3 ms·

It's an interesting question. Arguably, any model trained on copyrighted works could be said to be a derivative of those works, as the model is composed entirel

It's an interesting question. Arguably, any model trained on copyrighted works could be said to be a derivative of those works, as the model is composed entirely of the copyrighted works run through a deterministic computational process. But on the other hand, we certainly wouldn't apply that same standard to a human who "learned" from copyrighted works, so where do we draw the line?

Ensorceled 7y ago
> But on the other hand, we certainly wouldn't apply that same standard to a human who "learned" from copyrighted works, so where do we draw the line? Are you actually having trouble figuring out the line between commercial ML product and human being?
- Ajedi32 7y ago
  No, I'm talking about the line between learning and memorization. Not the overall line between human intelligence and modern-day AI (where there are certainly obvious distinctions). Let's say, for example, that a human who was not a native-born English speaker taught themselves to speak English solely by listening to hundreds of copyrighted English fiction audio books, and comparing them to translations of those same works of fiction in their own native language. (Maybe not the easiest way to learn English, but entirely possible.) Now suppose that human used their newfound mastery of the English language to teach someone else to speak English. Since their knowledge of English was solely derived from copyrighted works, would their knowledge of English be considered a derivative work and therefore subject to copyright? Obviously not. Now suppose the same thing occurred with an AI. The AI learned to translate some other language into English by reading hundreds of thousands of translated books in both languages. Is the AI model a derivative work, and therefore subject to copyright?
  - Ensorceled 7y ago
    There are ton of things humans can legally do that the computers/corporations can't. I can memorize a copyrighted work (make a mental copy) without triggering a copyright infringement. I can remember a friends email address without triggering GDPR. Why is this suddenly a difficult question just because it's machine learning versus machine copying/storing? That still leaves the question about should ML training sets do/should count as copyright infringement, but that's completely distinct from the idea that a human can do it without causing a copyright infringement.
6gvONxR4sf7o 7y ago
A JPEG is also sorta kinda a derivative work in the abstract. It learns the common patterns in an image and makes a new image with those patterns. A neural network that is hugely overfitting will do the same. The legal (and in this case maybe moral?) line isn't drawn based on the algorithm. The line is drawn based on the output. Is the output sufficiently original and transformative?

Ensorceled 7y ago

> But on the other hand, we certainly wouldn't apply that same standard to a human who "learned" from copyrighted works, so where do we draw the line? Are you actually having trouble figuring out the line between commercial ML product and human being?

Ajedi32 7y ago
No, I'm talking about the line between learning and memorization. Not the overall line between human intelligence and modern-day AI (where there are certainly obvious distinctions). Let's say, for example, that a human who was not a native-born English speaker taught themselves to speak English solely by listening to hundreds of copyrighted English fiction audio books, and comparing them to translations of those same works of fiction in their own native language. (Maybe not the easiest way to learn English, but entirely possible.) Now suppose that human used their newfound mastery of the English language to teach someone else to speak English. Since their knowledge of English was solely derived from copyrighted works, would their knowledge of English be considered a derivative work and therefore subject to copyright? Obviously not. Now suppose the same thing occurred with an AI. The AI learned to translate some other language into English by reading hundreds of thousands of translated books in both languages. Is the AI model a derivative work, and therefore subject to copyright?
- Ensorceled 7y ago
  There are ton of things humans can legally do that the computers/corporations can't. I can memorize a copyrighted work (make a mental copy) without triggering a copyright infringement. I can remember a friends email address without triggering GDPR. Why is this suddenly a difficult question just because it's machine learning versus machine copying/storing? That still leaves the question about should ML training sets do/should count as copyright infringement, but that's completely distinct from the idea that a human can do it without causing a copyright infringement.

6gvONxR4sf7o 7y ago

A JPEG is also sorta kinda a derivative work in the abstract. It learns the common patterns in an image and makes a new image with those patterns. A neural network that is hugely overfitting will do the same. The legal (and in this case maybe moral?) line isn't drawn based on the algorithm. The line is drawn based on the output. Is the output sufficiently original and transformative?