3 ms·

> The existing licenses cover AI training just fine [...] An AI product - more specifically, the model weights - is a derivative work of the original works used

> The existing licenses cover AI training just fine [...] An AI product - more specifically, the model weights - is a derivative work of the original works used for training The presumption is entirely debatable. A human is not considered derivative work of the original works they used for training. > Any other interpretation of the nature of copyright [...] which by definition, only covers human produced material Maybe copyright needs to leave the 1980s and evolve to deal with AI too? Maybe you do, too?

hulitu 3y ago
It deals with "AI" just fine. But not in the way Microsoft, for the extinguish phase, wants.
JoshTriplett 3y ago
A human is not copyrightable. And copyright is doing its job just fine in this context; AI training committing Open Source license violations en masse is the problem here. I'm all for copyright becoming substantially weaker or ceasing to exist, at which point AI training and lots of other things gets easier. As long as it does exist, however, AI training must respect it, and not become a copyright-violation laundering mechanism.
pessimizer 3y ago
> A human is not considered derivative work of the original works they used for training. No one has forced you to use humans as a comparison. A human is a citizen with other rights, and can own its own copyrights. Yet, you can still sue one for singing a song in front of other people. There were years of cinema that were distorted by the inability to have characters sing "Happy Birthday" to each other. edit: Suddenly, fair use now covers the ability to reproduce copyrighted material almost verbatim, but only when the new method to do so takes tens of millions of dollars of computer time to take advantage of.
- circuit10 3y ago
  AI models don't always copy things verbatim. Clearly if they do and then you use that it's copyright infringement, but Google isn't illegal just because you can search for code that's under a license agreement (not that AI models are search engines). Then again, you can't download Google's index, but you also can't for many AI models... there's a lot of nuance here, I don't think it's clear one way or the other

hulitu 3y ago

It deals with "AI" just fine. But not in the way Microsoft, for the extinguish phase, wants.

JoshTriplett 3y ago

A human is not copyrightable. And copyright is doing its job just fine in this context; AI training committing Open Source license violations en masse is the problem here. I'm all for copyright becoming substantially weaker or ceasing to exist, at which point AI training and lots of other things gets easier. As long as it does exist, however, AI training must respect it, and not become a copyright-violation laundering mechanism.

pessimizer 3y ago

> A human is not considered derivative work of the original works they used for training. No one has forced you to use humans as a comparison. A human is a citizen with other rights, and can own its own copyrights. Yet, you can still sue one for singing a song in front of other people. There were years of cinema that were distorted by the inability to have characters sing "Happy Birthday" to each other. edit: Suddenly, fair use now covers the ability to reproduce copyrighted material almost verbatim, but only when the new method to do so takes tens of millions of dollars of computer time to take advantage of.

circuit10 3y ago
AI models don't always copy things verbatim. Clearly if they do and then you use that it's copyright infringement, but Google isn't illegal just because you can search for code that's under a license agreement (not that AI models are search engines). Then again, you can't download Google's index, but you also can't for many AI models... there's a lot of nuance here, I don't think it's clear one way or the other