4 ms·

Can you block the training? It would seem you could block the generation of copyright violations. But training an AI is really the same as me listening to the

Can you block the training? It would seem you could block the generation of copyright violations. But training an AI is really the same as me listening to the music. The problem occurs if I copy the music.

EamonnMR 4y ago
> But training an AI is really the same as me listening to the music. That remains to be seen, legally. I think that if copyright law wants to keep existing it's going to have to treat training as a form of copying.
- nradov 4y ago
  Copyright law already has a section on derivative works. No new laws are needed.
  - kenjackson 4y ago
    But don’t you have to prove the derivative work based on the content of the work? For example you can’t argue that NBA Youngboy’s songs are derivative of the Beetles just because he heard them. You have to show in the work it is derivative. Likewise, you’d need to show the output of the AI is derivative.
    - nradov 4y ago
      Yes, exactly. Existing copyright law is adequate to cover that situation.
      - kenjackson 4y ago
        Funny how the same exact statement with different context (from my perspective) leads to different interpretations. I fully agree with your statement now. :-)
    - EamonnMR 4y ago
      Are you sure the weights themselves aren't a derivative work?
bheadmaster 4y ago
The difference between "recording" and "training" is really fuzzy. Not even the "exact reproduction" provides a well-defined line, as re-mixing and modifying a song is still considered a copyright violation. People love to draw the analogy to human listening, and while in principle I agree, the AI is still not human and the argument isn't necessarily valid - it needs arguments to why it's valid, too.
- tjr 4y ago
  People love to draw the analogy to human listening, and while in principle I agree, the AI is still not human and the argument isn't necessarily valid It could be worth considering: humans are [ostensibly, I claim] intelligent by default. There are obviously different details of scenarios, but humans can learn to speak by just being around their immediate family; they may only read a few dozen books in their lifetime; they can appreciate music listening to their very first album. AI (at least the current rendition of it) is trained on a massive collection of data before we could even start to claim they are intelligent. You can't take an empty neural network and have it listen to a single album and it gets anything worthwhile out of that. It can't read a few dozen books and be able to do much of anything. It needs a much, much larger data set before we would even think of saying it was intelligent. So I think, to compare a human listening to an album, and a trained AI system listening to an album, yes, those two things might be reasonably analogous. But to even get the AI system to that point, it needed to be built up using a huge quantity of data. Do the copyright holders on that data concur with that training usage? I think, in that respect, there is a difference compared to a human listening to a recording. Randomly pontificating in a discussion forum here; not offering a well-thought-out plan for everyone to live by!
  - bheadmaster 4y ago
    > So I think, to compare a human listening to an album, and a trained AI system listening to an album, yes, those two things might be reasonably analogous. This still relies on the implicit assumption that AI and human mind work the same way. While AI and humans may generate similar-looking results, and need a similar amount of data fed into them, is that really enough evidence that they work the same way, and should therefore be treated the same way? Additionally, this analogy is ignoring the fact that human mind cannot be replicated, scaled and automated, while an AI can. Isn't that aspect highly relevant in case of producing content?
triceratops 4y ago
> But training an AI is really the same as me listening to the music That's what we think right now. But copyright never anticipated that machines could listen to music, read books, or watch movies. The scale at which machines can consume content, and the extent to which they can remember it, might make it significantly different.
wgjordan 4y ago
> But training an AI is really the same as me listening to the music. The problem occurs if I copy the music. The trained AI model is a digital artifact that can be copied and distributed in a way that 'me listening to the music' cannot. The model contains detailed information about copyrighted content in a format capable of reconstituting infringing derivative works from a suitable prompt. It's 'really the same as' a huge content archive stored in a proprietary format with some lossy compression. > The problem occurs if I copy the music. Yes, I would argue that the 'problem' (copyright infringement) occurs when a trained AI model (essentially a big content archive) is copied and publicly distributed. For a hosted service, I would argue that the infringement occurs when the service copies and distributes an infringing derivative work in response to a prompt.

EamonnMR 4y ago

> But training an AI is really the same as me listening to the music. That remains to be seen, legally. I think that if copyright law wants to keep existing it's going to have to treat training as a form of copying.

nradov 4y ago
Copyright law already has a section on derivative works. No new laws are needed.
- kenjackson 4y ago
  But don’t you have to prove the derivative work based on the content of the work? For example you can’t argue that NBA Youngboy’s songs are derivative of the Beetles just because he heard them. You have to show in the work it is derivative. Likewise, you’d need to show the output of the AI is derivative.
  - nradov 4y ago
    Yes, exactly. Existing copyright law is adequate to cover that situation.
    - kenjackson 4y ago
      Funny how the same exact statement with different context (from my perspective) leads to different interpretations. I fully agree with your statement now. :-)
  - EamonnMR 4y ago
    Are you sure the weights themselves aren't a derivative work?

bheadmaster 4y ago

The difference between "recording" and "training" is really fuzzy. Not even the "exact reproduction" provides a well-defined line, as re-mixing and modifying a song is still considered a copyright violation. People love to draw the analogy to human listening, and while in principle I agree, the AI is still not human and the argument isn't necessarily valid - it needs arguments to why it's valid, too.

tjr 4y ago
People love to draw the analogy to human listening, and while in principle I agree, the AI is still not human and the argument isn't necessarily valid It could be worth considering: humans are [ostensibly, I claim] intelligent by default. There are obviously different details of scenarios, but humans can learn to speak by just being around their immediate family; they may only read a few dozen books in their lifetime; they can appreciate music listening to their very first album. AI (at least the current rendition of it) is trained on a massive collection of data before we could even start to claim they are intelligent. You can't take an empty neural network and have it listen to a single album and it gets anything worthwhile out of that. It can't read a few dozen books and be able to do much of anything. It needs a much, much larger data set before we would even think of saying it was intelligent. So I think, to compare a human listening to an album, and a trained AI system listening to an album, yes, those two things might be reasonably analogous. But to even get the AI system to that point, it needed to be built up using a huge quantity of data. Do the copyright holders on that data concur with that training usage? I think, in that respect, there is a difference compared to a human listening to a recording. Randomly pontificating in a discussion forum here; not offering a well-thought-out plan for everyone to live by!
- bheadmaster 4y ago
  > So I think, to compare a human listening to an album, and a trained AI system listening to an album, yes, those two things might be reasonably analogous. This still relies on the implicit assumption that AI and human mind work the same way. While AI and humans may generate similar-looking results, and need a similar amount of data fed into them, is that really enough evidence that they work the same way, and should therefore be treated the same way? Additionally, this analogy is ignoring the fact that human mind cannot be replicated, scaled and automated, while an AI can. Isn't that aspect highly relevant in case of producing content?

triceratops 4y ago

> But training an AI is really the same as me listening to the music That's what we think right now. But copyright never anticipated that machines could listen to music, read books, or watch movies. The scale at which machines can consume content, and the extent to which they can remember it, might make it significantly different.

wgjordan 4y ago

> But training an AI is really the same as me listening to the music. The problem occurs if I copy the music. The trained AI model is a digital artifact that can be copied and distributed in a way that 'me listening to the music' cannot. The model contains detailed information about copyrighted content in a format capable of reconstituting infringing derivative works from a suitable prompt. It's 'really the same as' a huge content archive stored in a proprietary format with some lossy compression. > The problem occurs if I copy the music. Yes, I would argue that the 'problem' (copyright infringement) occurs when a trained AI model (essentially a big content archive) is copied and publicly distributed. For a hosted service, I would argue that the infringement occurs when the service copies and distributes an infringing derivative work in response to a prompt.