“A superintelligent AI could cause massive harm as an unintended consequence of pursuing a human-set goal by creating instrumental subgoals that conflict with human well-being.”