OpenAI now tries to hide that ChatGPT was trained on copyrighted books, including J.K. Rowling’s Harry Potter series::A new research paper laid out ways in which AI developers should try and avoid showing LLMs have been trained on copyrighted material.
We have to distinguish between LLMs
They are not one and the same
Yeah, this headline is trying to make it seem like training on copyrighted material is or should be wrong.
Legally the output of the training could be considered a derived work. We treat brains differently here, that’s all.
I think the current intellectual property system makes no sense and AI is revealing that fact.
I think this brings up broader questions about the currently quite extreme interpretation of copyright. Personally I don’t think its wrong to sample from or create derivative works from something that is accessible. If its not behind lock and key, its free to use. If you have a problem with that, then put it behind lock and key. No one is forcing you to share your art with the world.
Most books are actually locked behind paywalls and not free to use? Or maybe I don’t understand what you meant?
Following that, if a sailor is the sea were to put a copy of a protected book on the internet and ChatGPT was trained on it, how that argument would go? The copyright owner didn’t place it there, so it’s not “their decision”. And savvy people can make sure it’s accessible if they want to.
My belief is, if they can use all non locked data for free, then the model should be shared for free too and it’s outputs shouldn’t be subject to copyright. Just for context
Should we distinguish it though? Why shouldn’t (and didn’t) artists have a say if their art is used to train LLMs? Just like publicly displayed art doesn’t provide a permission to copy it and use it in other unspecified purposes, it would be reasonable that the same would apply to AI training.
But it kinda does. If I see a van Gogh painting, I can be inspired to make a painting in the same style.
When “ai” “learns” from an image, it doesn’t copy the image or even parts of the image directly. It learns the patterns involved instead, over many pictures. Then it uses those patterns to make new images.
Ah, but that’s the thing. Training isn’t copying. It’s pattern recognition. If you train a model “The dog says woof” and then ask a model “What does the dog say”, it’s not guaranteed to say “woof”.
Similarly, just because a model was trained on Harry Potter, all that means is it has a good corpus of how the sentences in that book go.
Thus the distinction. Can I train on a comment section discussing the book?
Good news, they already do! Artists can license their work under a permissive license like the Creative Commons CC0 license. If not specified, rights are reserved to the creator.
I know, but one of the biggest conflicts between artists and AI developers is that they didn’t seek a license to use them for training. They just did it. So even if the end result is not an exact reproduction, it still relied on unauthorized use.
Unfortunately AI training sets don’t tend to respect those licenses. Since it’s near impossible to prove they used it without permission they’re SoL
Output from an AI has just been recently considered as not copyrightable.
I think it stemmed from the actors strikes recently.
It was stated that only work originating from a human can be copyrighted.
Where can I read more about this? I’ve seen it mentioned a few times, but never with any links.
They clearly only read the headline If they’re talking about the ruling that came out this week, that whole thing was about trying to give an AI authorship of a work generated solely by a machine and having the copyright go to the owner of the machine through the work-for-hire doctrine. So an AI itself can’t be authors or hold a copyright, but humans using them can still be copyright holders of any qualifying works.