One of the points being made is that in the worst-case scenario of getting Copilot to repeat back verbatim chunks of code from projects, something that's not it

One of the points being made is that in the worst-case scenario of getting Copilot to repeat back verbatim chunks of code from projects, something that's not its primary use case, it would be a situation similar to a copy machine. You can copy a page out of a book, or the whole book, and be covered under fair use. But you can't sell your copy on Amazon. And if you did, the copy machine nor Xerox ran afoul of copyright law. You could also use a copy machine to copy fragments of the Linux kernel source out of a book about the Linux source and use them to construct an entirely original work that's not considered derivative. The devil's in the details, but GitHub talks at some length about the plagiarization issue and their plans to detect and link back to where verbatim chunks exist in the training data to let the operator decide what to do soo.. IDK.