Discussion
It's not relevant for this purpose whether we're talking about a single page or a whole book, if the book was legally acquired to begin with. Simply put, training a general AI model is among the things you can legally do with a work you have a legal copy of.
The ruling does make it clear that the copying of works from illegal shadow libraries like LibGen is an entirely different matter.
He says that the test result *favours* a funding of no infringement. That's different. It means that less copying would have been more infringement.
To me, that makes no sense.
As Alsup quotes, what matters is "the amount and substantiality of what is thereby made accessible to a public". In Alsup's reasoning (which I agree with), you are focusing here on the wrong question -- what went into the model, vs. what it generates.
That said, I don't see the "less copying would have been more infringement" argument. Where do you see that in the text? AIUI, that would only be true if the model generated more infringing outputs.
Twice. Once for the "replacement* copies of the used books they bought and destroyed, and then again for the many copies they made of those copies.
That is very different from what the Google Books court said.
Can you quote the exact passage(s) you're referring to?
In the third factor test, Alsup says that plaintiffs' claim that they could have used other books, or no books at all, is simply not sufficient to hold against fair use: "Because using so many works was reasonably necessary, using any one work for actually training LLMs was about as reasonable as the next." Again, he points out that there were no "revelations to the public of any
portion of the works themselves".
(Copy from https://storage.courtlistener.com/recap/gov.uscourts.cand.434709/gov.uscourts.cand.434709.231.0_3.pdf, page numbers as in document.)
"The third factor thus *favors* fair use" and "The third fair use factor *favors* fair use"
(Emphasis in the original)
In my reading, this is Alsup simply applying the four factors and stating for each whether the specific circumstances weigh in favor of (or succinctly, "favor") a fair use interpretation. He is not saying that copying less would be more infringing -- just that much of the specific copying at issue here can be considered fair use.
When he says "Copying the entire work was exactly what this purpose required", he is not expressing that partial copying would be _more_ infringing.
As the amount was 100%, that means less copying would be less favourable.
One of the other factor tests he concludes a neutral result. But not here.
We'll have to agree to disagree -- you are misreading the use of the term "favor" here.
There are 3 options: Against, neutral, favour.
If the result is not "neutral", it means that the judge thinks it's relevant what specific amount was used or what the degree of substantiality is.
We have quite a bit of Fair Use precedence.
See Kelly v. Arriba Soft. You can't make a thumbnail without using the full picture. Third factor: Neutral!
Are you aware of any decision where a full work was used, and the third factor was decided as favorable to Far Use?
As I noted, Alsup draws a distinction between copying for training the model vs. copying as part of the output - and in doing so quotes Authors Guild v. Google:
> What matters in such cases is not so much “the amount and substantiality of the portion used” in making a copy, but rather the amount and substantiality of what is thereby made accessible to a public for which it may serve as a competing substitute.
That is the key point for the third factor.
Well, of course, there are 4 factors for a reason.
This is a good ruling, because even if you hate LLMs, other operations on legally acquired copies would almost certainly be adversely affected if LLM training was held not to be fair use. Indexing, summarization, improving translation models, etc.
I consider it a good ruling also because unlike a lot of folks on here, I don't hate LLMs -- they're an important foundational technology for things to come.
Just because you're no longer hungry after eating a package of wonder bread (a good result), doesn't make it a good meal. :-)
I see a badly argued ruling that screams for appeal, and that may be bad for Anthropic. It may be possible to get the same or a similar result with an actually good ruling (but of course that's more work), which could be more helpful to LLM developers.
A space for Bonfire maintainers and contributors to communicate
