If I download harry_potter_goblet_fire.txt off some dodgy site, then let's assume that owner of that site has infringed copyright by distributing it. If I upload it again to some other dodgy site, I would also infringe copyright in a similar same way. But that would be naughty so I'm not going to do that.
Let's say instead that I feed it into a bunch of janky pytorch scripts with a bunch of other text files, and out pops a bunch of weights. Twice.
The first model I build is a classifier. Its output is binary: is this text about wizards, yes/no.
The second model I build is an LLM. Its output is text, and (as in the article) you can get imperfect reproductions of parts of the training file out of it with the right prompts.
Now, I upload both those sets of weights to HuggingFace.
How many times am I supposed to have infringed copyright?
Is it:
A) Twice (at least), because the act of doing anything whatsoever with harry_potter_goblet_fire.txt without permission is verboten;
B) Once, because only one of the models is capable of reproducing the original (even if only approximately);
C) Zero, because neither model is capable of a reproduction that would compete with the original;
or
D) Zero, because I'm not the distributor of the file, and merely processing it - "format shifted" from the book, if you like - is not problematic in itself.
Logically I can see justifications for any of B) (tenuously), C), or D). Obviously publishers would want us to think that A) is right, but based on what? I see a lot of moral outrage, but very little actual argument. That makes me think there's nothing there.
As I understand things, the legal theory is: Everything is copyright infringement.
If I download an MP3 to my computer, the act of copying it to my hard disk is making a copy. The act of playing it copies it from my hard disk to RAM, then from RAM to the audio output; these are also copies.
That means you either need: A license (e.g. terms and conditions when you buy a legal MP3) or an "implied license" (this is what makes it legal to play CDs) or a right explicitly conferred by law (e.g. 17 U.S. Code § 117(a) lets you copy computer programs into RAM to run them) or a right of "fair use" (which allows e.g. Google copy web pages into their index)
So I'm afraid you've infringed copyright countless times - if you run several thousand epochs of training, you've infringed copyright several thousand times.
You may note that this is completely at odds with reality, and says almost nothing can happen without a copyright lawyer's permission. A cynic would say this suits the interests of copyright lawyers, if not reality, and nobody else gets to create legal theories about copyright infringement.
I’ll preface with IANACL, but you seem to be making a moral argument yourself about A), that it is not reasonable.
You have, I assume, a licensed copy of Harry Potter. That license restricts you from doing certain activities, like making (distributing? Lets go with distributing) derived works. Your models are derived works. Thus when you distribute your models, you’re violating the licence terms you “agreed” to when you acquired your copy of Harry Potter.
This is no judgement by me about whether that’s reasonable or not, just my understanding of the mechanics.
Rather than a moral argument, I think they're more just disagreeing that the spirit behind copyright law is being violated in that case. So while yes, there may be a rule in there that you may have agreed to, they disagree that such a rule is within the spirit of the law, and may reckon that it should not be a part of it even if it presently is. Like they explicitly mention how they're reflecting on the idea behind it all, rather than producing a legal analysis.
Or at least that's how I read it. I'm sure GP will clarify shortly.
More precisely, the reasoning hinges on the assertion that "Your models are derived works." which I doubt is so cut and dry. There are many processes that take external information as input. That alone clearly can't be sufficient to established that something is unambiguously a derived work.
I read a copyrighted book, it is lossy encoded into the weights of my brain. Am I a derivative work now? No. If that book inspires me to write another book in its genre, it will also not be a derivative work unless it adheres too closely to the original book.
In this scenario I personally have no physical, audio, or otherwise legitimately purchased copies of any Potters, Harry. I have entered into no direct licence agreement. If it makes things simpler, ignore the "format shifting" aside.
Nitpick: Can a license restrict you? I thought it only gave you additional rights, should you choose to accept it, but can't take away rights you have (e.g. the ability to make parodies from it). The restriction comes from IP laws themselves.
If I download harry_potter_goblet_fire.txt off some dodgy site, then let's assume that owner of that site has infringed copyright by distributing it. If I upload it again to some other dodgy site, I would also infringe copyright in a similar same way. But that would be naughty so I'm not going to do that.
Let's say instead that I feed it into a bunch of janky pytorch scripts with a bunch of other text files, and out pops a bunch of weights. Twice.
The first model I build is a classifier. Its output is binary: is this text about wizards, yes/no.
The second model I build is an LLM. Its output is text, and (as in the article) you can get imperfect reproductions of parts of the training file out of it with the right prompts.
Now, I upload both those sets of weights to HuggingFace.
How many times am I supposed to have infringed copyright?
Is it:
A) Twice (at least), because the act of doing anything whatsoever with harry_potter_goblet_fire.txt without permission is verboten;
B) Once, because only one of the models is capable of reproducing the original (even if only approximately);
C) Zero, because neither model is capable of a reproduction that would compete with the original;
or
D) Zero, because I'm not the distributor of the file, and merely processing it - "format shifted" from the book, if you like - is not problematic in itself.
Logically I can see justifications for any of B) (tenuously), C), or D). Obviously publishers would want us to think that A) is right, but based on what? I see a lot of moral outrage, but very little actual argument. That makes me think there's nothing there.