The Windows source code is almost certainly part of the training data given the leaks. You can use LLMs trained on that to make your own OS if you like.
For the same reason that I think it is (1) a bad idea to use an LLM to make your own OS, (2) when the AI is good enough to do that, it won't matter which OSes are or are not open-sourced, nor which OS's source code was used to train a model, as by that point the AI can make any OS for peanuts anyway — given those two points, I think the copyright holders fighting against their stuff teaching AI are missing the forest for the trees.
For the same reason that I think it is (1) a bad idea to use an LLM to make your own OS, (2) when the AI is good enough to do that, it won't matter which OSes are or are not open-sourced, nor which OS's source code was used to train a model, as by that point the AI can make any OS for peanuts anyway — given those two points, I think the copyright holders fighting against their stuff teaching AI are missing the forest for the trees.