4 ms·

I can see the argument that if you train on a currently sold copyrighted work you have to buy a normal copy, you can't just download a torrent and still call it

by andy99 3y ago

I can see the argument that if you train on a currently sold copyrighted work you have to buy a normal copy, you can't just download a torrent and still call it fair use (if you accept the concept of copyright at all...). So if someone could prove that Nvidia or whoever didn't own a copy (note who has the burden of proof) they could be liable for whatever an illegal downloading infringement is worth. Otherwise I agree that training on anything public is fair use.

stubish 3y ago
Owning a copy of a book doesn't give me the right to make copies and sell them or even give them away. You can't even do that with publicly available material unless it has been put in the public domain or exceptions like fair use or satire apply. The model contains a copy of the work, and you can't distribute or publish your model without violating copyright. Offering access to this model as a service is legally in dispute, with AI companies generally arguing that liability for copyright infringement lies with the end user.
- EMIRELADERO 3y ago
  > The model contains a copy of the work, and you can't distribute or publish your model without violating copyright. Does it? And if it did, why wouldn't fair use apply here as much as it would apply to the actual training/SaaS offering?
  - stubish 3y ago
    How is distributing a model containing copyright material any different to distributing a .zip file containing copyright material? Both contain an encoded, transformed copy of the work. And both can be used to reproduce the original work. The reproduction doesn't even need to be verbatim; just similar enough to convince a court that it is plagiarism. Fair use generally only applies for parts of a work, and not using the entire thing. AI companies are claiming fair use, claiming that the processing is transformative. This needs to be determined in court. I think the court cases currently underway will show it is not transformative, because the original work can be recovered. Fair use is also determined on a case by case basis, so I think courts won't have the power to grant blanket permission. New laws would need to be passed for that.
    - EMIRELADERO 3y ago
      > How is distributing a model containing copyright material any different to distributing a .zip file containing copyright material? Both contain an encoded, transformed copy of the work. And both can be used to reproduce the original work. The reproduction doesn't even need to be verbatim; just similar enough to convince a court that it is plagiarism. Because of the substantial non-infringing uses. A court wouldn't throw away a model that can do non-infringing stuff 90% of the time if it can infringe copyright the other 10%. > Fair use generally only applies for parts of a work, and not using the entire thing. AI companies are claiming fair use, claiming that the processing is transformative. This needs to be determined in court. I think the court cases currently underway will show it is not transformative, because the original work can be recovered. Fair use is also determined on a case by case basis, so I think courts won't have the power to grant blanket permission. New laws would need to be passed for that. Fair use can apply to any possible allegation of infringement, not just parts of a work, as long as courts find the facts are in favor of it. Whiile fair use is a case-by-case thing, courts can set precedent that will apply to all situations with similar fact patterns.
    - fragmede 3y ago
      There's a paper on this, and it's fairly convincing that they're able to plagiarize https://spectrum.ieee.org/amp/midjourney-copyright-2666872100 https://spectrum.ieee.org/amp/midjourney-copyright-266687210...
- Turing_Machine 3y ago
  > The model contains a copy of the work No, it absolutely does not. There's not enough space in even the larger models to contain verbatim copies of the training data, or anything remotely close to it.

by andy99 3y ago

stubish 3y ago

Owning a copy of a book doesn't give me the right to make copies and sell them or even give them away. You can't even do that with publicly available material unless it has been put in the public domain or exceptions like fair use or satire apply. The model contains a copy of the work, and you can't distribute or publish your model without violating copyright. Offering access to this model as a service is legally in dispute, with AI companies generally arguing that liability for copyright infringement lies with the end user.

EMIRELADERO 3y ago
> The model contains a copy of the work, and you can't distribute or publish your model without violating copyright. Does it? And if it did, why wouldn't fair use apply here as much as it would apply to the actual training/SaaS offering?
- stubish 3y ago
  How is distributing a model containing copyright material any different to distributing a .zip file containing copyright material? Both contain an encoded, transformed copy of the work. And both can be used to reproduce the original work. The reproduction doesn't even need to be verbatim; just similar enough to convince a court that it is plagiarism. Fair use generally only applies for parts of a work, and not using the entire thing. AI companies are claiming fair use, claiming that the processing is transformative. This needs to be determined in court. I think the court cases currently underway will show it is not transformative, because the original work can be recovered. Fair use is also determined on a case by case basis, so I think courts won't have the power to grant blanket permission. New laws would need to be passed for that.
  - EMIRELADERO 3y ago
    > How is distributing a model containing copyright material any different to distributing a .zip file containing copyright material? Both contain an encoded, transformed copy of the work. And both can be used to reproduce the original work. The reproduction doesn't even need to be verbatim; just similar enough to convince a court that it is plagiarism. Because of the substantial non-infringing uses. A court wouldn't throw away a model that can do non-infringing stuff 90% of the time if it can infringe copyright the other 10%. > Fair use generally only applies for parts of a work, and not using the entire thing. AI companies are claiming fair use, claiming that the processing is transformative. This needs to be determined in court. I think the court cases currently underway will show it is not transformative, because the original work can be recovered. Fair use is also determined on a case by case basis, so I think courts won't have the power to grant blanket permission. New laws would need to be passed for that. Fair use can apply to any possible allegation of infringement, not just parts of a work, as long as courts find the facts are in favor of it. Whiile fair use is a case-by-case thing, courts can set precedent that will apply to all situations with similar fact patterns.
  - fragmede 3y ago
    There's a paper on this, and it's fairly convincing that they're able to plagiarize https://spectrum.ieee.org/amp/midjourney-copyright-2666872100 https://spectrum.ieee.org/amp/midjourney-copyright-266687210...
Turing_Machine 3y ago
> The model contains a copy of the work No, it absolutely does not. There's not enough space in even the larger models to contain verbatim copies of the training data, or anything remotely close to it.