4 ms·

There is existing case law behind clean room reverse engineering (the "Chinese Wall" technique). Using generative AIs to scale up this process is inevitable. "

There is existing case law behind clean room reverse engineering (the "Chinese Wall" technique). Using generative AIs to scale up this process is inevitable. "Clean Room Defeats Software Infringement Claim in U.S. Federal Court" http://hoviblog.blogspot.de/2008/10/clean-room-defeats-software.html http://hoviblog.blogspot.de/2008/10/clean-room-defeats-softw... "Chinese Wall" https://en.wikipedia.org/wiki/Chinese_wall https://en.wikipedia.org/wiki/Chinese_wall

blibble 4y ago
the AI in this situation is no different than gzipping a copyrighted file, ungzipping it and then trying to claim the output is no longer copyrighted
- politician 4y ago
  There have to be two AIs, one examines the original and produces a description while the second examines the description and produces a new work. Neither the description nor the new work are assigned copyright because they were produced by an AI (see [1]). However, if subsequently a human then manipulates the new work, the derived work is automatically protected by copyright assigned to that person (see [1, 2]). Furthermore, the new author has a defense against a copyright infringement claim by the owner of the original based on existing case law shared previously. This has nothing to do with gzip. [1] https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence https://www.federalregister.gov/documents/2023/03/16/2023-05... [2] https://en.wikipedia.org/wiki/Berne_Convention https://en.wikipedia.org/wiki/Berne_Convention
  - blibble 4y ago
    no matter how many AIs you layer: it's still just a lossy compression function jpeg'ing frames of disney animation doesn't remove copyright (even if you do it twice)
    - CuriouslyC 4y ago
      Saying a model is like jpeg compression is like saying a van is like an elephant, in that maybe if you squint so hard you're almost blind it's kinda true.
    - politician 4y ago
      What matters here for copyright are the laws, not the math.
- CuriouslyC 4y ago
  It's quite different, because GZIP produces a lossy copy of an image, whereas models produce a distinct image that shares many aspects of style and composition. In one case it's obvious the product is the original image, in the other it's a derivative that may or may not be sufficiently different to be protected under fair use. I can write an algorithm that just generates random noise in the dimensions of art, and if I run it long enough it'll output things that are "close enough" to copyright works. There's no argument for that program being copyright infringement, and that holds for models as well.
  - blibble 4y ago
    > In one case it's obvious the product is the original image, in the other it's a derivative that may or may not be sufficiently different to be protected under fair use. right, so if it's a completely original work let's clear out the training set and let's see if it can do it without it no? so the output is a product of the input... a derivative work and if you run it again it produces exactly the same thing? (sans artificial random injection) sounds like a lossy compression function to me!
    - CuriouslyC 4y ago
      I'd like to see a human make art after you lobotomize them.
kmeisthax 4y ago
Clean room reverse engineering is a thing because software copyright was a mistake and the judicial system understands it was a mistake. In software, it is regular and common for programmers to use licensed software libraries with defined interfaces. In writing or art, the notion of a "compatible interface" isn't really there. If you wrote a short fan story with Marvel characters and you want to liberate it from Disney's ownership, you have to redesign all the characters; you can't just argue that you need a superhero with red nanotech armor and a drinking problem in exactly this particular shape for compatibility.
- politician 4y ago
  I'll agree to your premise, but I'm relying on the combination of the recent guidance from the Copyright Office on AI generated output (ie: the output cannot receive copyright) with Clean Room techniques, optionally followed by a human transformation. I agree that if the output is a character named Iron Man, then no dice; if, however, the output is a new song that sounds similar to the original song but isn't the same, then I suspect that this will clear the bar for copyright.

blibble 4y ago

the AI in this situation is no different than gzipping a copyrighted file, ungzipping it and then trying to claim the output is no longer copyrighted

politician 4y ago
There have to be two AIs, one examines the original and produces a description while the second examines the description and produces a new work. Neither the description nor the new work are assigned copyright because they were produced by an AI (see [1]). However, if subsequently a human then manipulates the new work, the derived work is automatically protected by copyright assigned to that person (see [1, 2]). Furthermore, the new author has a defense against a copyright infringement claim by the owner of the original based on existing case law shared previously. This has nothing to do with gzip. [1] https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence https://www.federalregister.gov/documents/2023/03/16/2023-05... [2] https://en.wikipedia.org/wiki/Berne_Convention https://en.wikipedia.org/wiki/Berne_Convention
- blibble 4y ago
  no matter how many AIs you layer: it's still just a lossy compression function jpeg'ing frames of disney animation doesn't remove copyright (even if you do it twice)
  - CuriouslyC 4y ago
    Saying a model is like jpeg compression is like saying a van is like an elephant, in that maybe if you squint so hard you're almost blind it's kinda true.
  - politician 4y ago
    What matters here for copyright are the laws, not the math.
CuriouslyC 4y ago
It's quite different, because GZIP produces a lossy copy of an image, whereas models produce a distinct image that shares many aspects of style and composition. In one case it's obvious the product is the original image, in the other it's a derivative that may or may not be sufficiently different to be protected under fair use. I can write an algorithm that just generates random noise in the dimensions of art, and if I run it long enough it'll output things that are "close enough" to copyright works. There's no argument for that program being copyright infringement, and that holds for models as well.
- blibble 4y ago
  > In one case it's obvious the product is the original image, in the other it's a derivative that may or may not be sufficiently different to be protected under fair use. right, so if it's a completely original work let's clear out the training set and let's see if it can do it without it no? so the output is a product of the input... a derivative work and if you run it again it produces exactly the same thing? (sans artificial random injection) sounds like a lossy compression function to me!
  - CuriouslyC 4y ago
    I'd like to see a human make art after you lobotomize them.

kmeisthax 4y ago

Clean room reverse engineering is a thing because software copyright was a mistake and the judicial system understands it was a mistake. In software, it is regular and common for programmers to use licensed software libraries with defined interfaces. In writing or art, the notion of a "compatible interface" isn't really there. If you wrote a short fan story with Marvel characters and you want to liberate it from Disney's ownership, you have to redesign all the characters; you can't just argue that you need a superhero with red nanotech armor and a drinking problem in exactly this particular shape for compatibility.

politician 4y ago
I'll agree to your premise, but I'm relying on the combination of the recent guidance from the Copyright Office on AI generated output (ie: the output cannot receive copyright) with Clean Room techniques, optionally followed by a human transformation. I agree that if the output is a character named Iron Man, then no dice; if, however, the output is a new song that sounds similar to the original song but isn't the same, then I suspect that this will clear the bar for copyright.