4 ms·

This comment is a bit hard to unpick. It's not clear that training violates copyright. However plagarism (created by any means AI or not) is already covered by

This comment is a bit hard to unpick. It's not clear that training violates copyright. However plagarism (created by any means AI or not) is already covered by existing IP law. (The difficult part is telling whether AI generated stuff infringes on the copyright of stuff in the training data. And the fact the onus is on the end-user of AI tools sucks...) However - if you're arguing that copyright law should be extended to prevent use in training - that's a much tougher sell and one where I'm not 100% sure of my position.

tpxl 4y ago
> It's not clear that training violates copyright What is perfectly clear are two things: 1. Github Copilot will sometimes produce non-trivial inputs it was trained on as outputs 2. Those outputs will not conform to the license That is, in my mind, a clear violation. This could easily be solved by making training (or transformation or whatever) explicit in the license. Edit: I single out Github copilot because such examples were found. I'm not sure you can find such examples for things such as Stable Diffusion or DALL-E, but the operating principle is the same.
- wongarsu 4y ago
  > I'm not sure you can find such examples for things such as Stable Diffusion or DALL-E DALL-E 1 will sometimes replicate watermarks (without being promted for it [1]). I'm not sure anyone has ever shown a 1-1 replication like what's shown with Copilot, but it's easy to get image models to create clearly derivative works (just ask for the Mona Lisa, or even a well-known modern work). 1: https://ibb.co/gz5RDkB https://ibb.co/gz5RDkB
ginko 4y ago
>It's not clear that training violates copyright. Training a model essentially creates a derivative work from the input data. GPL licensed code requires keeping the same license for derivative work. BSD-style licenses at the very least require retaining the copyright statement. It feels pretty clear to me that the trained model and code generated from it would have to adhere to the original license. In a way ML is a bit like lossy compression. Would you think JPEG encoding and decoding an image strips the result of its original copyright?
- andybak 4y ago
  I agree that this is one likely interpretation but it's not the only one and it's yet to be tested in the courts.
- wongarsu 4y ago
  There is a point where a derivative work is so transformative that the original copyright no longer matters. In the US you would argue via fair use, where the precedent is "The more transformative the new work, the less will be the significance of other factors, like commercialism, that may weigh against a finding of fair use" [1], and other jurisdictions have equivalent ways to excempt works that are different or transformative enough. So the big question is how transformativ a given model is. And at least for language models I'm not sure we even understand Transformer models enough to give a qualified answer. 1: https://en.wikipedia.org/wiki/Transformative_use https://en.wikipedia.org/wiki/Transformative_use
  - staindk 4y ago
    I think this is true, but currently you can coax e.g. stable diffusion to generate art in the style of a specific person, and if you do things right, it will even have a smudge-y version of their monogram/signature in the bottom right. Pretty clear indication that something is not yet 'transformative enough', IMO. Same happens with Copilot and such - you can get it to spit out tens of lines of code verbatim pulled from someone else's project, with comments and all. These kinds of things just don't sit right with me.
    - andybak 4y ago
      > I think this is true, but currently you can coax e.g. stable diffusion to generate art in the style of a specific person, and if you do things right, it will even have a smudge-y version of their monogram/signature in the bottom right. You're conflating two things here: 1. SD can generate art in the style of a specific person 2. It also adds signatures to some pieces These are independent things. Sometimes it does 1 without 2 and sometimes 2 without 1. And the signature can appear on the wrong artist - or the signature is a weird mix of different signatures. It's just that it's been trained on "paintings with signatures" so it includes them. It's not the damning evidence some claim it to be.

tpxl 4y ago

> It's not clear that training violates copyright What is perfectly clear are two things: 1. Github Copilot will sometimes produce non-trivial inputs it was trained on as outputs 2. Those outputs will not conform to the license That is, in my mind, a clear violation. This could easily be solved by making training (or transformation or whatever) explicit in the license. Edit: I single out Github copilot because such examples were found. I'm not sure you can find such examples for things such as Stable Diffusion or DALL-E, but the operating principle is the same.

wongarsu 4y ago
> I'm not sure you can find such examples for things such as Stable Diffusion or DALL-E DALL-E 1 will sometimes replicate watermarks (without being promted for it [1]). I'm not sure anyone has ever shown a 1-1 replication like what's shown with Copilot, but it's easy to get image models to create clearly derivative works (just ask for the Mona Lisa, or even a well-known modern work). 1: https://ibb.co/gz5RDkB https://ibb.co/gz5RDkB

ginko 4y ago

>It's not clear that training violates copyright. Training a model essentially creates a derivative work from the input data. GPL licensed code requires keeping the same license for derivative work. BSD-style licenses at the very least require retaining the copyright statement. It feels pretty clear to me that the trained model and code generated from it would have to adhere to the original license. In a way ML is a bit like lossy compression. Would you think JPEG encoding and decoding an image strips the result of its original copyright?

andybak 4y ago
I agree that this is one likely interpretation but it's not the only one and it's yet to be tested in the courts.
wongarsu 4y ago
There is a point where a derivative work is so transformative that the original copyright no longer matters. In the US you would argue via fair use, where the precedent is "The more transformative the new work, the less will be the significance of other factors, like commercialism, that may weigh against a finding of fair use" [1], and other jurisdictions have equivalent ways to excempt works that are different or transformative enough. So the big question is how transformativ a given model is. And at least for language models I'm not sure we even understand Transformer models enough to give a qualified answer. 1: https://en.wikipedia.org/wiki/Transformative_use https://en.wikipedia.org/wiki/Transformative_use
- staindk 4y ago
  I think this is true, but currently you can coax e.g. stable diffusion to generate art in the style of a specific person, and if you do things right, it will even have a smudge-y version of their monogram/signature in the bottom right. Pretty clear indication that something is not yet 'transformative enough', IMO. Same happens with Copilot and such - you can get it to spit out tens of lines of code verbatim pulled from someone else's project, with comments and all. These kinds of things just don't sit right with me.
  - andybak 4y ago
    > I think this is true, but currently you can coax e.g. stable diffusion to generate art in the style of a specific person, and if you do things right, it will even have a smudge-y version of their monogram/signature in the bottom right. You're conflating two things here: 1. SD can generate art in the style of a specific person 2. It also adds signatures to some pieces These are independent things. Sometimes it does 1 without 2 and sometimes 2 without 1. And the signature can appear on the wrong artist - or the signature is a weird mix of different signatures. It's just that it's been trained on "paintings with signatures" so it includes them. It's not the damning evidence some claim it to be.