumstances have now fundamentally changed, and while I won't be so bold as to pretend to know what the ideal legal framework is going forward, I think it's also a much bolder statement than people think to say that fair use as currently applied to humans should apply equally to AI and that this should be accepted without question.
I can spit out copyrighted work verbatim.
“No Lieutenant, your men are already dead”
See?
But AI isn’t all about generating creative works. It’s a store of information that I can query - a bit like searching Google; but understands semantics, and is interactive. It can translate my own text for me - in which case all the creativity comes from me, and I use it just for its knowledge of language. Many people use it to generate boilerplate code, which is pretty generic and wouldn’t usually be subject to copyright.
This is how I use the AI: I learn from it. Honestly I just never got the bug on wanting it to generate creative works I can sell. I guess I’d rather sell my own creative output, you know? It’s more fun than ordering a robot to be creative for me.
I have used it as a collaborator when doing creative work. It's a great brainstorming buddy, and I use it to generate rough drafts of stuff. Usually I use it while developing roleplaying scenarios for TTRPGs I run for my friends. Generative AI is great for illustrating those scenarios, too.
I know it inherently seems like a bad idea to fix an AI problem with more AI, but it seems applicable to me here. I believe it should be technically feasible to incorporate into the model something which checks if the result is too similar to source content as part of the regression.
My gut would be that this would, at least in the short term, make responses worse on the whole, so would probably require legal action or pressure to have it implemented.
The key element here is that an LLM does not actually have access to its training data, and at least as of now, I'm skeptical that it's technologically feasible to search through the entire training corpus, which is an absolutely enormous amount of data, for every query, in order to determine potential copyright violations, especially when you don't know exactly which portions of the response you need to use in your search. Even then, that only catches verbatim (or near verbatim) violations, and plenty of copyright questions are a lot fuzzier.
For instance, say you tell GPT to generate a fan fiction story involving a romance between Draco Malfoy and Harry Potter. This would unquestionably violate JK Rowling's copyright on the characters if you published the output for commercial gain, but you might be okay if you just plop it on a fan fic site for free. You're unquestionably okay if you never publish it at all and just keep it to yourself (well, a lawyer might still argue that this harms JK Rowling by damaging her profit if she were to publish a Malfoy-Harry romance, since people can just generate their own instead of buying hers, but that's a messier question). But, it's also possible that, in the process of generating this story, GPT might unwittingly directly copy chunks of renowned fan fiction masterpiece My Immortal. Should GPT allow this, or would the copyright-management AI strike it? Legally, it's something of a murky question.
For yet another angle, there is of course a whole host of public domain text out there. GPT probably knows the text of the Lord's Prayer, for instance, and so even though that output would perfectly match some training material, it's legally perfectly okay. So, a copyright police AI would need to know the copyright status of all its training material, which is not something you can super easily determine by just ingesting the broad internet.
skeptical that it’s technologically feasible to search through the entire training corpus, which is an absolutely enormous amount of data
Google, DuckDuckGo, Bing, etc. do it all the time.
I don’t see why it wouldn’t be able to. That’s a Big Data problem, but we’ve gotten very very good at searches. Bing, for instance, conducts a web search on each prompt in order to give you a citation for what it says, which is pretty close to what I’m suggesting.
As far as comparing to see if the text is too similar, I’m not suggesting a simple comparison or even an Expert Machine; I believe that’s something that can be trained. GANs already have a discriminator that’s essentially measuring how close to generated content is to “truth.” This is extremely similar to that.
I completely agree that categorizing input training data by whether or not it is copyrighted is not easy, but it is possible, and I think something that could be legislated. The AI you would have as a result would inherently not be as good as it is in the current unregulated form, but that’s not necessarily a worse situation given the controversies.
On top of that, one of the common defenses for AI is that it is learning from material just as humans do, but humans also can differentiate between copyrighted and public works. For the defense to be properly analogous, it would make sense to me that it would need some notion of that as well.
It's actually the other way around, Bing does websearches based on what you've asked it and then the answer it generates can incorporate information that was returned by the websearching. This is why you can ask it about current events that weren't in its training data, for example - it looks the information up, puts it into its context, and then generates the response that you see. Sort of like if I asked you to write a paragraph about something that you didn't know about, you'd go look the information up first.
but humans also can differentiate between copyrighted and public works
Not really. Here's a short paragraph about sailboats. Is it copyrighted?
Sailboats, those graceful dancers of the open seas, epitomize the harmonious marriage of nature and human ingenuity. Their billowing sails, like ethereal wings, catch the breath of the wind, propelling them across the endless expanse of the ocean. Each vessel bears the scars of countless journeys, a testament to the resilience of both sailor and ship.
Thank you for your thoroughly analytical take on the subject. Solid points all around.
can you add signature instead
MudMan@kbin.social 9 months ago
I'm gonna say those circumstances changed when digital copies and the Internet became a thing, but at least we're having the conversation now, I suppose.
I agree that ML image and text generation can create something that breaks copyright. You for sure can duplicate images or use copyrighted characterrs. This is also true of Youtube videos and Tiktoks and a lot of human-created art. I think it's a fascinated question to ponder whether the infraction is in what the tool generates (i.e. did it make a picture of Spider-Man and sell it to you for money, whcih is under copyright and thus can't be used that way) or is the infraction in the ingest that enables it to do that (i.e. it learned on pictures of Spider-Man available on the Internet, and thus all output is tainted because the images are copyrighted).
The first option makes more sense to me than the second, but if I'm being honest I don't know if the entire framework makes sense at this point at all.
lily33@lemm.ee 9 months ago
The infraction should be in what’s generated. Because the interest by itself also enables many legitimate, non-infracting uses - which don’t involve generating creative work at all - or where the creative input comes from the user.
MudMan@kbin.social 9 months ago
I don't disagree on principle, but I do think it requires some thought.
Also, that's still a pretty significant backstop. You basically would need models to have a way to check generated content for copyright, in the way Youtube does, for instance. And that is already a big debate, whether enforcing that requirement is affordable to anybody but the big companies.
But hey, maybe we can solve both issues the same way. We sure as hell need a better way to handle mass human-produced content and its interactions with IP. The current system does not work and it grandfathers in the big players in UGC, so whatever we come up with should work for both human and computer-generated content.