3 ms·

I don't think this legal analysis is correct. The most actionable copyright problem isn't anything that OpenAI's models do in generating content (whether you co

I don't think this legal analysis is correct. The most actionable copyright problem isn't anything that OpenAI's models do in generating content (whether you consider it transformative or otherwise). The problem is that in creating the model, OpenAI's developers had to copy (into GPU memory, at least) creative works protected by copyright. That copy was not fair use under any of the long-established criteria. The closest precedent is probably the 9th Circuit's ruling in MAI Systems[1]. Given the Supreme Court's recent ruling against the estate of Andy Warhol, I'd bet on them endorsing the key holdings of that precedent in a hypothetical Silverman v. OpenAI. [1] https://en.wikipedia.org/wiki/MAI_Systems_Corp._v._Peak_Computer,_Inc https://en.wikipedia.org/wiki/MAI_Systems_Corp._v._Peak_Comp....

andybak 3y ago
Is that the right sense of "copying" though? I thought "distributing" was regarded as a better way to think about the law in this area.
- twoodfin 3y ago
  Take a look at the circumstances of MAI Systems. The infringement found was a third-party maintenance engineer copying the user’s licensed software within the user’s own machine for unlicensed use. Congress later carved out this particular case as fair use. There is no such carve out for copying novels to train LLMs.
BaseballPhysics 3y ago
Beat me to it. That said, I'm not sure the caselaw is that clear. For example, in Authors Guild v Google, the courts found Google's use of copyrighted works to build a search index as fair use: https://en.m.wikipedia.org/wiki/Authors_Guild,_Inc._v._Google,_Inc https://en.m.wikipedia.org/wiki/Authors_Guild,_Inc._v._Googl.... In particular, of the four factor test (https://en.m.wikipedia.org/wiki/Fair_use#U.S._fair_use_factors https://en.m.wikipedia.org/wiki/Fair_use#U.S._fair_use_facto...), in Authors Guild the judge found that the authors didn't experience economic damage from the copying. In the case of LLMs, can the affected artists argue for economic damages? That's not obvious to me.
- twoodfin 3y ago
  Google Books was a wildly distinct use, though. There’s plenty of support in copyright law for fair use claims around indexing and excerpting for such purposes. As for economic damages, that seems easy: One obvious use of the model is to create derivative works that could economically compete with the original creators’ works. Simpler examples make this clear: I love Disney animated films. I love them so much I build a model & train it on the complete corpus of Disney animated films. My model is a success, and can produce new films which more or less echo the style, characters, themes of the Disney films. This is not what Google was doing. The creative work was the well-organized, efficient index, not new books or a program to write them.
ejz 3y ago
I don't think this case is about GPU copying, since that's not what's alleged in the complaint. Mostly, they are alleging that Silverman's book was in a training set that contained an unauthorized copy. I wrote this post in a way that was meant to be brief, clear, and honestly a little snarky. If I was writing a brief, I'd have gone through a more detailed analysis, but I think fails for a number of reasons, including fair use for the copying, transformative use for the AI system, and frankly a lack of damages because the LLM isn't distributing copies of her work (which is one reason they are trying to establish Silverman as a lead plaintiff in a class action suit). I glossed over a lot of this discussion to keep the post brief. But of course, this post is not legal advice, it's just a quick (and hopefully fun!) blog. I am much more knowledgeable about patent law than copyright law, by the way, and acknowledge any flaws in my post. Most importantly, I wanted to highlight what I view as a real difference between what copyright holders want from copyright and the policy goals of copyright. What I believe they want is a patent, but what they have is a copyright. They are intended to achieve very different policy goals and as such have different structures.
- twoodfin 3y ago
  The fundamental problem: a) Training the model involved making a non-transient copy of Silverman’s work (e.g. in an encoded form into the GPU). b) That copy was not a fair use. The purpose of the copy was to train an ML model to generate similar content. OpenAI offers commercial access to the resulting generation mechanism, and indisputably some of what people are paying for is content generation to compete with creators like Silverman. If OpenAI were only training models for research purposes they’d have a much stronger case for fair use.

andybak 3y ago

Is that the right sense of "copying" though? I thought "distributing" was regarded as a better way to think about the law in this area.

twoodfin 3y ago
Take a look at the circumstances of MAI Systems. The infringement found was a third-party maintenance engineer copying the user’s licensed software within the user’s own machine for unlicensed use. Congress later carved out this particular case as fair use. There is no such carve out for copying novels to train LLMs.

BaseballPhysics 3y ago

Beat me to it. That said, I'm not sure the caselaw is that clear. For example, in Authors Guild v Google, the courts found Google's use of copyrighted works to build a search index as fair use: https://en.m.wikipedia.org/wiki/Authors_Guild,_Inc._v._Google,_Inc https://en.m.wikipedia.org/wiki/Authors_Guild,_Inc._v._Googl.... In particular, of the four factor test (https://en.m.wikipedia.org/wiki/Fair_use#U.S._fair_use_factors https://en.m.wikipedia.org/wiki/Fair_use#U.S._fair_use_facto...), in Authors Guild the judge found that the authors didn't experience economic damage from the copying. In the case of LLMs, can the affected artists argue for economic damages? That's not obvious to me.

twoodfin 3y ago
Google Books was a wildly distinct use, though. There’s plenty of support in copyright law for fair use claims around indexing and excerpting for such purposes. As for economic damages, that seems easy: One obvious use of the model is to create derivative works that could economically compete with the original creators’ works. Simpler examples make this clear: I love Disney animated films. I love them so much I build a model & train it on the complete corpus of Disney animated films. My model is a success, and can produce new films which more or less echo the style, characters, themes of the Disney films. This is not what Google was doing. The creative work was the well-organized, efficient index, not new books or a program to write them.

ejz 3y ago

I don't think this case is about GPU copying, since that's not what's alleged in the complaint. Mostly, they are alleging that Silverman's book was in a training set that contained an unauthorized copy. I wrote this post in a way that was meant to be brief, clear, and honestly a little snarky. If I was writing a brief, I'd have gone through a more detailed analysis, but I think fails for a number of reasons, including fair use for the copying, transformative use for the AI system, and frankly a lack of damages because the LLM isn't distributing copies of her work (which is one reason they are trying to establish Silverman as a lead plaintiff in a class action suit). I glossed over a lot of this discussion to keep the post brief. But of course, this post is not legal advice, it's just a quick (and hopefully fun!) blog. I am much more knowledgeable about patent law than copyright law, by the way, and acknowledge any flaws in my post. Most importantly, I wanted to highlight what I view as a real difference between what copyright holders want from copyright and the policy goals of copyright. What I believe they want is a patent, but what they have is a copyright. They are intended to achieve very different policy goals and as such have different structures.

twoodfin 3y ago
The fundamental problem: a) Training the model involved making a non-transient copy of Silverman’s work (e.g. in an encoded form into the GPU). b) That copy was not a fair use. The purpose of the copy was to train an ML model to generate similar content. OpenAI offers commercial access to the resulting generation mechanism, and indisputably some of what people are paying for is content generation to compete with creators like Silverman. If OpenAI were only training models for research purposes they’d have a much stronger case for fair use.