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.
If I memorize the text of Harry Potter, my brain does not thereby become a copyright infringement.
A copyright infringement only occurs if I then reproduce that text, e.g. by writing it down or reciting it in a public performance.
Training an LLM from a corpus that includes a piece of copyrighted material does not necessarily produce a work that is legally a derivative work of that copyrighted material.
If the developers have taken steps to ensure that the LLM cannot recite copyrighted material, that should count in their favor, not against them. Calling it “hiding” is backwards.
You are a human, you are allowed to create derivative works under the law. Copyright law as it relates to machines regurgitating what humans have created is fundamentally different. Future legislation will have to address a lot of the nuance of this issue.
If Google took samples from millions of different songs that were under copyright and created a website that allowed users to mix them together into new songs, they would be sued into oblivion before you could say “unauthorized reproduction.”
You simply cannot compare one single person memorizing a book to corporations feeding literally millions of pieces of copyrighted material into a blender and acting like the resulting sausage is fine because “only a few rats fell into the vat, what’s the big deal”
Terrible analogy.
Which one? And why exactly?
The analogy talks about mixing samples of music together to make new music, but that’s not what is happening in real life.
The computers learn human language from the source material, but they are not referencing the source material when creating responses. They create new, original responses which do not appear in any of the source material.
“Learn” is debatable in this usage. It is trained on data and the model creates a set of values that you can apply that produce an output similar to human speach. It’s just doing math though. It’s not like a human learns. It doesn’t care about context or meaning or anything else.
Okay, but in the context of this conversation about copyright I don’t think the learning part is as important as the reproduction part.
you bought the book to memorize from, anyway.
No, I shoplifted it from an Aldi
Let’s not pretend that LLMs are like people where you’d read a bunch of books and draw inspiration from them. An LLM does not think nor does it have an actual creative process like we do. It should still be a breach of copyright.
… you’re getting into philosophical territory here. The plain fact is that LLMs generate cohesive text that is original and doesn’t occur in their training sets, and it’s very hard if not impossible to get them to quote back copyrighted source material to you verbatim. Whether you want to call that “creativity” or not is up to you, but it certainly seems to disqualify the notion that LLMs commit copyright infringement.
This topic is fascinating.
I really do think i understand both sides here and want to find the hard line that seperates man from machine.
But it feels, to me, that some philosophical discussion may be required. Art is not something that is just manufactured. “Created” is the word to use without quotation marks. Or maybe not, i don’t know…
I wasn’t referring to whether the LLM commits copyright infringement when creating a text (though that’s an interesting topic as well), but rather the act of feeding it the texts. My point was that it is not like us in a sense that we read and draw inspiration from it. It’s just taking texts and digesting them. And also, from a privacy standpoint, I feel kind of disgusted at the thought of LLMs having used comments such as these ones (not exactly these, but you get it), for this purpose as well, without any sort of permission on our part.
That’s mainly my issue, the fact that they have done so the usual capitalistic way: it’s easier to ask for forgiveness than to ask for permission.
I think you’re putting too much faith in humans here. As best we can tell the only difference between how we compute and what these models do is scale and complexity. Your brain often lies to you and makes up reasoning behind your actions after the fact. We’re just complex networks doing math.
*could