3 ms·
There's a part that I don't understand. If some software is mirrored on github by someone that isn't the copyright owner, it seems like github shouldn't be able
by Zababa 5y ago
There's a part that I don't understand. If some software is mirrored on github by someone that isn't the copyright owner, it seems like github shouldn't be able to use it. Yet they said nothing about that specifically. In that case, is the only option to put code somewhere else than github under a license that forbids reuploading to github, and issue DMCAs when/if people reupload your code? It also sounds like when code is removed through DMCA, it should be removed from the training set and they should retrain copilot.
- dragonwriter 5y ago> If some software is mirrored on github by someone that isn't the copyright owner, it seems like github shouldn't be able to use it. If they don't need permission from the copyright owner, either via license of GitHub T&C, because it's fair use, which is their overt legal theory, then why would it matter legally whether the code was posted to GitHub at all, much less by whom? Ingesting only code form GitHub is a practical convenience that has nothing to do with their legal theory of the right to do it. > Yet they said nothing about that specifically Their theory of fair use means they have the right to ingest any code, irrespective of who owns it and what conditions (if any) it is licensed under or where (or even if) it is hosted online. They don't need a separate justification for your scenario if the theory they’ve cited is correct. > In that case, is the only option to put code somewhere else than github under a license that forbids reuploading to github, and issue DMCAs when/if people reupload your code Nope, that doesn't help at all, legally; it may help practically as long as they are just using GitHub hosted code and not consuming code from other public hosting platforms, but it has no bearing on their legal theory of why they can ingest code without additional permissions.
- Zababa 5y agoThanks for the clarification. So according to their theory, I could train a model on any code, even private Microsoft code, and that would be okay? That sounds surprising to me.