When you've reverted the same "correction" google docs made to a word at least ten times in the same doc on the same day, why is this never the kind of learning the machines are engaged in