Post

I think you missed perhaps the best guess for copyright (at least in the US), which is not engineer intuitive: It may be that LLMs are not derived works of the training data (indeed they may be not works at all!), but having one write recognizably Harry Potter makes the output a derived work of HP.

That kind of copyright does not work at all for software. I can write software that is basically identical to yours, have same function and variable names and functionality. But, as long, as I have not seen your source code, there is no copyright violation. en.wikipedia.org/wiki/Google_....