Unsurprising... and also unwise
Modern aircraft are 'fly-by-wire' meaning the pilot commands are translated to electronic signals and the actual mechanical movements of the plane are controlled through the electronics. On the other hand they have highly accurate simulators which log far more missions than real flight hours, but with close to identical electronic systems. So it stands to reason that
a) fitting an aircraft to accept commands directly from the electronics is presumably quite easy technically.
b) they have tons and tons of training data that is both very clean, and also harvested from their top pilots (unlike other models trained on random datasets trawled from the interwebs.
So its completely unsurprising not only that they did this but that the AI could beat a real pilot in a dogfight.
Question is though, if they don't have anything in their training set for unexpected stuff happening or shit going wrong, how ill the AI respond.
That's quite apart from the generally bad idea of letting a computer autonomously control highly dangerous weapons