I think this interpretation works if the code being generated is seen as essentially being retrieved by a lossy lookup function.
But another interpretation is that the generic structure of the code was learned from the works, which is not copywritable. And that generic structure was used to synthesize new code, in much the same way a human who had seen a pattern in a proprietary codebase years ago was able to use that pattern in their own code. I am not a lawyer but most licenses do not prohibit that in my experience. More often in my experience this is what is happening with generative ai.
The tricky bit is that the ai can probably do both in the eyes of copywrite law, since the boundary seems to be very context dependent and existing models don’t have any concept of how much you need to compress and forget the specific details so that it is seen as novel by the courts. The model can memorize significant parts of some inputs despite not having nearly enough space for memorizing the input set, so the first interpretation is possible even if it isn’t the typical output. There isn’t really a kind of “courts will see this as novel” regularizer and there might need to be?