• There’s a few logical counterpoints to this:

    1. Training on copyrighted materials is not necessarily something that copyright law prohibits. NYT vs OpenAI should hopefully help clarify this. I think there’s a reasonable argument that training is fair use. Which would mean that training on GPL code doesn’t automatically trigger anything in the GPL license regarding derivative works.

    2. Copyright covers the expression, in general unless it’s a reproduction, a lot of code won’t be the same expression of things or is the simplest obvious way to do something, which means it may not be infringing in the first place. This isn’t necessarily even a fair use question.

    3. Even when AI generates some non-copyrightable code, there’s still the question of whether there’s enough human contribution to push it into being an expression of human creativity. This is hard to determine in bulk and if you get it wrong by assuming that all AI generated code is uncopyrightable then you risk being sued. E.g. if I take something public domain and paint over it, the work is copyrightable, even though the thing behind it is not. This is the concept of a work containing both copyrightable and non-copyrightable material.

    4. Even if AI generated code isn’t copyrightable, that doesn’t mean there’s any obligation to publish it or make it open source. Public domain and open source aren’t the same thing, and nothing about a work being uncopyrightable means it has to be made publicly available.


    score 2 · canonical