That’s like saying a pencil is a copyright violation. Tools can be used to violate copyright law, but that’s the use of the tool, not the creation of it.
I follow that up with a clarification in the next sentence.
So these companies that have feed data into their models that they have not acquired the licensing rights to, should not be allowed to continue onward until that has been rectified.
Judge William Alsup of the U.S. District Court for the Northern District of California in the Anthropic case indicated the following:
Training LLMs = Fair Use (in general)
Scanning Purchased Books = Fair Use (broadly)
Pirated Copies and Indefinite Retention = Not Transformative
Training LLMs typically is transformative because LLMs rarely give regurgitated answers, that is a copy of the data is not stored internally to the model. That storage is important later. But Judge Alsup indicated that there was narrow application to this fair use because it required a license agreement (which Anthropic had procured) to scan works into their model.
Buying a book and typing a report on it is not much different than what LLMs do and thus Judge Alsup indicated that such was also fair use and “quintessentially transformative”.
Where Judge Alsup drew the line are books and works that were scanned into the model without any permission to do so, be it they obtain that permission from the author or the publisher. Additionally, Anthropic stored the books within their system for additionally training on iterative models. This is not allowed. A model must be augmented by itself or new agreements obtained to start a new.
So you are correct that LLMs are indeed transformative and are permitted under a fair use defense. But there’s limits to that applicability. And again to turn around to what I personally believe. I think all of this is non-sense and more reasons why copyright doesn’t make sense in this age.
Also, I should note, that the output of a model can be subject to copyright violation. Just like you can use Photoshop to make something close enough to an original to get in trouble with trademark, so too can you use image generation to make a copy of something and it too would find you in trouble.
That’s not clear at all, though. Training a model is the very definition of transformational, which current copyright law acknowledges and allows.
It’s not transformational if it can quote the original input verbatim, though.
That’s like saying a pencil is a copyright violation. Tools can be used to violate copyright law, but that’s the use of the tool, not the creation of it.
Lots of people can quote things verbatim if asked, that’s not necessarily a problem.
I follow that up with a clarification in the next sentence.
Judge William Alsup of the U.S. District Court for the Northern District of California in the Anthropic case indicated the following:
Training LLMs typically is transformative because LLMs rarely give regurgitated answers, that is a copy of the data is not stored internally to the model. That storage is important later. But Judge Alsup indicated that there was narrow application to this fair use because it required a license agreement (which Anthropic had procured) to scan works into their model.
Buying a book and typing a report on it is not much different than what LLMs do and thus Judge Alsup indicated that such was also fair use and “quintessentially transformative”.
Where Judge Alsup drew the line are books and works that were scanned into the model without any permission to do so, be it they obtain that permission from the author or the publisher. Additionally, Anthropic stored the books within their system for additionally training on iterative models. This is not allowed. A model must be augmented by itself or new agreements obtained to start a new.
So you are correct that LLMs are indeed transformative and are permitted under a fair use defense. But there’s limits to that applicability. And again to turn around to what I personally believe. I think all of this is non-sense and more reasons why copyright doesn’t make sense in this age.
Also, I should note, that the output of a model can be subject to copyright violation. Just like you can use Photoshop to make something close enough to an original to get in trouble with trademark, so too can you use image generation to make a copy of something and it too would find you in trouble.