If you acquire the material illegally that a separate issue that training AI doesn't give you any protection against.
And as long as OpenAI have an office in Japan they can absolutely legally train the models, no?
Also, the notion of the downloading itself being an illegal act is not universal as others have pointed out.
Making and having your own copies, and doing what you want with them, has always been fine. At worst it's a grey area, but in many cases it's been protected as fair use.
Whether people have been sued for downloading works they don't have the right to copy onto their machines is irrelevant to whether it is actually illegal. And it certainly has nothing to do with fair use, which is about copyrighted works that you actually do have some right to.
IANAL so I'm not going to tell anyone what does and what does not constitute fair use in what jurisdictions.
BTW people are being sued for distribution because they make great examples because their offenses and thus the damages are much greater.
I'm not familiar with the state of play outside the US, but the US is one of the stricter jurisdictions in this regard, for reasons that have mostly to do with sophisticated corruption.
I'm responding to the "for me not for thee" and the top comment about there being an inconsistency between the treatment of large companies and the treatment of individuals in this case.
Unless people are typically punished for downloading and using copyrighted content, there is no such inconsistency. They are not, so there is not.
Copyright troll lawsuits have been fairly public and widely covered in the tech press, and most criminal prosecutions come with a formulaic, gloating press release from the law enforcement folks responsible. So it's pretty easy to follow this stuff.
The DMCA does allow harassment by copyright holders to individuals suspected of infringement. It's just that most people like authors wouldn't blow their legal budget suing kids.
Everyone who was sued by the RIAA was done so for possessing the music in a publicly accessible method. Distribution was never actually proven in any of the cases (including the high profile losses). Defendants who argued that accessibility does not qualify as distribution actually won their cases. Most who argued against the validity of the evidence acquisition also won their cases.
'Acquiring' is more difficult to pursue legally. It's easier to go after distribution. In this case, Meta or OpenAI did not distribute anything because they are not chumps. They can go after whoever posted the dataset containing books. Not sure if that is eleuther or just some random person on the internet. In either case, the strategy of going after the rich companies won't work.
Ask it to write a book similar to Harry Potter, and it'll make an attempt at it. But human writers absolutely do read Harry Potter and write similar books, and that's perfectly legal. There have probably been thousands of published books inspired by Lord of the Rings.
No amount of whining and hand wringing from engineers will ever make this true. This is for the courts to decide.
A reasonable interpretation, in my eyes, is that the training process is a black box which takes in copyrighted works and produces a training model. The training model is a derivative work of the inputs. It therefore violates the copyrights of a large number of rights holders. The outputs of the model are derivative works which also violate copyright.
And anyone using or training a model trained on works for which they do not have the rights? Completely fucked. Or at least, they must accept this as a real risk.
I won't be particularly thrilled if that turns out to be the case, but I wouldn't be surprised if it does.
But as you say, we won't know until it's tested in court. And even then, often court cases around a complex topic like this will end up with a ruling that only clarifies a narrow aspect of it. So it might take many related court cases before we have a pretty good understanding of where the law stands. And then, of course, the law could change.
Now, can you get it to output a derivative work? Maybe. Is every output a derivative work? Maybe not.
What is the blackbox “limit” here? Is the mean value of all images in imagenet (which contains many copyrighted images) violating copyright? Is the character count of sarah silverman’s books? What about a prime number representing them - https://en.wikipedia.org/wiki/Illegal_number?wprov=sfti1
Training is much more similar to a character count than an illegal prime in my view, and thus, is almost certainly going to be okay/found to be okay. If not, something like, 90% of all models used today had some component trained on copyrighted data of some form.
Its unmistakably not a derivative work of the inputs individually or collectively, since a derivative work must be itself an distinct work of authorship (the same as the work of authorship requirement for copyright), and the output of a purely mechanical process is not.
The collection of inputs itself might be a derivative work of the individual inputs, before considering Fair Use.
Please omit flamebait and swipes, as the site guidelines ask: https://news.ycombinator.com/newsguidelines.html. Your comment would have been fine without that bit.