6 ms·

You seem to be conflating two things. Is the model itself a reproduction or is it capable or making reproductions? I think there is a strong argument for bot t

by shkkmo 3y ago

You seem to be conflating two things. Is the model itself a reproduction or is it capable or making reproductions? I think there is a strong argument for bot treating a ML model that is a mathematical amalgamation of a huge variety of material and is thus transformed as a new work. Now, if you use this system to make a reproduction of one of the works it was trained on, that doesn't "wash" the IP, that reproduction would still face all the same tests for infringement as any other work.

danaris 3y ago
Is an MP3 file itself a reproduction of the audio? Or is it merely capable of producing the reproduction? Now, it's trivially true that an MP3 file is not, itself, physically audio. It is physically bits in a digital storage medium. Without the right software and the right commands, there is no way to reproduce the recorded audio from the MP3 file. Now, that software is pretty ubiquitous today, so we have come to think of MP3 files as being the audio...but in a technical sense they do actually share a lot with the models behind LLMs and other generative ML projects. In another 10 years, the software to reproduce elements from the training set of an LLM may be as ubiquitous as MP3 players are today. (Well, that's actually pretty unlikely, given how many devices can play MP3s, but such software could well be baked into the OSes of major desktop and mobile operating systems by that time, anyway.) I think it is an oversimplification to call an LLM a "lossy compression" of the training data, but being an oversimplification doesn't mean there's not some very real truth to it—possibly enough that the law would find it to be a reasonable analogy.
- hn_acker 3y ago
  > Is an MP3 file itself a reproduction of the audio? Or is it merely capable of producing the reproduction? Regardless, when the MP3 file is played using a program following the MP3 standard, the result will be a sound identical or substantially similar to the sound that was encoded into the MP3 file. The purpose of the MP3 standard is to encode audio into a file (which will be called an MP3 file) which when played using a program following the MP3 standard produces a sound similar to the original audio. An AI model is meant to be created by aggregating multiple works, but the purpose of the model is not necessarily to produce something substantially similar to any work in the set. In my framing from the previous paragraph, the purpose of an AI model that produces images is to produce an image that can be opened using JPG/PNG/WEBP/AVIF image viewer, where the new image isn't necessarily supposed to be but can be similar to an existing work. If you were to train an AI model on literally only one work, then you would get something analogous in purpose to an MP3 file, since why would you use the model to produce something nonsimilar to the single work in the training set? Pasting what I wrote in a different comment: My argument is that similarity of new works (including AI model outputs) to actual works (not to non-existing works, existing styles, or theoretical aggregations of multiple works) is a prerequisite to infringement. From an article about the substantial similarity test in the US [1]: > To win a claim of copyright infringement in civil or criminal court, a plaintiff must show he or she owns a valid copyright, the defendant actually copied the work, and the level of copying amounts to misappropriation.[1][3] In order to get an infringing output, the user usually has to include a reference to an existing author or an existing work in the prompt. Sometimes that's not the case (which I've worried about with respect to Copilot), but in order to damn the model as a whole, you would have to establish that in over some percentage of cases the model produces infringing outputs for prompts which don't reference a particular author (whether individual or collective), a particular work, or a style strongly associated with a single or a few authors. [1] https://en.wikipedia.org/wiki/Substantial_similarity https://en.wikipedia.org/wiki/Substantial_similarity
  - danaris 3y ago
    > but the purpose of the model is not necessarily to produce something substantially similar to any work in the set. I think the point is that the purpose here is not relevant if the function includes the ability to do so. Even if the function also includes the ability to produce things that are, in their totality, substantially different from any given work in the training set.
    - hn_acker 3y ago
      > I think the point is that the purpose here is not relevant if the function includes the ability to do so. Even if the function also includes the ability to produce things that are, in their totality, substantially different from any given work in the training set. Before I get into the ability issue, keep in mind that regardless of the future of AI-related law, AI companies already have an incentive to prevent models from producing outputs substantially similar to existing copyrighted works, and the companies already do go out of their way to do the same. And I reiterate that generating something similar to a work in the training set will usually require the user of the model to reference a person (or group), a copyrightable work, or a style strongly associated with a certain person (or group). If the user writes a prompt like that then the user should bears the most responsibility if the result infringes on the copyright to a work in the training set. But more importantly, whether the output infringes doesn't matter until someone publishes it. That someone will be the user who submitted the prompt (and/or the model service provider if the service provider automatically publishes the output by default). Now I'll address the ability issue. > I think the point is that the purpose here is not relevant if the function includes the ability to do so. Even if the function also includes the ability to produce things that are, in their totality, substantially different from any given work in the training set. In the context of regulating AI, I reword `the ability to produce something substantially similar to a work in the training set` (using backticks because you didn't write those exact words) to `the ability to produce an output that infringes on copyright`, and I replace the main subject `model` with `program`. How do you write a law to regulate 1. a program (the program, not the users of the program) with `the ability to produce an output that infringes on copyright` without also regulating 2. a program with the ability to download a copyrighted work from a website that doesn't intend for users to download the work in such a way, while keeping in mind that users of this latter program can distribute the work in ways that infringe on copyright? For example, how do you regulate an image generation model without also regulating youtube-dl? You might point out, in the former case the model produces an infringing work while in the latter case no infringement happens until the user distributes the downloaded work or an infringing derivative. But in the latter case as well, there isn't infringement until the user distributes something relating to the output. Don't confuse my claim with "it's not illegal if no one finds out". If you make a movie adaptation out of someone else's book without permission, you aren't committing copyright infringement unless and until you actually distribute or show your movie adaptation to someone! My general point is that you can't treat the model's mere ability to infringe on copyright as justification to make the model or the model maker (rather than additionally or only a user of the model) liable for infringing outputs. To regulate a particular AI model in a way that preserves freedom of expression, you would have to demonstrate that the model consistently produces infringing outputs X percent of the time when given prompts which don't reference a person/group, a copyrightable work, or a characteristic/style strongly associated with a person/group. I don't know what X percent should be either, but surely 1% is not sufficient.
- shkkmo 3y ago
  An MP3 or any similar format, is a piece of data that exists to create an audio reproduction of a specific work. This single and clearly intentional purpose is quite different than the quite varied purpose of a trained generalist ML model. This "transformational" aspect and lack of intent make it much harder to argue that fair use doesn't apply to training the model. > but in a technical sense they do actually share a lot with the models behind LLMs and other generative ML projects How so? The "encoding", "decoding" and storage all seem very, very different. > such software could well be baked into the OSes of major desktop and mobile operating systems by that time That seems silly. However, if it were to come to pass then models trained with a clear intent for the main use to be reproducing the works it was trained, then fair use in training the model becomes harder to argue. I would argue that if the liability remains with the users of the model, then there will be demand for tools that check generated works for possibility of infringement. However, the more the intended use(s) of the trained model differs from that of the copyrighted training material, the stronger that argument for fair use becomes. I would argue that our approach to ML models should be more akin to a CD burner. It is something that could be used for piracy but if we don't put the responsibility for that on the manufacturers but rather on the users who choose to pirate. This makes sense because the fair use argument of a teacher using a ML model to generate content for their class is very different from a studio using a ML model to make content for a commerical work. If we make the model responsible for the possibility of making infringing content then we take away that flexibility.
  - danaris 3y ago
    The problem with likening taking other people's work as training data and reproducing it with an LLM to ripping a CD and sharing the files is the direction of power imbalance. In the music analogy, the "owner" of the original content whose copyright is being potentially infringed is a megacorporation making billions of dollars every quarter, with more lawyers than God. The person potentially infringing that copyright is you. It's me. It's random people everywhere. In the LLM case, the "owner" of the original content whose copyright is being potentially infringed is you. It's me. It's Neil Gaiman. It's Toby Fox. It's not just anybody, it's everybody. The entity potentially infringing that copyright is a megacorporation making billions of dollars every quarter, with more lawyers than God. I hope this makes clear why I think one is a problem and the other isn't.

by shkkmo 3y ago

danaris 3y ago

Is an MP3 file itself a reproduction of the audio? Or is it merely capable of producing the reproduction? Now, it's trivially true that an MP3 file is not, itself, physically audio. It is physically bits in a digital storage medium. Without the right software and the right commands, there is no way to reproduce the recorded audio from the MP3 file. Now, that software is pretty ubiquitous today, so we have come to think of MP3 files as being the audio...but in a technical sense they do actually share a lot with the models behind LLMs and other generative ML projects. In another 10 years, the software to reproduce elements from the training set of an LLM may be as ubiquitous as MP3 players are today. (Well, that's actually pretty unlikely, given how many devices can play MP3s, but such software could well be baked into the OSes of major desktop and mobile operating systems by that time, anyway.) I think it is an oversimplification to call an LLM a "lossy compression" of the training data, but being an oversimplification doesn't mean there's not some very real truth to it—possibly enough that the law would find it to be a reasonable analogy.

hn_acker 3y ago
> Is an MP3 file itself a reproduction of the audio? Or is it merely capable of producing the reproduction? Regardless, when the MP3 file is played using a program following the MP3 standard, the result will be a sound identical or substantially similar to the sound that was encoded into the MP3 file. The purpose of the MP3 standard is to encode audio into a file (which will be called an MP3 file) which when played using a program following the MP3 standard produces a sound similar to the original audio. An AI model is meant to be created by aggregating multiple works, but the purpose of the model is not necessarily to produce something substantially similar to any work in the set. In my framing from the previous paragraph, the purpose of an AI model that produces images is to produce an image that can be opened using JPG/PNG/WEBP/AVIF image viewer, where the new image isn't necessarily supposed to be but can be similar to an existing work. If you were to train an AI model on literally only one work, then you would get something analogous in purpose to an MP3 file, since why would you use the model to produce something nonsimilar to the single work in the training set? Pasting what I wrote in a different comment: My argument is that similarity of new works (including AI model outputs) to actual works (not to non-existing works, existing styles, or theoretical aggregations of multiple works) is a prerequisite to infringement. From an article about the substantial similarity test in the US [1]: > To win a claim of copyright infringement in civil or criminal court, a plaintiff must show he or she owns a valid copyright, the defendant actually copied the work, and the level of copying amounts to misappropriation.[1][3] In order to get an infringing output, the user usually has to include a reference to an existing author or an existing work in the prompt. Sometimes that's not the case (which I've worried about with respect to Copilot), but in order to damn the model as a whole, you would have to establish that in over some percentage of cases the model produces infringing outputs for prompts which don't reference a particular author (whether individual or collective), a particular work, or a style strongly associated with a single or a few authors. [1] https://en.wikipedia.org/wiki/Substantial_similarity https://en.wikipedia.org/wiki/Substantial_similarity
- danaris 3y ago
  > but the purpose of the model is not necessarily to produce something substantially similar to any work in the set. I think the point is that the purpose here is not relevant if the function includes the ability to do so. Even if the function also includes the ability to produce things that are, in their totality, substantially different from any given work in the training set.
  - hn_acker 3y ago
    > I think the point is that the purpose here is not relevant if the function includes the ability to do so. Even if the function also includes the ability to produce things that are, in their totality, substantially different from any given work in the training set. Before I get into the ability issue, keep in mind that regardless of the future of AI-related law, AI companies already have an incentive to prevent models from producing outputs substantially similar to existing copyrighted works, and the companies already do go out of their way to do the same. And I reiterate that generating something similar to a work in the training set will usually require the user of the model to reference a person (or group), a copyrightable work, or a style strongly associated with a certain person (or group). If the user writes a prompt like that then the user should bears the most responsibility if the result infringes on the copyright to a work in the training set. But more importantly, whether the output infringes doesn't matter until someone publishes it. That someone will be the user who submitted the prompt (and/or the model service provider if the service provider automatically publishes the output by default). Now I'll address the ability issue. > I think the point is that the purpose here is not relevant if the function includes the ability to do so. Even if the function also includes the ability to produce things that are, in their totality, substantially different from any given work in the training set. In the context of regulating AI, I reword `the ability to produce something substantially similar to a work in the training set` (using backticks because you didn't write those exact words) to `the ability to produce an output that infringes on copyright`, and I replace the main subject `model` with `program`. How do you write a law to regulate 1. a program (the program, not the users of the program) with `the ability to produce an output that infringes on copyright` without also regulating 2. a program with the ability to download a copyrighted work from a website that doesn't intend for users to download the work in such a way, while keeping in mind that users of this latter program can distribute the work in ways that infringe on copyright? For example, how do you regulate an image generation model without also regulating youtube-dl? You might point out, in the former case the model produces an infringing work while in the latter case no infringement happens until the user distributes the downloaded work or an infringing derivative. But in the latter case as well, there isn't infringement until the user distributes something relating to the output. Don't confuse my claim with "it's not illegal if no one finds out". If you make a movie adaptation out of someone else's book without permission, you aren't committing copyright infringement unless and until you actually distribute or show your movie adaptation to someone! My general point is that you can't treat the model's mere ability to infringe on copyright as justification to make the model or the model maker (rather than additionally or only a user of the model) liable for infringing outputs. To regulate a particular AI model in a way that preserves freedom of expression, you would have to demonstrate that the model consistently produces infringing outputs X percent of the time when given prompts which don't reference a person/group, a copyrightable work, or a characteristic/style strongly associated with a person/group. I don't know what X percent should be either, but surely 1% is not sufficient.
shkkmo 3y ago
An MP3 or any similar format, is a piece of data that exists to create an audio reproduction of a specific work. This single and clearly intentional purpose is quite different than the quite varied purpose of a trained generalist ML model. This "transformational" aspect and lack of intent make it much harder to argue that fair use doesn't apply to training the model. > but in a technical sense they do actually share a lot with the models behind LLMs and other generative ML projects How so? The "encoding", "decoding" and storage all seem very, very different. > such software could well be baked into the OSes of major desktop and mobile operating systems by that time That seems silly. However, if it were to come to pass then models trained with a clear intent for the main use to be reproducing the works it was trained, then fair use in training the model becomes harder to argue. I would argue that if the liability remains with the users of the model, then there will be demand for tools that check generated works for possibility of infringement. However, the more the intended use(s) of the trained model differs from that of the copyrighted training material, the stronger that argument for fair use becomes. I would argue that our approach to ML models should be more akin to a CD burner. It is something that could be used for piracy but if we don't put the responsibility for that on the manufacturers but rather on the users who choose to pirate. This makes sense because the fair use argument of a teacher using a ML model to generate content for their class is very different from a studio using a ML model to make content for a commerical work. If we make the model responsible for the possibility of making infringing content then we take away that flexibility.
- danaris 3y ago
  The problem with likening taking other people's work as training data and reproducing it with an LLM to ripping a CD and sharing the files is the direction of power imbalance. In the music analogy, the "owner" of the original content whose copyright is being potentially infringed is a megacorporation making billions of dollars every quarter, with more lawyers than God. The person potentially infringing that copyright is you. It's me. It's random people everywhere. In the LLM case, the "owner" of the original content whose copyright is being potentially infringed is you. It's me. It's Neil Gaiman. It's Toby Fox. It's not just anybody, it's everybody. The entity potentially infringing that copyright is a megacorporation making billions of dollars every quarter, with more lawyers than God. I hope this makes clear why I think one is a problem and the other isn't.