Members of an AI watchdog group have warned in a new book that computational systems could eventually become self-aware to an extent that would be deadly to humanity