It seems to me rather less likely that someone at Microsoft knowingly and deliberately took his specific diagram and "ran it through an AI image generator" than that someone asked an AI image generator to produce a diagram with a similar concept, and it responded with a chunk of mostly-memorized data, which the operator believed to be a novel creation. How many such diagrams were there likely to have been, in the training set? Is overfitting really so unlikely?

The author of the Microsoft article most likely failed to credit or link back to his original diagram because they had no idea it existed.

How you commit plagiarism is less important than the fact that you commit plagiarism.

Yes, but from OP's perspective this is a distinction without a difference.