Training is almost certainly fair use, so it's exactly a transitory copy in service of fair use. Training, other than the brief "transitory copy" you mention is not copying, it's making a minuscule algorithmic adjustment based on fleeting exposure to the data.
Congress took the circuit holding in MAI Systems seriously enough to carve out a new fair use exception for copying software—entirely within the memory system of a licensed user—in service of debugging it.
If it took an act of Congress to make “unlicensed” debugging a fair use copy…
If Microsoft truly believes that the trained output doesn't violate copyright then it should be forced to prove that by training it on all its internal source code, including Windows.