OpenAI's justification for why training data is fair use, not infringement [pdf]

Re:OpenAI's justification for why training data is fair use,... by Certhas on Thursday October 05, 2023 @02:02PM (#37781851)

So we need new legislation.

Edit: To be more explicit, on my view both the parent and the op are right, and they don't contradict each other: We don't need new copyright legislation. We need new legal concepts.

I think both people are right: There is an overwhelming need to address the situation of artists and writers ("content creators"). Copyright is not being used because it's the right tool to do so, but because it's the only even vaguely plausible one. It's also clear that it's not sufficient.

[ Parent | on HN ]

Re:OpenAI's justification for why training data is fair use,... by troupe on Thursday October 05, 2023 @03:20PM (#37783010)

We have legal concepts that fit. Ideas cannot be copyrighted. The particular expression of an idea can be copyrighted, but an idea itself is in the public domain. The ideas you or an LLM get from reading a book are in the public domain while the book (the expression of those ideas) is not.

In what way is that not sufficient?

[ Parent | on HN ]

Re:OpenAI's justification for why training data is fair use,... by pclmulqdq on Thursday October 05, 2023 @05:09PM (#37784325)

There is no legal concept for a machine to have ideas. I agree with you that the current legal framework is clear, but I disagree as to the outcome: without new legislation, the creation of an LLM on unlicensed content seems to be the creation of a derived work without permission of the owners whose work you are deriving from, so the creation of the LLM itself is copyright infringement.

[ Parent | on HN ]

Re:OpenAI's justification for why training data is fair use,... by troupe on Thursday October 05, 2023 @05:50PM (#37784726)

But we already have a way to deal with derived work. Reading a bunch of books and then writing with the ideas you learned is perfectly fine. Copying a paragraph without citing it is not. Mimic the style of authors you want to emulate is fine. Passing off their book as your own is not.

If creating an LLM is copyright infringement, then we could say that a human reading a book is copyright infringement.

[ Parent | on HN ]

Re:OpenAI's justification for why training data is fair use,... by Certhas on Friday October 06, 2023 @03:42AM (#37788145)

And thus copyright is not a good framework for LLMs...

It's obvious that this technology was not considered when developing current legal concepts, so if the concepts produce unsatisfactory outcomes, they need to be reevaluated.

[ Parent | on HN ]

Re:OpenAI's justification for why training data is fair use,... by sensanaty on Friday October 06, 2023 @07:37AM (#37789383)

But a human reading !== A computer "reading".

Why do people so easily and readily conflate the two? Is it just because it makes the AI hype sound more real, if you compare it 1:1 with a human brain?

Machines and humans are not equal or even in the ballpark of being similar in any way other than superfluous ones, and should (and already are) held by different standards, both legal and otherwise.

[ Parent | on HN ]

Re:OpenAI's justification for why training data is fair use,... by Certhas on Friday October 06, 2023 @03:28AM (#37788070)

I don't understand your argument. I concur with the GP (https://news.ycombinator.com/item?id=37780772) that there are legitimate issues around LLMs that need addressing. I say that current legal principles seem inadequate to address these issues. Your "counterpoint" seems to merely reiterate the latter. You seem to agree that these concepts can not be used effectively to address issues around LLMs pointed out, as what the LLM extracts is not copyrightable.

Do you simply disagree that there are issues worth addressing?

Also, typically the entity that carries ideas can not be owned by a cooperation. So there is a completely new relationship between property and ideas at play.

[ Parent | on HN ]