4 ms·

There are two legitimate points where copyright violation can occur with LLMs. (Not arguing the merits of copyright, just based on the concept as it is). One

by andy99 1y ago

There are two legitimate points where copyright violation can occur with LLMs. (Not arguing the merits of copyright, just based on the concept as it is). One is when copyrighted material is "pirated" for use in training, i.e. you torrent "the pile" instead of paying to acquire the books. The other is when a user uses an LLM to generate a work that violates copyright. Training itself isn't a violation, that's common sense. I am aware of lots of copyrighted things, and I could generate a work that violates copyright. My knowing this in and of itself isn't a violation. The fact that an LLM agrees to help someone violate copyright is a failure mode, on par with telling them how to make meth or whatever other things their creators don't want them doing. There's a good argument for hardening them against requests to generate copyrighted content, and this already happens.

singleshot_ 1y ago
This is an interesting comment which avoids the issue: when a user uses an LLM to violate copyright, who is liable, and how would you justify your answer?
- tux1968 1y ago
  Not OP, but I would say the answer is the same as it would be if you substitute the LLM with a live human person who has memorized a section of a book and can recall it perfectly when asked.
  - diputsmonro 1y ago
    It depends on where the money changes hands, IMO (which is basically what I think youre getting at). If you pay someone to perfectly recite a copywrited work (as you pay ChatGPT to do), then it would definitely be a violation. The situation is similar with image generation. An artist can draw a picture of Mickey Mouse without any issue. But if you pay an artist to draw you the same picture, that would also be a violation. With generative tools, the users are not themselves writers or artists using tools - they are effectively commissioners, commissioning custom artwork from an LLM and paying the operator. If someone built a machine that you put a quartner in, cranked a handle, and then printed out pictures of the Disney character you choose, then Disney is right in demanding them to stop (or more likely, force a license deal). Whatever technology drives the machine, whether an AI model or an image database or a mechanical turk, is largely immaterial.
    - fc417fc802 1y ago
      > An artist can draw a picture of Mickey Mouse without any issue. But if you pay an artist to draw you the same picture, that would also be a violation. I don't believe that's correct. The issue is not money changing hands but rather the reproduction itself. Even if I give it away for free I'm still violating IP law. There's also a fundamental issue with your argument - LLMs aren't recognized as having legal agency. If I pay an artist to violate IP law then the artist, being a human, is presumably at fault in addition to myself. Same for a company (owned by people). But tools are different. If I vandalize someone's car with a hammer the hardware store isn't at fault for selling it to me. I'm at fault for how I chose to use the tool that I purchased (or rented access to in the case of a hosted LLM). > If someone built a machine that you put a quartner in This is a flawed example because the machine was designed with the specific intention of reproducing a copyrighted work. That is different from a general purpose tool which can potentially be misused by the wielder.
      - fragmede 1y ago
        but without money changing hands, it becomes a whole lot less interesting. at the end of it, the boy who lived isn't a story about broke washed up nobody with an uncouth uncle, standing in line for the soup kitchen; if money goes away, if it turns out this whole capitalism was for a tada, what then?
- quesera 1y ago
  Why would the answer here be any different than when using a photocopier, brain, or other tool for the same purpose?
- jplusequalt 1y ago
  The company who trained the LLM. They're the one's who used the copyrighted material in their training set. Claiming they were unaware is not an excuse.
  - fc417fc802 1y ago
    It's an interesting conundrum. If I take an extremely large panoramic photograph and then fail to censor out small copyrighted sections of it, am I violating copyright law? It's not a perfect analogy by any means but it does serve to illustrate the difference in intent between distributing a particular work versus creating something that happens to incorporate copyrighted material verbatim but doesn't have any inherent need to or purpose in doing so.
    - dijksterhuis 1y ago
      > If I take an extremely large panoramic photograph and then fail to censor out small copyrighted sections of it, am I violating copyright law? It depends. Is it just under copyright, or is the featured location trademarked too? Is the photograph for commercial purposes? Is the featured location generally accepted as being part of a cityscape / landscape? * Eiffel tower: https://wiki.gettyimages.com/897/ https://wiki.gettyimages.com/897/ * Millennium-wheel: https://wiki.gettyimages.com/british-airways-london-eye-millennium-wheel/ https://wiki.gettyimages.com/british-airways-london-eye-mill... * Pro sports venues: https://wiki.gettyimages.com/pro-sport-stadiums-and-venues/ https://wiki.gettyimages.com/pro-sport-stadiums-and-venues/ * Hollywood sign: https://wiki.gettyimages.com/hollywood-sign/ https://wiki.gettyimages.com/hollywood-sign/ https://www.youtube.com/watch?v=KUdQ7gxU6Rg https://www.youtube.com/watch?v=KUdQ7gxU6Rg
201984 1y ago
If the data of a copyrighted work is effectively embedded in the model weights, does that not make the LLM itself an illegal copy? The weights are just a binary file after all.
- dboreham 1y ago
  The Dude Doctrine applies here: "That's just, like uh, your opinion, man".
TGower 1y ago
The only way to include a book in a training dataset for LLMs without violating copyright law is to contact the rights holder and buy a license to do so. Buying an ebook license off Amazon isn't enough for this, and creating a digital copy from a physical copy for your commercial use is also against the law. A good rule of thumb is if it would be illegal for a company to distribute the digital file to empolyees for training, it's definetally illegal to train an AI the company will own on it.
- Ekaros 1y ago
  It is widely accepted in many jurisdictions that different types of uses have different types of copy right schemes. Especially true for video and music content. You can't just take DVD/Blueray copy of movie and show it to movie theatre in many places. Or copy a CD and play it on radio. I see no reason why training AI should be treated like human reading. Especially if it is repeated. And more so if copies are illegally acquired like torrents.

by andy99 1y ago

singleshot_ 1y ago

This is an interesting comment which avoids the issue: when a user uses an LLM to violate copyright, who is liable, and how would you justify your answer?

tux1968 1y ago
Not OP, but I would say the answer is the same as it would be if you substitute the LLM with a live human person who has memorized a section of a book and can recall it perfectly when asked.
- diputsmonro 1y ago
  It depends on where the money changes hands, IMO (which is basically what I think youre getting at). If you pay someone to perfectly recite a copywrited work (as you pay ChatGPT to do), then it would definitely be a violation. The situation is similar with image generation. An artist can draw a picture of Mickey Mouse without any issue. But if you pay an artist to draw you the same picture, that would also be a violation. With generative tools, the users are not themselves writers or artists using tools - they are effectively commissioners, commissioning custom artwork from an LLM and paying the operator. If someone built a machine that you put a quartner in, cranked a handle, and then printed out pictures of the Disney character you choose, then Disney is right in demanding them to stop (or more likely, force a license deal). Whatever technology drives the machine, whether an AI model or an image database or a mechanical turk, is largely immaterial.
  - fc417fc802 1y ago
    > An artist can draw a picture of Mickey Mouse without any issue. But if you pay an artist to draw you the same picture, that would also be a violation. I don't believe that's correct. The issue is not money changing hands but rather the reproduction itself. Even if I give it away for free I'm still violating IP law. There's also a fundamental issue with your argument - LLMs aren't recognized as having legal agency. If I pay an artist to violate IP law then the artist, being a human, is presumably at fault in addition to myself. Same for a company (owned by people). But tools are different. If I vandalize someone's car with a hammer the hardware store isn't at fault for selling it to me. I'm at fault for how I chose to use the tool that I purchased (or rented access to in the case of a hosted LLM). > If someone built a machine that you put a quartner in This is a flawed example because the machine was designed with the specific intention of reproducing a copyrighted work. That is different from a general purpose tool which can potentially be misused by the wielder.
    - fragmede 1y ago
      but without money changing hands, it becomes a whole lot less interesting. at the end of it, the boy who lived isn't a story about broke washed up nobody with an uncouth uncle, standing in line for the soup kitchen; if money goes away, if it turns out this whole capitalism was for a tada, what then?
quesera 1y ago
Why would the answer here be any different than when using a photocopier, brain, or other tool for the same purpose?
jplusequalt 1y ago
The company who trained the LLM. They're the one's who used the copyrighted material in their training set. Claiming they were unaware is not an excuse.
- fc417fc802 1y ago
  It's an interesting conundrum. If I take an extremely large panoramic photograph and then fail to censor out small copyrighted sections of it, am I violating copyright law? It's not a perfect analogy by any means but it does serve to illustrate the difference in intent between distributing a particular work versus creating something that happens to incorporate copyrighted material verbatim but doesn't have any inherent need to or purpose in doing so.
  - dijksterhuis 1y ago
    > If I take an extremely large panoramic photograph and then fail to censor out small copyrighted sections of it, am I violating copyright law? It depends. Is it just under copyright, or is the featured location trademarked too? Is the photograph for commercial purposes? Is the featured location generally accepted as being part of a cityscape / landscape? * Eiffel tower: https://wiki.gettyimages.com/897/ https://wiki.gettyimages.com/897/ * Millennium-wheel: https://wiki.gettyimages.com/british-airways-london-eye-millennium-wheel/ https://wiki.gettyimages.com/british-airways-london-eye-mill... * Pro sports venues: https://wiki.gettyimages.com/pro-sport-stadiums-and-venues/ https://wiki.gettyimages.com/pro-sport-stadiums-and-venues/ * Hollywood sign: https://wiki.gettyimages.com/hollywood-sign/ https://wiki.gettyimages.com/hollywood-sign/ https://www.youtube.com/watch?v=KUdQ7gxU6Rg https://www.youtube.com/watch?v=KUdQ7gxU6Rg

201984 1y ago

If the data of a copyrighted work is effectively embedded in the model weights, does that not make the LLM itself an illegal copy? The weights are just a binary file after all.

dboreham 1y ago
The Dude Doctrine applies here: "That's just, like uh, your opinion, man".

TGower 1y ago

The only way to include a book in a training dataset for LLMs without violating copyright law is to contact the rights holder and buy a license to do so. Buying an ebook license off Amazon isn't enough for this, and creating a digital copy from a physical copy for your commercial use is also against the law. A good rule of thumb is if it would be illegal for a company to distribute the digital file to empolyees for training, it's definetally illegal to train an AI the company will own on it.

Ekaros 1y ago
It is widely accepted in many jurisdictions that different types of uses have different types of copy right schemes. Especially true for video and music content. You can't just take DVD/Blueray copy of movie and show it to movie theatre in many places. Or copy a CD and play it on radio. I see no reason why training AI should be treated like human reading. Especially if it is repeated. And more so if copies are illegally acquired like torrents.