3 ms·

This is not the legal argument used for AI training. The legal argument that is used is that the way AI training works is in fact specifically permitted by copy

This is not the legal argument used for AI training. The legal argument that is used is that the way AI training works is in fact specifically permitted by copyright law (in the EU), or by fair use (in the US). This HN thread has a bit more detail, including links: https://news.ycombinator.com/item?id=37879938 https://news.ycombinator.com/item?id=37879938 If you don't want to click through and just read the argument presented by OpenAI on this topic, you can find it here: https://www.uspto.gov/sites/default/files/documents/OpenAI_RFC-84-FR-58141.pdf https://www.uspto.gov/sites/default/files/documents/OpenAI_R...

xigoi 3y ago
I'm not talking about how it works, but how it should work.
- Kim_Bruning 3y ago
  > "If you train on copyrighted data because you don't believe in copyright" OpenAI may or may not believe in copyright, but they don't believe it applies in their case. From a different angle the US Copyright Office tends to agree, not accepting copyright registration for AI works. [2] We may end up agreeing that AI models and their output should be open source or even public domain. However, I would not argue that on the basis of "because the makers do not believe in copyright". I think the copyright argument might be orthogonal in general. [2] 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...
  - xigoi 3y ago
    The linked thread argues that if you don't keep or distribute an exact copy of the copyrighted work, then you're fine. So by this logic, if I pirate a movie, compress it with a lossy algorithm and send it to my friends, that is also fine, since they can't recover the original movie from what I sent them?
    - Kim_Bruning 3y ago
      It's certainly not allowed to be an exact copy. It's also not simply allowed to be an inexact copy, mind. For your specific example, it depends on how inexact the "copy" actually is. The four factors of fair use are: 1) the purpose and character of your use 2) the nature of the copyrighted work 3) the amount and substantiality of the portion taken, and 4) the effect of the use upon the potential market. At some point your algorithm might be so lossy that the court rules that the amount and substantiality of what is taken is sufficiently low that your output may count as fair use. (eg. if the end result is something like [1] ) Say we argue for the sake of argument that -say- Stable Diffusion is some sort of lossy compression algorithm. it turns out that -on average- SD only notes down a few bits per image that it trains on. Thus it easily meets the amount and substantiality factor. If you argue that SD is not a compression algorithm (and I really don't think it is), you may start to meet some of the other factors as well. The analysis under EU law is a bit different, since there's no fair use per-se. But there's a separate rule for scraping which applies. [1] https://fairuse.stanford.edu/2014/12/22/much-photo-need-alter-avoid-copyright-infringement-hint-cheshire-cat/ https://fairuse.stanford.edu/2014/12/22/much-photo-need-alte...

xigoi 3y ago

I'm not talking about how it works, but how it should work.

Kim_Bruning 3y ago
> "If you train on copyrighted data because you don't believe in copyright" OpenAI may or may not believe in copyright, but they don't believe it applies in their case. From a different angle the US Copyright Office tends to agree, not accepting copyright registration for AI works. [2] We may end up agreeing that AI models and their output should be open source or even public domain. However, I would not argue that on the basis of "because the makers do not believe in copyright". I think the copyright argument might be orthogonal in general. [2] 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...
- xigoi 3y ago
  The linked thread argues that if you don't keep or distribute an exact copy of the copyrighted work, then you're fine. So by this logic, if I pirate a movie, compress it with a lossy algorithm and send it to my friends, that is also fine, since they can't recover the original movie from what I sent them?
  - Kim_Bruning 3y ago
    It's certainly not allowed to be an exact copy. It's also not simply allowed to be an inexact copy, mind. For your specific example, it depends on how inexact the "copy" actually is. The four factors of fair use are: 1) the purpose and character of your use 2) the nature of the copyrighted work 3) the amount and substantiality of the portion taken, and 4) the effect of the use upon the potential market. At some point your algorithm might be so lossy that the court rules that the amount and substantiality of what is taken is sufficiently low that your output may count as fair use. (eg. if the end result is something like [1] ) Say we argue for the sake of argument that -say- Stable Diffusion is some sort of lossy compression algorithm. it turns out that -on average- SD only notes down a few bits per image that it trains on. Thus it easily meets the amount and substantiality factor. If you argue that SD is not a compression algorithm (and I really don't think it is), you may start to meet some of the other factors as well. The analysis under EU law is a bit different, since there's no fair use per-se. But there's a separate rule for scraping which applies. [1] https://fairuse.stanford.edu/2014/12/22/much-photo-need-alter-avoid-copyright-infringement-hint-cheshire-cat/ https://fairuse.stanford.edu/2014/12/22/much-photo-need-alte...