4 ms·

> Lossy or not, the training data provides value. If we ignore the issue of machine learning for now; It's not the job of copyright to prevent people extractin

by phire 3y ago

> Lossy or not, the training data provides value. If we ignore the issue of machine learning for now; It's not the job of copyright to prevent people extracting value from a copyrighted work. If it was, then it would be possible for copyright holders to launch lawsuits that block entities from using the knowledge that was published in copyrighted reference material. Or the rights holder of a cookbook would be able to block people from making the recipe. We have other sections of law that provide protection along these lines. Patents give their holders a monopoly to extract value from the given invention, a much stronger protection that copyright law. But in exchange Patents are limited to 21 years, much be publicly documented and only certain types of things are covered. Trade secret laws can be used to protect recipes, formulas and other processes, but only as long as the holder makes reasonable efforts to keep it secret. The owner can't have it both ways, have the IP publicly known and protected by trade secret law. The only purpose of copyright is to give the holder a monopoly over the reproduction of a work. The definition of reproduction might be quite wide these days: A performance of a work is a reproduction, a cover of a song is a reproduction, distribution is reproduction in the modern age of computers... etc, but the limit of copyright law is reproduction. Coming back to machine learning, there is a somewhat open question if training counts as a form of reproduction (well, outside of Japan). But we can't use your proposed "extraction of value" metric as a way to decide that. Personally, I would argue that training a machine learning model is roughly equivalent to a human brain consuming copyrighted works, and should be treated the same in law. The fact that a machine learning model is (sometimes) capable of recreating a copyrighted work later shouldn't be held against them, as a human brain is also fully capable of recreating copyrighted works from memory. From a legal perspective, recreating a copyrighted work from memory will not save a human from a copyright infringement lawsuit, it's simply not a defence. The copyright infringement happens when the work is recreated.

belorn 3y ago
The purpose of copyright is to create moral and economical incentives to authors to create new copyrightable works, but giving those authors a time limited state granted monopoly. Reproduction is only one aspect to this. As an example, applying a song to a video require addition permissions even if the party has permissions for reproduction. The owners to a record can also disallow the use of a song in a political event based on the context of morality, even if the politician has bought permissions to play the song in public. I would personally also argue that machine learning model is roughly equivalent to a compression algorithm. Converting a 4k video to a 420p video is just as lossy as feeding a learning model from a 4k video and ask it to reproduce the 420p video. It has nothing in common with how a human brain is consuming content or learns information. No person can produce a 420p video just by consuming a 4k video, nor can any machine learning model gain the emotional constructs and social contexts that human brains get from learning.
- phire 3y ago
  Your examples are not inherent rights that copyright law explictly grants to rights holders. They are clever side effect of how the holder licenses out their monopoly on reproduction. Holders rarely grant unrestricted reproduction rights to anyone. Reproduction rights licenses always come with a bunch of explicit restrictions, for example: "You may reproduce this novel, in print, unmodified, only for retail sale, in North America, on this quality of paper, for the next 5 years" and so on. The party has the license to reproduce the song as a standalone audio recording, but attaching it to a video and reproducing the combined work isn't covered and the party must enter into negotiations with the rights holder for a new license. Such licenses often only grant the rights to reproduce it with that exact video and not a different one later, which allows the rights holder to gain control over which videos their song is attached to. Same thing with holding morality over political events. The rights holder was careful to add a bunch of restrictions to that public performance license they sell. Sure, the politician might have bought a licence, but they forgot to check the small print that blocks their type of event from actually using it. ----- Machine learning is kind of like compression, yes... It can be a useful analogy at times. But it is absolutely nothing like lossy video compression. It's not compressing a single file or object. The only way you could train it on a 4k video and get a 420p video out is if that model was extremely over-fitted. The resulting model would likely be bigger than a 420p h264 video file and useless for anything else. The way that machine learning is like a compression is that it find common patterns across it's entire training set and merges them in very lossy ways. And it's actually very much like how a human brain works. Your brain doesn't start from scratch for every single human face you recognise. Instead, your brain has built up a generic understanding of the average human face, grouping by clusters of features. Then to remember a given human face it just remembers which cluster of features it's close to and then how it differs... Which is a form of lossy compression. > No person can produce a 420p video just by consuming a 4k video But many people do remember entire songs, complete with lyrics and music. And people with musical skills can (and often do) reproduce that song from memory as a cover... Which is copyright infringement if preformed publicly or otherwise distributed. > nor can any machine learning model gain the emotional constructs and social contexts that human brains get from learning. There are many things which large LLMs like chatgpt are absolutely incapable of doing. People do over hype their capabilities. But in my experiments, chatgpt is actually quite good at tasks that require interpreting emotions and social contexts. Does it actually understand these emotions and social contexts? shrug. But if it doesn't actually understand that just proves that true understanding isn't actually needed to preform useful tasks in those areas.
ChuckkH 3y ago
Copyright law (EU, US, UK, international under the Berne Convention) covers reproduction, distribution and exhibition. That's exactly what those FBI warnings on VHS movies used to say. Distribution and exhibition are prohibited along with reproduction. In all the cases I mentioned, the only legal way to make any exception to that is if the copying does not harm the interests of the author or reduce the market value of the work. These are the actual laws. "training a machine learning model is roughly equivalent to a human brain consuming copyrighted works" A few clear differences: 1. The person "training" a machine learning model doesn't even need to view the work. They copy a file. They do not study or learn from it. 2. A human brain doesn't rely on a 100% verbatim digital copy of the work. To the extent that a brain "makes a copy" of what it observes, it is impossible for it not to. 3. Copyright law (almost everywhere) explicitly applies to making digital copies of binary files of a work (without which it is not possible to "train" a model using the work). Nowhere does it ever apply to a human brain when a person looks at the work. Not all the things you mentioned are considered "reproduction". A cover of a song is a derivative work, and requires compensation. Showing a movie is exhibition, and is explicitly addressed in copyright laws. These things are not just considered "some form of reproduction". The laws actually exist and are easy to find and read.

by phire 3y ago

belorn 3y ago

The purpose of copyright is to create moral and economical incentives to authors to create new copyrightable works, but giving those authors a time limited state granted monopoly. Reproduction is only one aspect to this. As an example, applying a song to a video require addition permissions even if the party has permissions for reproduction. The owners to a record can also disallow the use of a song in a political event based on the context of morality, even if the politician has bought permissions to play the song in public. I would personally also argue that machine learning model is roughly equivalent to a compression algorithm. Converting a 4k video to a 420p video is just as lossy as feeding a learning model from a 4k video and ask it to reproduce the 420p video. It has nothing in common with how a human brain is consuming content or learns information. No person can produce a 420p video just by consuming a 4k video, nor can any machine learning model gain the emotional constructs and social contexts that human brains get from learning.

phire 3y ago
Your examples are not inherent rights that copyright law explictly grants to rights holders. They are clever side effect of how the holder licenses out their monopoly on reproduction. Holders rarely grant unrestricted reproduction rights to anyone. Reproduction rights licenses always come with a bunch of explicit restrictions, for example: "You may reproduce this novel, in print, unmodified, only for retail sale, in North America, on this quality of paper, for the next 5 years" and so on. The party has the license to reproduce the song as a standalone audio recording, but attaching it to a video and reproducing the combined work isn't covered and the party must enter into negotiations with the rights holder for a new license. Such licenses often only grant the rights to reproduce it with that exact video and not a different one later, which allows the rights holder to gain control over which videos their song is attached to. Same thing with holding morality over political events. The rights holder was careful to add a bunch of restrictions to that public performance license they sell. Sure, the politician might have bought a licence, but they forgot to check the small print that blocks their type of event from actually using it. ----- Machine learning is kind of like compression, yes... It can be a useful analogy at times. But it is absolutely nothing like lossy video compression. It's not compressing a single file or object. The only way you could train it on a 4k video and get a 420p video out is if that model was extremely over-fitted. The resulting model would likely be bigger than a 420p h264 video file and useless for anything else. The way that machine learning is like a compression is that it find common patterns across it's entire training set and merges them in very lossy ways. And it's actually very much like how a human brain works. Your brain doesn't start from scratch for every single human face you recognise. Instead, your brain has built up a generic understanding of the average human face, grouping by clusters of features. Then to remember a given human face it just remembers which cluster of features it's close to and then how it differs... Which is a form of lossy compression. > No person can produce a 420p video just by consuming a 4k video But many people do remember entire songs, complete with lyrics and music. And people with musical skills can (and often do) reproduce that song from memory as a cover... Which is copyright infringement if preformed publicly or otherwise distributed. > nor can any machine learning model gain the emotional constructs and social contexts that human brains get from learning. There are many things which large LLMs like chatgpt are absolutely incapable of doing. People do over hype their capabilities. But in my experiments, chatgpt is actually quite good at tasks that require interpreting emotions and social contexts. Does it actually understand these emotions and social contexts? shrug. But if it doesn't actually understand that just proves that true understanding isn't actually needed to preform useful tasks in those areas.

ChuckkH 3y ago

Copyright law (EU, US, UK, international under the Berne Convention) covers reproduction, distribution and exhibition. That's exactly what those FBI warnings on VHS movies used to say. Distribution and exhibition are prohibited along with reproduction. In all the cases I mentioned, the only legal way to make any exception to that is if the copying does not harm the interests of the author or reduce the market value of the work. These are the actual laws. "training a machine learning model is roughly equivalent to a human brain consuming copyrighted works" A few clear differences: 1. The person "training" a machine learning model doesn't even need to view the work. They copy a file. They do not study or learn from it. 2. A human brain doesn't rely on a 100% verbatim digital copy of the work. To the extent that a brain "makes a copy" of what it observes, it is impossible for it not to. 3. Copyright law (almost everywhere) explicitly applies to making digital copies of binary files of a work (without which it is not possible to "train" a model using the work). Nowhere does it ever apply to a human brain when a person looks at the work. Not all the things you mentioned are considered "reproduction". A cover of a song is a derivative work, and requires compensation. Showing a movie is exhibition, and is explicitly addressed in copyright laws. These things are not just considered "some form of reproduction". The laws actually exist and are easy to find and read.