In the US copyright violation is a strict liability statute. Regardless of whether or not a court directly confirms or denies Microsoft's right to use code in that way, the end developer is still liable for whatever he or she uses.

But that's orthogonal to being able to use whatever you want as training data for an AI.

Has the exact issue of a remixing AI been tested in court? No. But everything even remotely similar has been deemed legal. Considering the legal and financial backing on both sides of the issue I expect it to go Microsoft's way even if it does end up before a judge.

It clearly fails the

> the revelations did not provide a significant market substitute for the protected aspects of the originals.

test of your cited case law though. The courts clearly drew a line at developing AI to inject snippets of copy-written works in similar copy-written works. And in context it would be the developer of the AI at fault (in addition to the end users who also used it to infringe other works in the creation of partially derived works; multiple parties can be at fault).

Basically the courts are making it pretty clear that they would have been against what Google had made if it were suggesting phrases in a plugin for a word processing program to create books that would compete with the original books. But being a separate domain of simply collating existing books and providing better search for their corpus (which led you to the original) was allowed.