4 ms·

Copyright requires permission to produce derivative works. If I take a picture of Mickey Mouse, add 1 to each RGB value, it would still be covered under copyri

Copyright requires permission to produce derivative works. If I take a picture of Mickey Mouse, add 1 to each RGB value, it would still be covered under copyright. I think the term 'learn' is doing the heavy lifting in your argument here. Specifically, your argument seems to be: 1. We use the term 'learn' as shorthand for 'fitting a machine learning algorithm to data'. 2. We also use the term 'learn' for a human looking at copyrighted material to practice their craft. 3. Humans using copyrighted material to learn doesn't make their works derivative. 4. Therefore, machine learning algorithms that are trained on copyrighted material aren't producing derivative works. This argument holds an implicit assumption: if we use the same term for a process, it should covered by the same rules/laws. Which I think isn't very sound.

andybak 4y ago
I'm not sure "derivative work" has the legal meaning that would support your argument. To quote Wikipedia: > In copyright law, a derivative work is an expressive creation that includes major copyrightable elements of an original The key part here is "major copyrightable".
- karpierz 4y ago
  Can you be more specific? I'm not sure how this affects the argument I've laid out. To be clear, my argument doesn't show that statistical generative approaches to art are violating copyright by using copyrighted data. It only shows that the argument "humans learn on copyrighted stuff therefore it's legal for AI to do it" isn't a sound argument. Another argument that relies on the definition of derivative work might be able to show that you can train a model on copyrighted data without producing works that violate copyright.
  - andybak 4y ago
    I might have misunderstood you. "humans learn on copyrighted stuff therefore it's legal for AI to do it" - Learning can never be a violation of copyright. A work (digital or physical) can potentially violate copyright. So there are two potential "work" here: 1. The model itself 2. Images created with the model. I would argue that (1) can't violate copyright as it's not a derivative work (2) may or may not violate copyright but if it does, then it does so to exactly the same degree as it would if it wasn't AI generated. In other words "You're not allowed to plagarize copyrighted works no matter what production method you use".
    - karpierz 4y ago
      That's fair, I should've been more precise. We can break this into (1) and (2): (1) The model is a copyright violation as it contains untransformed copyrighted material. You can show this by simply asking for a picture of Mickey Mouse. To preempt a few (obvious) arguments about what it means to be "transformed", keep in mind an encrypted zip file containing images of Mickey Mouse would not be considered transformative, even if it happens to contain images that aren't copyrighted. (2) > then it does so to exactly the same degree as it would if it wasn't AI generated. This is the conclusion of an argument, but I'd need to see how you show that "it does so to exactly the same degree". Right now, it looks like you're taking that as a base assumption.
      - andybak 4y ago
        > (1) The model is a copyright violation as it contains untransformed copyrighted material. You can show this by simply asking for a picture of Mickey Mouse. This is a strong argument and gave me pause. I think if Disney sued they could probably win. However - examples such as this are probably quite rare. Mickey Mouse is so uniquitous that is massively overrepresented in the training and data and that has resulted in overfitting. For the vast majority of artists and almost certainly for the plaintiffs in this case, the demonstration wouldn't work. You most likely cannot produce an infringing image from a concise text prompt. Arguably one of the outcomes of legal action might be the need for a lot of effort to remove examples similar to the Mickey Mouse example. This would be a good thing but wouldn't prevent AI image generation from broadly doing the things that the plaintiffs want to stop it doing. > (2) but I'd need to see how you show that "it does so to exactly the same degree". Not sure I follow? If I produce an infringing image and a court agrees it's infringing then I'm liable. It doesn't matter how the image was produced. I don't see what there is to dispute here?
        karpierz 4y ago
        I think we're agreed on (2) For (1), I think it's handy to think of an encrypted zip file that contains a bunch of images. If you give it a valid password, it'll give you a corresponding image. If it produces an image that violates copyright, it clearly violates copyright, even if most of the images contained in it don't. It's possible that if the model required sufficient instruction to reproduce Mickey Mouse ("create a line from (0,20) to (20,30) of thickness 2, then a line from ..."), and could not do so without that, then the argument wouldn't apply. And there's definitely a grey area once we're outside of the territory of reproducing images. But we definitely aren't anywhere close to that with these models, and looking at the "AI Safety" work, it doesn't seem possible to get these models to "behave". > This would be a good thing but wouldn't prevent AI image generation from broadly doing the things that the plaintiffs want to stop it doing. Yeah, I agree that the plaintiffs aren't suing solely because the model can easily regenerate their images. But I think we'll find its insanely hard to get these models to avoid producing copyrighted images. They may be forced to license the training images.
AbrahamParangi 4y ago
However, what stable diffusion does bears no resemblance to "add 1 to each RGB value" and bears substantial resemblance to the human process of "learning". Without litigating if the model is "learning" in a human analogous way, instead I think the burden is on showing that the model is "substantially copying" and to an equal or greater degree than what would be needed to consider a work derivative.
limitedsupply 4y ago
Another counterargument is that these models do not generate anything without a prompt. So in a way it’s like a very smart photoshop: if you, a designer, use tools to create a character that resembles Mickey Mouse - this is on you, not the tool. It’s going to be interesting to hear this being argued in court.
- karpierz 4y ago
  A password-protected zip file containing Mickey Mouse photos also doesn't produce copyrighted material without the right prompt. Yet, the distributor of the zip file is definitely responsible for copyright violation. And to preempt a likely counterargument, intent is not a required factor for establishing copyright violation.
  - limitedsupply 4y ago
    The important distinction is that the model does not contain a copy of a copyrighted material. It contains data trained on copyrighted material (like our brains).
    - karpierz 4y ago
      You're making a distinction that isn't reflected in copyright law. If I take a JPEG of Mickey Mouse and then turn it into a PNG, it's not a "copy", as the bits are different. But it still contains copyrighted material. You can try to argue that the bits of PNG itself aren't an image of Mickey Mouse, but rather the algorithm that reads the PNG produces an image of Mickey Mouse. But that isn't really a relevant distinction in so far as copyright is concerned. In addition, this statement is false: > The important distinction is that the model does not contain a copy of a copyrighted material. It has been shown repeatedly that the model produces copies of training data. The copies are of course not stored as JPEGs/PNGs in the model, but they are retrievable from the model, given the correct password (prompt).
      - limitedsupply 4y ago
godelski 4y ago
> Copyright requires permission to produce derivative works. This then would put the responsibility on the creator, not the software. If I use a prompt "Mickey Mouse fighting Thanos" then the images produced are clearly derivative. But is this fair use? There's plenty of fan art out there like that. In fact, I typed that string into google and got fan artwork of Mickey as Thanos and the infinity stones are different media centers that Disney controls[0]. Making the image quite a high quality parity. I'm not a lawyer, but I don't feel like this image should be illegal. If so, doesn't that make this conversation far more complex than what you're writing here? I'd agree if I was trying to sell pictures of Mickey Mouse and claim to be Disney (and that there was no case for parody), but this does seem like a really hard line to draw and I don't think people are being honest about that. [0] https://www.reddit.com/r/thanosdidnothingwrong/comments/9z7a2d/mickey_gauntlet/ https://www.reddit.com/r/thanosdidnothingwrong/comments/9z7a...
- williamcotton 4y ago
  Exactly, the person coaxing copyrighted material from the latent space is liable. Now, a tool that burdens an individual with a torrent of copy-written material might not be worth using, but that is separate from the data embedded in the latent space of the model itself.
  - godelski 4y ago
    > the person coaxing copyrighted material from the latent space I'd be careful with this terminology. These models are able to generate more diversity in data than they are able to memorize. So I just want to make sure readers do not read this as "copying and combining things to make an image" in a copy/paste manner rather than the same manner that humans do (very different mechanisms) of having a broader understanding of shapes and fuzzy concepts. But I'm also a generative modeling researcher and so I'm biased to be pedantic.
    - williamcotton 4y ago
      Thanks for the clarification!

andybak 4y ago

I'm not sure "derivative work" has the legal meaning that would support your argument. To quote Wikipedia: > In copyright law, a derivative work is an expressive creation that includes major copyrightable elements of an original The key part here is "major copyrightable".

karpierz 4y ago
Can you be more specific? I'm not sure how this affects the argument I've laid out. To be clear, my argument doesn't show that statistical generative approaches to art are violating copyright by using copyrighted data. It only shows that the argument "humans learn on copyrighted stuff therefore it's legal for AI to do it" isn't a sound argument. Another argument that relies on the definition of derivative work might be able to show that you can train a model on copyrighted data without producing works that violate copyright.
- andybak 4y ago
  I might have misunderstood you. "humans learn on copyrighted stuff therefore it's legal for AI to do it" - Learning can never be a violation of copyright. A work (digital or physical) can potentially violate copyright. So there are two potential "work" here: 1. The model itself 2. Images created with the model. I would argue that (1) can't violate copyright as it's not a derivative work (2) may or may not violate copyright but if it does, then it does so to exactly the same degree as it would if it wasn't AI generated. In other words "You're not allowed to plagarize copyrighted works no matter what production method you use".
  - karpierz 4y ago
    That's fair, I should've been more precise. We can break this into (1) and (2): (1) The model is a copyright violation as it contains untransformed copyrighted material. You can show this by simply asking for a picture of Mickey Mouse. To preempt a few (obvious) arguments about what it means to be "transformed", keep in mind an encrypted zip file containing images of Mickey Mouse would not be considered transformative, even if it happens to contain images that aren't copyrighted. (2) > then it does so to exactly the same degree as it would if it wasn't AI generated. This is the conclusion of an argument, but I'd need to see how you show that "it does so to exactly the same degree". Right now, it looks like you're taking that as a base assumption.
    - andybak 4y ago
      > (1) The model is a copyright violation as it contains untransformed copyrighted material. You can show this by simply asking for a picture of Mickey Mouse. This is a strong argument and gave me pause. I think if Disney sued they could probably win. However - examples such as this are probably quite rare. Mickey Mouse is so uniquitous that is massively overrepresented in the training and data and that has resulted in overfitting. For the vast majority of artists and almost certainly for the plaintiffs in this case, the demonstration wouldn't work. You most likely cannot produce an infringing image from a concise text prompt. Arguably one of the outcomes of legal action might be the need for a lot of effort to remove examples similar to the Mickey Mouse example. This would be a good thing but wouldn't prevent AI image generation from broadly doing the things that the plaintiffs want to stop it doing. > (2) but I'd need to see how you show that "it does so to exactly the same degree". Not sure I follow? If I produce an infringing image and a court agrees it's infringing then I'm liable. It doesn't matter how the image was produced. I don't see what there is to dispute here?
      - karpierz 4y ago
        I think we're agreed on (2) For (1), I think it's handy to think of an encrypted zip file that contains a bunch of images. If you give it a valid password, it'll give you a corresponding image. If it produces an image that violates copyright, it clearly violates copyright, even if most of the images contained in it don't. It's possible that if the model required sufficient instruction to reproduce Mickey Mouse ("create a line from (0,20) to (20,30) of thickness 2, then a line from ..."), and could not do so without that, then the argument wouldn't apply. And there's definitely a grey area once we're outside of the territory of reproducing images. But we definitely aren't anywhere close to that with these models, and looking at the "AI Safety" work, it doesn't seem possible to get these models to "behave". > This would be a good thing but wouldn't prevent AI image generation from broadly doing the things that the plaintiffs want to stop it doing. Yeah, I agree that the plaintiffs aren't suing solely because the model can easily regenerate their images. But I think we'll find its insanely hard to get these models to avoid producing copyrighted images. They may be forced to license the training images.

AbrahamParangi 4y ago

However, what stable diffusion does bears no resemblance to "add 1 to each RGB value" and bears substantial resemblance to the human process of "learning". Without litigating if the model is "learning" in a human analogous way, instead I think the burden is on showing that the model is "substantially copying" and to an equal or greater degree than what would be needed to consider a work derivative.

limitedsupply 4y ago

Another counterargument is that these models do not generate anything without a prompt. So in a way it’s like a very smart photoshop: if you, a designer, use tools to create a character that resembles Mickey Mouse - this is on you, not the tool. It’s going to be interesting to hear this being argued in court.

karpierz 4y ago
A password-protected zip file containing Mickey Mouse photos also doesn't produce copyrighted material without the right prompt. Yet, the distributor of the zip file is definitely responsible for copyright violation. And to preempt a likely counterargument, intent is not a required factor for establishing copyright violation.
- limitedsupply 4y ago
  The important distinction is that the model does not contain a copy of a copyrighted material. It contains data trained on copyrighted material (like our brains).
  - karpierz 4y ago
    You're making a distinction that isn't reflected in copyright law. If I take a JPEG of Mickey Mouse and then turn it into a PNG, it's not a "copy", as the bits are different. But it still contains copyrighted material. You can try to argue that the bits of PNG itself aren't an image of Mickey Mouse, but rather the algorithm that reads the PNG produces an image of Mickey Mouse. But that isn't really a relevant distinction in so far as copyright is concerned. In addition, this statement is false: > The important distinction is that the model does not contain a copy of a copyrighted material. It has been shown repeatedly that the model produces copies of training data. The copies are of course not stored as JPEGs/PNGs in the model, but they are retrievable from the model, given the correct password (prompt).
    - limitedsupply 4y ago

godelski 4y ago

> Copyright requires permission to produce derivative works. This then would put the responsibility on the creator, not the software. If I use a prompt "Mickey Mouse fighting Thanos" then the images produced are clearly derivative. But is this fair use? There's plenty of fan art out there like that. In fact, I typed that string into google and got fan artwork of Mickey as Thanos and the infinity stones are different media centers that Disney controls[0]. Making the image quite a high quality parity. I'm not a lawyer, but I don't feel like this image should be illegal. If so, doesn't that make this conversation far more complex than what you're writing here? I'd agree if I was trying to sell pictures of Mickey Mouse and claim to be Disney (and that there was no case for parody), but this does seem like a really hard line to draw and I don't think people are being honest about that. [0] https://www.reddit.com/r/thanosdidnothingwrong/comments/9z7a2d/mickey_gauntlet/ https://www.reddit.com/r/thanosdidnothingwrong/comments/9z7a...

williamcotton 4y ago
Exactly, the person coaxing copyrighted material from the latent space is liable. Now, a tool that burdens an individual with a torrent of copy-written material might not be worth using, but that is separate from the data embedded in the latent space of the model itself.
- godelski 4y ago
  > the person coaxing copyrighted material from the latent space I'd be careful with this terminology. These models are able to generate more diversity in data than they are able to memorize. So I just want to make sure readers do not read this as "copying and combining things to make an image" in a copy/paste manner rather than the same manner that humans do (very different mechanisms) of having a broader understanding of shapes and fuzzy concepts. But I'm also a generative modeling researcher and so I'm biased to be pedantic.
  - williamcotton 4y ago
    Thanks for the clarification!