6 ms·

But the copy used during training is itself transient, so all this boils down really to the question of whether training a machine is fair use. Which can't be a

But the copy used during training is itself transient, so all this boils down really to the question of whether training a machine is fair use. Which can't be answered here exactly because the concept of fair use is deliberately vague, so this will boil down to a lawsuit and probably go to the Supremes. The USA will work something out that's reasonable as they always do and, lacking AI companies and often the concept of fair use to begin with, the rest of the world will never work out the necessary case law and fall even further behind. Possible exception: UK, which does have some significant LLM companies and also fair use law, albeit as is often the case with the UK these firms are US/UK hybrids with significant presence in both countries and the legal HQ is usually in the USA.

bandrami 3y ago
It's not even a question of "fair use" because OpenAI isn't providing anybody a copy. Probability models really just aren't copyrightable to begin with.
- mike_hearn 3y ago
  Fair use also covers generating derived works, and arguably the AI is a derived work.
  - bandrami 3y ago
    That would be hard to argue. You can't point to a sequence of words that's in the original that was copied into the OpenAI output.
    - mike_hearn 3y ago
      That's not how derivative works are defined. For example, translations to other languages or extensive summarizations (condensations) count as derivative works even though no words are directly copied.
- dragonwriter 3y ago
  > It’s not even a question of “fair use” because OpenAI isn’t providing anybody a copy “Fair Use” can apply to essentially any of the exclusive rights under copyright, including making (with or without distributing) a copy or derivative work. > Probability models really just aren’t copyrightable to begin with. “Probability models” that are built from copyrightable works either: (1) have a sufficient human creative input to be copyrightable on its own (in which caee it still may infringe copyrights applicable to its source material as a derivative work), or (2) do not have a sufficient human creative input to be copyrightable, and thus are a form of mechanical copy of their data from which they are developed (which, to the extent it either is, in aggregate, a copyrighted work, and/or contains copies of other copyrighted works, is protected by one or more copyrights, which do not cease to apply to the mechanical copy, and which an unlicensed mechanical copy would violate unless it fell into an exception like Fair Use.)
  - bandrami 3y ago
    But OpenAI is neither copying nor deriving a work. Style (which can be described as a probability model) is not copyrightable. You're halfway there with #2. The output is not copyrightable, but unless you can actually point to a sequence of words from the original it can't be infringing.
    - dragonwriter 3y ago
      > But OpenAI is neither copying nor deriving a work. They absolutely are copying in the course of training the model, and they are doing something that often looks a lot like copying when producing output with the model. > Style (which can be described as a probability model) is not copyrightable. Style is not all that can be described in an LLM's "probability model", otherwise LLM models would never be able to reproduce content from the training set, which they do. > The output is not copyrightable, but unless you can actually point to a sequence of words from the original it can’t be infringing. While verbatim copying (“a sequence of words from the original”) is infringing, mechanical copying with a mechanical substitution filter (even a wickedly complex one) is also copying, and violates copyright. And ChatGPT is quite capable of producing verbatim text from its training set.

bandrami 3y ago

It's not even a question of "fair use" because OpenAI isn't providing anybody a copy. Probability models really just aren't copyrightable to begin with.

mike_hearn 3y ago
Fair use also covers generating derived works, and arguably the AI is a derived work.
- bandrami 3y ago
  That would be hard to argue. You can't point to a sequence of words that's in the original that was copied into the OpenAI output.
  - mike_hearn 3y ago
    That's not how derivative works are defined. For example, translations to other languages or extensive summarizations (condensations) count as derivative works even though no words are directly copied.
dragonwriter 3y ago
> It’s not even a question of “fair use” because OpenAI isn’t providing anybody a copy “Fair Use” can apply to essentially any of the exclusive rights under copyright, including making (with or without distributing) a copy or derivative work. > Probability models really just aren’t copyrightable to begin with. “Probability models” that are built from copyrightable works either: (1) have a sufficient human creative input to be copyrightable on its own (in which caee it still may infringe copyrights applicable to its source material as a derivative work), or (2) do not have a sufficient human creative input to be copyrightable, and thus are a form of mechanical copy of their data from which they are developed (which, to the extent it either is, in aggregate, a copyrighted work, and/or contains copies of other copyrighted works, is protected by one or more copyrights, which do not cease to apply to the mechanical copy, and which an unlicensed mechanical copy would violate unless it fell into an exception like Fair Use.)
- bandrami 3y ago
  But OpenAI is neither copying nor deriving a work. Style (which can be described as a probability model) is not copyrightable. You're halfway there with #2. The output is not copyrightable, but unless you can actually point to a sequence of words from the original it can't be infringing.
  - dragonwriter 3y ago
    > But OpenAI is neither copying nor deriving a work. They absolutely are copying in the course of training the model, and they are doing something that often looks a lot like copying when producing output with the model. > Style (which can be described as a probability model) is not copyrightable. Style is not all that can be described in an LLM's "probability model", otherwise LLM models would never be able to reproduce content from the training set, which they do. > The output is not copyrightable, but unless you can actually point to a sequence of words from the original it can’t be infringing. While verbatim copying (“a sequence of words from the original”) is infringing, mechanical copying with a mechanical substitution filter (even a wickedly complex one) is also copying, and violates copyright. And ChatGPT is quite capable of producing verbatim text from its training set.