3 ms·

There's going to be an argument about "derivative works". If I feed all 7 books of Harry Potter into an AI, and only that, and then it spits out something - Tha

There's going to be an argument about "derivative works". If I feed all 7 books of Harry Potter into an AI, and only that, and then it spits out something - That's clearly a derivative work. If I feed all of the New York Times Bestsellers for the past 50 years into an AI and it spits out something, there's much less of an argument that it's derivative of any one work. I've seen stuff like https://www.thiswaifudoesnotexist.net/ https://www.thiswaifudoesnotexist.net/ - It's both impressive and not. Sure, it's "new" and "Unique" every time, but if you were to take a trained eye to it, you'd see that a bunch of what it's done are very simple derivatives of an already derivative artform. Do I know which image it's stolen the basic shape from and which it's stolen the color palette? No, but if I spent a few hours looking through the training dataset I'd probably find another image that looked almost exactly like the one I'm looking at. These things are impressive because they can fool a cursory glance, and quite frankly, that's often enough.

PeterisP 7y ago
"If I feed all 7 books of Harry Potter into an AI, and only that, and then it spits out something - That's clearly a derivative work." does not match the current legal interpretation in the industry. The general assumption is that machine learning models trained on some data are (usually - details matter) not considered derivative work from that data; so unless you have specific contractual restrictions (e.g. you got the date bacause you have a contract with the data owner saying what you'll do/not do with the models derived from it) then copyright does not prohibit you from using and distributing these models. That's the case even if it's trained on a particular dataset with a single, clear copyright owner. It probably starts with old precedent on copyrightability of statistics derived from creative work - things like word frequency, most common words, etc are not considered derivative works. Then we have statistical language models, of the kind that were used in statistical machine translation before the neural approaches overtook everything - and again, the established interpretation there is that statical models trained on a corpus are not considered derivative work from that data, because they essentially are a trivial extension of word sequence frequency counts... but they already are in the category of "feed all 7 books of Harry Potter into an AI, and only that, and then it spits out something", e.g. a simple hidden Markov model or trigram language model from Harry Potter books will easily hallucinate (very lousy quality) Potter-themed text, but the model that can do that is not considered derivative work. Though this all may vary in different jurisdictions.
- GauntletWizard 7y ago
  "does not match the current legal interpretation in the industry." is the complete proof that the industry is fraudulent, vile, and sociopathic.
gwern 7y ago
> No, but if I spent a few hours looking through the training dataset I'd probably find another image that looked almost exactly like the one I'm looking at. If you look at nearest-neighbor lookups in papers like BigGAN, you'll see that for all the apparent 'simple derivatives', you will see that it is actually quite difficult to find "another image that looked almost exactly like the one I'm looking at". Similar, certainly, recognizably the same character, absolutely (particularly well-represented ones like Touhou characters), but exactly? And if they were purely copying specific points, the interpolations would not work. (It would also be a lot easier to reverse a given image into the latent vector to edit it.)

PeterisP 7y ago

"If I feed all 7 books of Harry Potter into an AI, and only that, and then it spits out something - That's clearly a derivative work." does not match the current legal interpretation in the industry. The general assumption is that machine learning models trained on some data are (usually - details matter) not considered derivative work from that data; so unless you have specific contractual restrictions (e.g. you got the date bacause you have a contract with the data owner saying what you'll do/not do with the models derived from it) then copyright does not prohibit you from using and distributing these models. That's the case even if it's trained on a particular dataset with a single, clear copyright owner. It probably starts with old precedent on copyrightability of statistics derived from creative work - things like word frequency, most common words, etc are not considered derivative works. Then we have statistical language models, of the kind that were used in statistical machine translation before the neural approaches overtook everything - and again, the established interpretation there is that statical models trained on a corpus are not considered derivative work from that data, because they essentially are a trivial extension of word sequence frequency counts... but they already are in the category of "feed all 7 books of Harry Potter into an AI, and only that, and then it spits out something", e.g. a simple hidden Markov model or trigram language model from Harry Potter books will easily hallucinate (very lousy quality) Potter-themed text, but the model that can do that is not considered derivative work. Though this all may vary in different jurisdictions.

GauntletWizard 7y ago
"does not match the current legal interpretation in the industry." is the complete proof that the industry is fraudulent, vile, and sociopathic.

gwern 7y ago

> No, but if I spent a few hours looking through the training dataset I'd probably find another image that looked almost exactly like the one I'm looking at. If you look at nearest-neighbor lookups in papers like BigGAN, you'll see that for all the apparent 'simple derivatives', you will see that it is actually quite difficult to find "another image that looked almost exactly like the one I'm looking at". Similar, certainly, recognizably the same character, absolutely (particularly well-represented ones like Touhou characters), but exactly? And if they were purely copying specific points, the interpolations would not work. (It would also be a lot easier to reverse a given image into the latent vector to edit it.)