8 ms·

> artists had their art trained on without their permission and violating the copyright Okay, I'll bite. Since when does copyright require permission to learn

> artists had their art trained on without their permission and violating the copyright Okay, I'll bite. Since when does copyright require permission to learn from the copyrighted material? I don't need your permission for that, and neither should an AI. Or alternately, assuming you reject the "human-AI analogy": since when do I need permission to collect statistical information about copyrighted material? Should Disney be able to sue me for telling you how much screen time Jafar had in Disney's Aladdin, or what percentage of pixels in the first 5 minutes were a shade of blue?

johnnymorgan 4y ago
This, training is passive knowledge. Basically this is saying if you look at it, and then learn to draw you owe Disney money. Disney would love that but it's ludicrous
Snild 4y ago
Machines are not people, and don't enjoy the same rights as us. (And I would not call it "learning", at least not in the same meaning of the word as for humans, but that's a more philosophical discussion.)
- stale2002 4y ago
  > Machines are not people, and don't enjoy the same rights as us Fortunately, a machine isn't doing anything on its own. It is the person who uses the machine that is doing something. And it is the human who is allowed to use tools to create transformative work.
visarga 4y ago
Nobody's arguing that near-duplications or trademark violations are not bad. But I think they want a broader power grab - to extend copyright to art styles. In other words to protect not just the expression but also the ideas themselves. They want something like patent + trademark + copyright all in one.
karpierz 4y ago
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".
fxtentacle 4y ago
The illegal part is actually SD selling the images which contain copyrighted material. If you were to paint Mickey and sell those paintings, Disney might sue you. Why should SD (the company) be allowed to do that?
- andybak 4y ago
  I don't understand this comment. SD aren't selling images of Mickey Mouse. Can you explain your point a bit more clearly?
- sebzim4500 4y ago
  Can you link to an example where Stability AI are selling images of Mickey Mouse? Sounds pretty hard to believe tbh, but you seem confident.
  - fxtentacle 4y ago
    When you type in "Mickey Mouse" as a prompt and pay for their generation tool, they are in effect selling you a Mickey Mouse image.
notahacker 4y ago
You definitely need permission to make copies of people's art on your computers to use whilst learning though (copyright holders tend to be OK with or completely unaware of that, but it's a right they reserve) More to the point, this feels like the "have your cake and eat it" of legal arguments. If we're going to anthropomorphise AI to the extent we argue storing data in its data corpus without authorisation is just like allowing a human to look at your legally owned stuff, and reinterpreting that data in a [sometimes not particularly] lossy fashion as a set of weights and biases is just like a human learning, we should probably follow the "AI is just like a person" argument through to its logical conclusion and hold it legally responsible for not respecting copyright law if it ever produces outputs with trademarked images or uncanny resemblance to copyrighted material that aren't deemed "fair use", just like a person would be. (tbf, it's not too hard to imagine a world in which ChatGPT says "I'm sorry, but I am not able to write stories involving Mickey Mouse, because Mickey Mouse is the intellectual property of the Disney Corporation. Respecting intellectual property rights is very important", and I'm not sure that's an improvement...)
amelius 4y ago
> Since when does copyright require permission to learn from the copyrighted material? Ah, but you are anthropomorphizing an algorithm. What is happening inside the computer you call learning, but it is just some mathematical function. Other people might not agree that this is learning, and might call it "mixing" or "copying".
- sebzim4500 4y ago
  Fine then. Say I went through Star Wars and counted how many blue pixels there were in each frame and published a list of the counts. Would you claim that list would be infringing? Doing statistical analysis does not violate copyright, or Google would have to shut down tomorrow.
  - notahacker 4y ago
    > Fine then. Say I went through Star Wars and counted how many blue pixels there were in each frame and published a list of the counts. Would you claim that list would be infringing? If you were instructed to counted blue pixels by somebody who used your lists for a SaaS product which could generate high fidelity representations of Death Stars, wookies and Harrison Ford (someone else probably counted the other colour pixels!) every time a customer of theirs typed in "Star Wars", quite possibly.
- johnthewise 4y ago
  The algorithm itself tries to learn to remove noise. This part of algorithm is undeniably a learning process, evidenced by the fact that it gets better along the training process. It just happens to be that, in order to remove noise, you need to have an understanding of the underlying thing(probability distribution), so in order to learn to remove noise, it needs to learn the images. I don't understand how a mathematical function by definition can't learn, since it's just a language we use to describe what's going on.

johnnymorgan 4y ago

This, training is passive knowledge. Basically this is saying if you look at it, and then learn to draw you owe Disney money. Disney would love that but it's ludicrous

Snild 4y ago

Machines are not people, and don't enjoy the same rights as us. (And I would not call it "learning", at least not in the same meaning of the word as for humans, but that's a more philosophical discussion.)

stale2002 4y ago
> Machines are not people, and don't enjoy the same rights as us Fortunately, a machine isn't doing anything on its own. It is the person who uses the machine that is doing something. And it is the human who is allowed to use tools to create transformative work.

visarga 4y ago

Nobody's arguing that near-duplications or trademark violations are not bad. But I think they want a broader power grab - to extend copyright to art styles. In other words to protect not just the expression but also the ideas themselves. They want something like patent + trademark + copyright all in one.

karpierz 4y ago

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".

fxtentacle 4y ago

The illegal part is actually SD selling the images which contain copyrighted material. If you were to paint Mickey and sell those paintings, Disney might sue you. Why should SD (the company) be allowed to do that?

andybak 4y ago
I don't understand this comment. SD aren't selling images of Mickey Mouse. Can you explain your point a bit more clearly?
sebzim4500 4y ago
Can you link to an example where Stability AI are selling images of Mickey Mouse? Sounds pretty hard to believe tbh, but you seem confident.
- fxtentacle 4y ago
  When you type in "Mickey Mouse" as a prompt and pay for their generation tool, they are in effect selling you a Mickey Mouse image.

notahacker 4y ago

You definitely need permission to make copies of people's art on your computers to use whilst learning though (copyright holders tend to be OK with or completely unaware of that, but it's a right they reserve) More to the point, this feels like the "have your cake and eat it" of legal arguments. If we're going to anthropomorphise AI to the extent we argue storing data in its data corpus without authorisation is just like allowing a human to look at your legally owned stuff, and reinterpreting that data in a [sometimes not particularly] lossy fashion as a set of weights and biases is just like a human learning, we should probably follow the "AI is just like a person" argument through to its logical conclusion and hold it legally responsible for not respecting copyright law if it ever produces outputs with trademarked images or uncanny resemblance to copyrighted material that aren't deemed "fair use", just like a person would be. (tbf, it's not too hard to imagine a world in which ChatGPT says "I'm sorry, but I am not able to write stories involving Mickey Mouse, because Mickey Mouse is the intellectual property of the Disney Corporation. Respecting intellectual property rights is very important", and I'm not sure that's an improvement...)

amelius 4y ago

> Since when does copyright require permission to learn from the copyrighted material? Ah, but you are anthropomorphizing an algorithm. What is happening inside the computer you call learning, but it is just some mathematical function. Other people might not agree that this is learning, and might call it "mixing" or "copying".

sebzim4500 4y ago
Fine then. Say I went through Star Wars and counted how many blue pixels there were in each frame and published a list of the counts. Would you claim that list would be infringing? Doing statistical analysis does not violate copyright, or Google would have to shut down tomorrow.
- notahacker 4y ago
  > Fine then. Say I went through Star Wars and counted how many blue pixels there were in each frame and published a list of the counts. Would you claim that list would be infringing? If you were instructed to counted blue pixels by somebody who used your lists for a SaaS product which could generate high fidelity representations of Death Stars, wookies and Harrison Ford (someone else probably counted the other colour pixels!) every time a customer of theirs typed in "Star Wars", quite possibly.
johnthewise 4y ago
The algorithm itself tries to learn to remove noise. This part of algorithm is undeniably a learning process, evidenced by the fact that it gets better along the training process. It just happens to be that, in order to remove noise, you need to have an understanding of the underlying thing(probability distribution), so in order to learn to remove noise, it needs to learn the images. I don't understand how a mathematical function by definition can't learn, since it's just a language we use to describe what's going on.