Going by the uncanny-valley accuracy of their suggested auto-replies, they're training their natural language neural nets with every message sent over their wires.