4 ms·

> I sincerely hope that one of the many court cases produces a verdict that says AI-generated code is, in fact, subject to the licenses of the inputs. Such a s

> I sincerely hope that one of the many court cases produces a verdict that says AI-generated code is, in fact, subject to the licenses of the inputs. Such a strict interpretation of copyright would kill Open Source, as any attempt at reverse engineering would be disallowed by that. Furthermore it would mean that only big and rich companies would have AI, as they can do behind closest doors whatever they want with AI, Open Source licenses only cover redistribution after all, which internal use doesn't fall under. Meanwhile any attempt at publicly available open AI models would instantly get killed by copyright claims.

JoshTriplett 3y ago
> Such a strict interpretation of copyright would kill Open Source, as any attempt at reverse engineering would be disallowed by that. Not at all, for multiple reasons. Reverse-engineering already has that problem: if any of the people who do the reverse engineering also work on the code, it's entirely possible for some of the reverse-engineered code to end up in the new code, making the new code a derivative work of the original. There are standard ways to carefully avoid that: https://en.wikipedia.org/wiki/Clean_room_design https://en.wikipedia.org/wiki/Clean_room_design (Also, reverse engineering is a small fraction of Open Source.) > Furthermore it would mean that only big and rich companies would have AI, as they can do behind closest doors whatever they want with AI, Open Source licenses only cover redistribution after all, which internal use doesn't fall under. The moment they distribute anything written by the model, the same problem applies. And if they don't redistribute anything written by the model, then sure, they can do anything they like, just as you're free to internally combine GPLed and proprietary code if you never ship the result. (Note, though, that many companies have figured out it's a bad idea to do this, because it creates a combination you cannot ever distribute, and circumstances might change in the future to lead you to want to distribute it.) > Meanwhile any attempt at publicly available open AI models would instantly get killed by copyright claims. No, they just need to actually pay attention to the licenses of work they train on. Train on permissively licensed code, document every codebase trained on, and record the licenses and copyright notices.

JoshTriplett 3y ago

> Such a strict interpretation of copyright would kill Open Source, as any attempt at reverse engineering would be disallowed by that. Not at all, for multiple reasons. Reverse-engineering already has that problem: if any of the people who do the reverse engineering also work on the code, it's entirely possible for some of the reverse-engineered code to end up in the new code, making the new code a derivative work of the original. There are standard ways to carefully avoid that: https://en.wikipedia.org/wiki/Clean_room_design https://en.wikipedia.org/wiki/Clean_room_design (Also, reverse engineering is a small fraction of Open Source.) > Furthermore it would mean that only big and rich companies would have AI, as they can do behind closest doors whatever they want with AI, Open Source licenses only cover redistribution after all, which internal use doesn't fall under. The moment they distribute anything written by the model, the same problem applies. And if they don't redistribute anything written by the model, then sure, they can do anything they like, just as you're free to internally combine GPLed and proprietary code if you never ship the result. (Note, though, that many companies have figured out it's a bad idea to do this, because it creates a combination you cannot ever distribute, and circumstances might change in the future to lead you to want to distribute it.) > Meanwhile any attempt at publicly available open AI models would instantly get killed by copyright claims. No, they just need to actually pay attention to the licenses of work they train on. Train on permissively licensed code, document every codebase trained on, and record the licenses and copyright notices.