4 ms·

Legally I believe this distinction is quite clear (IANAL) between these two cases. You can memorize a whole book, and it's not copyright infringement. If you,

by new299 7y ago

Legally I believe this distinction is quite clear (IANAL) between these two cases. You can memorize a whole book, and it's not copyright infringement. If you, create a compressed zip of book you that infringes if you don't own the copyright. If a trained model really doesn't infringe, then that's it copyright is dead. Because it's not difficult to overfit a model such that it can replicate the original work. And if that model is no longer under the original copyright it can be shared without restriction.

chii 7y ago
> If you, create a compressed zip of book you that infringes if you don't own the copyright. of course you don't - a compresssed zip, or any transformation of the text is a derivative work. But the neural network you use the text to train is not a derivative work unless you intentionally only used this single book to train it for the express purpose of reproducing the original text. In which case, that is no different than doing a format transformation, and that is protected under copyright. What i m saying is that any book that is used as a source of knowledge cannot claim derivative work on the resulting culmination of knowledge.
- new299 7y ago
  Why do you believe it's not a derivative work? Let's say a neural network was trained on two books, and it ends up reproducing large fragments of text from those two books. Is that a derivative work? Or 3 Or 4... And reproducing how much of the original work exactly, under what conditions. Does the size of the model factor in to the analysis? It doesn't seem as clear cut as you suggest to me.
  - roywiggins 7y ago
    Yeah, if you trained an AI to write Harry Potter fanfic (and remove all the trademarks by hand, no Harry Potter(TM) etc), is the result a derivative work? It seems like it could be, even if it's generic child wizards at a generic wizarding school. Fanfic is generally a derivative work.

by new299 7y ago

chii 7y ago

> If you, create a compressed zip of book you that infringes if you don't own the copyright. of course you don't - a compresssed zip, or any transformation of the text is a derivative work. But the neural network you use the text to train is not a derivative work unless you intentionally only used this single book to train it for the express purpose of reproducing the original text. In which case, that is no different than doing a format transformation, and that is protected under copyright. What i m saying is that any book that is used as a source of knowledge cannot claim derivative work on the resulting culmination of knowledge.

new299 7y ago
Why do you believe it's not a derivative work? Let's say a neural network was trained on two books, and it ends up reproducing large fragments of text from those two books. Is that a derivative work? Or 3 Or 4... And reproducing how much of the original work exactly, under what conditions. Does the size of the model factor in to the analysis? It doesn't seem as clear cut as you suggest to me.
- roywiggins 7y ago
  Yeah, if you trained an AI to write Harry Potter fanfic (and remove all the trademarks by hand, no Harry Potter(TM) etc), is the result a derivative work? It seems like it could be, even if it's generic child wizards at a generic wizarding school. Fanfic is generally a derivative work.