Artificial intelligence systems are more harsh on people who violated hypothetical rules or policies than judgement calls from humans, a study from MIT researchers shows.