If you’re executing a NN algorithm in your mind, or via pen & paper, then you are copying from the training samples, because that’s what the algorithm does. During training you compute errors against the samples, and update your weights to reduce error. During inference or generation, you use the weights (the results you remembered across all your training data) to produce an output. When your training samples are clustered in the latent space, the network will only remember an average of the samples, but samples that are sparse and don’t have close neighbors are sometimes remembered verbatim because there’s nothing nearby to average from. You can legally run the algorithm all you want on your own. Once you run it and then distribute the output, it might be in violation of Copyright Law if you accidentally reproduced one of the samples. Same is true for traditional human learning, you can free copy ideas legally, but reproducing too closely something that someone else made may be against the law, even if it was accidental.
Thoughts are never illegal wrt US Copyright Law. It’s a straw man to insist on making this point.
> In other words, the end user of the model is the one to be held responsible if they reproduce and distribute the copyrighted material.
No, this is false because it is the creators of the model that 1) did not legally acquire the source material and 2) distributed the network that contains latent copies of the source material that end users can use to reproduce works from.
This is incorrect. As another poster mentioned, it is not illegal to read a stolen book. It is only illegal to steal the book.
Secondly the source material is acquired legally since it is open to consumption on the open internet.
Thirdly model does not contain “latent copies of the source material”. By using a simple test (currently legal standard) that if I showed you the node weights and counts of the network no person even trained in the art can identify it to a specific piece of work. Therefore it is at best a derivative, reasonably distinct.
Nope, this is strawman and continuing to demonstrate a misunderstanding of Copyright Law. There is no such legal standard, where did you get that? If the network can reproduce a work, then it does in fact contain a latent copy. Arguing that you can’t see it by inspecting node weights is straw man. You cannot argue that you’re not copying music if you use a new compression algorithm and then suggest it’s distinct and derivative because nobody can read the raw compressed data. That’s not how Copyright Law works. If you can approximately re-perform someone else’s work, you’re in violation. This is true even if you have to run a black-box program to produce the output.
> no person even trained in the art can identify it to a specific piece of work
Ironically, you’re actually admitting that even AI researchers can’t prove the network won’t reproduce someone’s work.
The rest you seem to now be looking for a snarky gotcha, which if you don’t want to have a discussion, then I’m uninterested in discussing further. I made clear above and in a sibling comment that remixes are gray area, and this question is complicated. That said, even if AI people do acquire source material legally, they are in fact copying it and distributing it, and that part alone can potentially violate US Copyright Law. This isn’t even up for debate, so I don’t know why you’re attempting to suggest otherwise. The lawsuits mentioned in the article were brought on evidence that networks violated copyrights of specific existing works, and lots of people have found specific examples of violations.
1) the creating of the model is does not violate copyright. Claiming otherwise means running same algorithm in meatspace would violate copyright laws, which implies thoughts violates laws which is absurd.
2) distribution of the model does not violate copyright laws because the models themselves do not contain latent copies of the work. The model itself is not the work nor a recognizable copy of it nor can it be reconstituted back to the work. It is a tool more analogous to photoshop where the tool can be used to reproduce copyrighted work, yes, by the end user (where I believe the responsibility lies), but the tool itself is not copyrighted work. Microsoft word can be used to generate copyrighted books, if I’m correct. Or I can hire smarter tool: a human writer to produce copyrighted works. Is the writer-for-hire illegal himself as a person? I believe the law will eventually understand this to be the case.
> nor can it be reconstituted back to the work
This is false. It has already happened multiple times that networks reproduced copyrighted material.
Secondly you seem to be conflating the “tool itself” to “what the tool can do” to be strongly equivalent. I.e if the tool has the capability to violate laws, then the existence and distribution of the tool itself also violates said law. (Not so)
> if the tool has the capability to violate laws, then the existence and distribution of the tool itself violates said law.
That’s right if you remove the word “existence”. Distribution of a NN model that violates copyright by reproducing copyrighted works is illegal. That part has been my point in this thread, it seems like you understand now and we agree.
It’s “existence” is not illegal under US Copyright Law unless you didn’t have the legal right to use the training material, and in that case it’s illegal to use the material whether you used a computer or your brain, it doesn’t matter how you created the neural network (or even whether you created a neural network), the violation there isn’t the act of creating the network, it’s the act of stealing and using material you don’t have permission to use.
This whole discussion would be a lot less frustrating for you if instead of making assumptions and logic arguments about brains and computers, you took some time to read the copyright legal code. https://www.copyright.gov/
Cars, phones, guns, knives (practically anything) can be used to generate activities that break the law. They are perfectly legal to distribute. The onus on the legality of the activity lies with the end user.
> If you understand that the model is a tool, and that as a tool it can be used to generate activity that can violate laws and be used for other perfectly legal activities, then as a broad principle the distribution of said tool is not a violation of said laws
That statement is incorrect, the logic is flawed. Just because a tool has both legal and illegal uses does not necessarily have any bearing whatsoever on whether the tool’s distribution is legal. Tools that are illegal to distribute can have legal uses, and that does not make them legal to distribute.
You are talking in circles. My last response here.
A silly example. Making GPT write a rap battle between Keynes and Mises goes beyond a performative remix, it is transformational work, nothing is copied explicitly. If a human were to write it that would not violate copyright.
I think that to tackle this we need a new lens other than copyright in the long term.
The argument that NNs aren’t memorizing is definitely debatable and not necessarily true. They are designed to memorize deltas and averages from examples. They are, at the most fundamental level, building high dimensional splines to approximate their training data, and intentionally trying to minimize the error between the output and the examples. It’s fair to say that “usually” they don’t remember any single training sample, but it’s very easy for NNs to accidentally remember outliers verbatim. The whole reason the lawsuits mentioned in the article are happening is because we keep finding more and more examples where the network has reproduced someone’s specific work in large part. If we’re going to claim that today’s AI is producing original work, then we have to guarantee it, not just assert that it doesn’t usually happen.
> a rap battle between Keynes and Mises goes beyond a performative remix, it is a transformational work, nothing is copied explicitly.
I don’t buy that the work can be called transformational just because the remix doesn’t have any recognizable snippets. GPT is in fact copying individual words explicitly, and it’s putting words together by studying the statistical occurrence of words in context of other words.
> I think that to tackle this we need a new lens other than copyright
I totally agree with that. This question is legitimately hard. We do need a new lens, but we might have to keep and respect the old one too at the same time. I feel like AI work should acknowledge that difficulty and step up to lead the curation of training sets that are legal wrt copyright by design, rather than ignoring the concerns of the very people who made the work they are leveraging.
This comment fundamentally and dangerously misunderstands Copyright Law. Insights are not copyrighted, nor are they copyrightable. Copyright law controls who gets to distribute a specific “fixation” or performance of work. It is not, and never was about preventing the spread of ideas. Authors and artists have always intended for you to read/observe/listen to their work when you legally acquire a copy. They just want you to not copy it verbatim, but go do your own original work if you want to distribute or sell something.
The whole problem is that today’s NNs are specifically designed to remember and remix only the fixed performative parts of the work, and they, unlike humans, don’t understand the insights at all. They are just deterministic machines that copy and remix at a large scale. As such, it’s pretty clear the people training AI today should expect to have to ask permission before “training” (copying) other people’s work.