The output isn't guaranteed to be free of copyright from its training materials. It just usually is. There have been clear demonstrations of it regurgitating code from the training set verbatim, which would of course still be covered by the original license.

Microsoft isn't going to train Copilot on Windows code for the same reason it didn't train it on private repos: the code is private and they don't want to risk leaking private code.

I imagine there would be no problem training it on e.g. the Unreal Engine code which is not open source but is available to read.

The big practical issue is that there's no warning when Copilot produces code that might violate copyright, so you're taking a bit of a risk if you use it to generate huge chunks of code as-is. I imagine they are working on a solution to that though. It's not difficult conceptually.