5 ms·

> In principle, an AI learning from a scientific textbook is no different than a human student doing the same. In principle human brains don't count as a copy

> In principle, an AI learning from a scientific textbook is no different than a human student doing the same. In principle human brains don't count as a copy of a copyrighted work, whereas the AI model that is capable of regurgitating the original-- I don't understand how that's not a copy[1] or derivative work which is a good indication the judges presented with these questions won't understand either. So, no, a brain with knowledge is a brain, and an AI model with knowledge is a potentially infringing copy or derivative work. [1] https://www.law.cornell.edu/definitions/uscode.php?width=840&height=800&iframe=true&def_id=17-USC-2024104691-364936160&term_occur=999&term_src= https://www.law.cornell.edu/definitions/uscode.php?width=840...

worstspotgain 2y ago
I can memorize and recite a poem, but in the context of a scientific textbook, neither I nor an AI will remember the actual words used, just key facts like E = mc^2. If you limit your critique to the AI's ability to recite a poem or lyrics verbatim, then you have a point. However, I used the example of a scientific textbook for that exact reason. To the extent that the AI only paraphrases or cites facts from, it's not a copy. The derivative work issue is more interesting. The AI can generate a derivative work, but won't do so without a prompt asking for one. If you ask for an Oz-like bedtime story, you should get an Anxious Lynx and not a Cowardly Lion. If you ask for a story with Oz characters, the result will be a derivative work, but it will include your prompt. The work was not in the AI, though the ability to make it was. In this sense, the AI is more like an instrument that can be used to violate copyright law, like an employee who has read Oz would be. It's hard to find a close inanimate analogue, but it's probably on the same shelf as search engines and typewriters.
- dahart 2y ago
  > neither I nor an AI will remember the actual words used […] To the extent that the AI only paraphrases Many AIs have in fact reproduced copyrighted work in large portions, it’s one of the reasons there are a bunch of lawsuits. > AI can generate a derivative work, but won’t do so without a prompt asking for one This is an odd and vague claim. Which AI are you talking about? And why do you claim there’s any generative AI in existence that won’t remix from it’s training examples? Their design from the ground up is to create derivative works. > AI is more like an instrument that can be used to violate copyright law If the training data wasn’t legally acquired, using it to train in the first place may violate the law. People are trying to change copyright’s interpretation, but pending the outcome of that it’s pretty clear that today’s AI may have violated copyright law before generating any output.
  - worstspotgain 2y ago
    > This is an odd and vague claim. Which AI are you talking about? Let's exclude maliciously crafted prompts. Given an "innocent" prompt that is not meant to elicit a copyright violation, what is the probability than an LLM will generate something that will both reproduce a copyrighted work and fail the fair use tests? It's obviously non-zero, but it's exceedingly small. I'd say it's around the same probability as Google Street View accidentally reproducing a copyrighted map that was displayed in a storefront window. A claim that Google has to pre-compensate all map publishers because of this possibility is just an attack on Street View per se. > And why do you claim there’s any generate AI in existence that won’t remix from it’s training examples? Their design from the ground up is to create derivative works. Just because something sounds like a derivative work doesn't mean it will infringe copyright or fail the fair use tests (see sibling threads.)
    - dahart 2y ago
      I reject the analogy to Street View, due to the fact that a sign in a window isn’t Street View’s primary use, it’s an accident (plus the store owner is responsible for ensuring permitted use of publicly visible signs and is culpable for copyright violations, and that photography of anything in public view is permitted by law - see the famous “photographer’s rights”), whereas with AI, statistically reproducing the training examples is the primary intent of it’s design. Reproducing derivative remix works is its function and nature through and through. Whether the output is actually recognizable as one of the individual inputs is a different question from whether the output is a derivative work. > Just because something sounds like a derivative work doesn’t mean it will infringe copyright or fail the fair use tests That’s true! Remixes where the source comes from multiple songs have a history of escaping copyright claims more often than derivative works of a single song. But that doesn’t change the fact that using copyrighted data without permission and without Fair Use exemption to train AI is currently illegal. And it doesn’t change the fact that when AI reproduces substantial portions of any given work, regardless of what the probability was, it is illegal. Claims that the probability of it are “exceedingly small” don’t stand up to the (now many!) actual cases filed in court demonstrating AI reproducing copyrighted works, and not because the chances were small, but because it simply breaks the law.
FeepingCreature 2y ago
So if I can regurgitate the textbook, my brain violates copyright? I think the more sensible take is that the reproduction by the LLM violates copyright, but not the act of training nor the LLM weights themselves.
- yencabulator 2y ago
  If your shared copies of your brain in that state, it seems like that would. Lots of LLM models can be downloaded. If you had memorized the textbook from a stash of torrented epubs, as a commercial activity with intent to monetize, that likely wouldn't go over well for you either. The "big AI" companies are trying to create a world of rules for thee not for me by claiming that when copyright infringement is large scale enough, it somehow becomes fair use.
  - worstspotgain 2y ago
    You're never going to get a textbook out of an LLM, or a chapter, or a page. Big-IP lawyers probably spend weeks engineering prompts to find enough stuff that comes out un-paraphrased. Poems and lyrics, maybe. The legal test for fair use considers "the amount and substantiality of the portion used in relation to the copyrighted work as a whole." In this case it's the portion regurgitated near-verbatim, since paraphrasing and summarizing is fair use. You can write Cliff Notes without the author's permission. LLMs are probably capable of failing the fair use tests for less than 0.1% of the stuff they're exposed to.
- dahart 2y ago
  Yes, that is what copyright law already says (that your brain reproducing a work is the violation, not your brain itself), which is why it took time for the lawsuits to build momentum - they had to find evidence of copyright violating reproduction. The part you’re missing here is that acquiring and using copyrighted data without permission also violates copyright. Just like copying and watching a movie you didn’t pay for or legally acquire violates copyright. It’s a worse offense to distribute a movie you don’t have the rights to, but copying one and consuming it without permission is an offense nonetheless. The act of training on data not legally acquired, therefore, certainly can be a copyright violation.
zmmmmm 2y ago
You make a good argument. The sad thing is that I could see this might put an end to open models while still allowing big tech to host them. Because creation of a derivative work isn't a problem itself, it only infringes at the point / time of distribution, which is the whole point of open models. But big tech could probably argue that their training of a private model doesn't infringe per se, and it is then the users who then create prompts that cause the regurgitation to occur that hit the tripwire of triggering copyright. Presumably, big tech can solve that, either with terms of use (telling the user it's their responsibility / liability) or YouTube style where a Content ID like system skewers anything that it can't prove isn't infringing before it gets to the user.
mr_toad 2y ago
A regression model - an equation for a line - is obviously not a copy of the data points that it is fit to. Yet, occasionally you might be able to get it to reproduce a point in the original data, but only if that point happens to be right on the line. It doesn’t mean that point is “in” the model, a regression model contains no points, just a slope and an intercept. An AI is nothing more than a large network of regression models, a very large set of equations. It is not a copy of anything.
- sirwhinesalot 2y ago
  A JPEG doesn't include a single pixel of the original image.
tensor 2y ago
Art forgeries are an example of a human brain regurgitating the original content. Someone reading an article and rewriting it with some differences is a derivative work. So yes, a brain with knowledge is potentially infringing a copy or a derivative work. Human brains are often treated specially in law, but I don't think they should be.

worstspotgain 2y ago

I can memorize and recite a poem, but in the context of a scientific textbook, neither I nor an AI will remember the actual words used, just key facts like E = mc^2. If you limit your critique to the AI's ability to recite a poem or lyrics verbatim, then you have a point. However, I used the example of a scientific textbook for that exact reason. To the extent that the AI only paraphrases or cites facts from, it's not a copy. The derivative work issue is more interesting. The AI can generate a derivative work, but won't do so without a prompt asking for one. If you ask for an Oz-like bedtime story, you should get an Anxious Lynx and not a Cowardly Lion. If you ask for a story with Oz characters, the result will be a derivative work, but it will include your prompt. The work was not in the AI, though the ability to make it was. In this sense, the AI is more like an instrument that can be used to violate copyright law, like an employee who has read Oz would be. It's hard to find a close inanimate analogue, but it's probably on the same shelf as search engines and typewriters.

dahart 2y ago
> neither I nor an AI will remember the actual words used […] To the extent that the AI only paraphrases Many AIs have in fact reproduced copyrighted work in large portions, it’s one of the reasons there are a bunch of lawsuits. > AI can generate a derivative work, but won’t do so without a prompt asking for one This is an odd and vague claim. Which AI are you talking about? And why do you claim there’s any generative AI in existence that won’t remix from it’s training examples? Their design from the ground up is to create derivative works. > AI is more like an instrument that can be used to violate copyright law If the training data wasn’t legally acquired, using it to train in the first place may violate the law. People are trying to change copyright’s interpretation, but pending the outcome of that it’s pretty clear that today’s AI may have violated copyright law before generating any output.
- worstspotgain 2y ago
  > This is an odd and vague claim. Which AI are you talking about? Let's exclude maliciously crafted prompts. Given an "innocent" prompt that is not meant to elicit a copyright violation, what is the probability than an LLM will generate something that will both reproduce a copyrighted work and fail the fair use tests? It's obviously non-zero, but it's exceedingly small. I'd say it's around the same probability as Google Street View accidentally reproducing a copyrighted map that was displayed in a storefront window. A claim that Google has to pre-compensate all map publishers because of this possibility is just an attack on Street View per se. > And why do you claim there’s any generate AI in existence that won’t remix from it’s training examples? Their design from the ground up is to create derivative works. Just because something sounds like a derivative work doesn't mean it will infringe copyright or fail the fair use tests (see sibling threads.)
  - dahart 2y ago
    I reject the analogy to Street View, due to the fact that a sign in a window isn’t Street View’s primary use, it’s an accident (plus the store owner is responsible for ensuring permitted use of publicly visible signs and is culpable for copyright violations, and that photography of anything in public view is permitted by law - see the famous “photographer’s rights”), whereas with AI, statistically reproducing the training examples is the primary intent of it’s design. Reproducing derivative remix works is its function and nature through and through. Whether the output is actually recognizable as one of the individual inputs is a different question from whether the output is a derivative work. > Just because something sounds like a derivative work doesn’t mean it will infringe copyright or fail the fair use tests That’s true! Remixes where the source comes from multiple songs have a history of escaping copyright claims more often than derivative works of a single song. But that doesn’t change the fact that using copyrighted data without permission and without Fair Use exemption to train AI is currently illegal. And it doesn’t change the fact that when AI reproduces substantial portions of any given work, regardless of what the probability was, it is illegal. Claims that the probability of it are “exceedingly small” don’t stand up to the (now many!) actual cases filed in court demonstrating AI reproducing copyrighted works, and not because the chances were small, but because it simply breaks the law.

FeepingCreature 2y ago

So if I can regurgitate the textbook, my brain violates copyright? I think the more sensible take is that the reproduction by the LLM violates copyright, but not the act of training nor the LLM weights themselves.

yencabulator 2y ago
If your shared copies of your brain in that state, it seems like that would. Lots of LLM models can be downloaded. If you had memorized the textbook from a stash of torrented epubs, as a commercial activity with intent to monetize, that likely wouldn't go over well for you either. The "big AI" companies are trying to create a world of rules for thee not for me by claiming that when copyright infringement is large scale enough, it somehow becomes fair use.
- worstspotgain 2y ago
  You're never going to get a textbook out of an LLM, or a chapter, or a page. Big-IP lawyers probably spend weeks engineering prompts to find enough stuff that comes out un-paraphrased. Poems and lyrics, maybe. The legal test for fair use considers "the amount and substantiality of the portion used in relation to the copyrighted work as a whole." In this case it's the portion regurgitated near-verbatim, since paraphrasing and summarizing is fair use. You can write Cliff Notes without the author's permission. LLMs are probably capable of failing the fair use tests for less than 0.1% of the stuff they're exposed to.
dahart 2y ago
Yes, that is what copyright law already says (that your brain reproducing a work is the violation, not your brain itself), which is why it took time for the lawsuits to build momentum - they had to find evidence of copyright violating reproduction. The part you’re missing here is that acquiring and using copyrighted data without permission also violates copyright. Just like copying and watching a movie you didn’t pay for or legally acquire violates copyright. It’s a worse offense to distribute a movie you don’t have the rights to, but copying one and consuming it without permission is an offense nonetheless. The act of training on data not legally acquired, therefore, certainly can be a copyright violation.

zmmmmm 2y ago

You make a good argument. The sad thing is that I could see this might put an end to open models while still allowing big tech to host them. Because creation of a derivative work isn't a problem itself, it only infringes at the point / time of distribution, which is the whole point of open models. But big tech could probably argue that their training of a private model doesn't infringe per se, and it is then the users who then create prompts that cause the regurgitation to occur that hit the tripwire of triggering copyright. Presumably, big tech can solve that, either with terms of use (telling the user it's their responsibility / liability) or YouTube style where a Content ID like system skewers anything that it can't prove isn't infringing before it gets to the user.

mr_toad 2y ago

A regression model - an equation for a line - is obviously not a copy of the data points that it is fit to. Yet, occasionally you might be able to get it to reproduce a point in the original data, but only if that point happens to be right on the line. It doesn’t mean that point is “in” the model, a regression model contains no points, just a slope and an intercept. An AI is nothing more than a large network of regression models, a very large set of equations. It is not a copy of anything.

sirwhinesalot 2y ago
A JPEG doesn't include a single pixel of the original image.

tensor 2y ago

Art forgeries are an example of a human brain regurgitating the original content. Someone reading an article and rewriting it with some differences is a derivative work. So yes, a brain with knowledge is potentially infringing a copy or a derivative work. Human brains are often treated specially in law, but I don't think they should be.