MIT researchers have detailed AI’s failure to grasp negation, a basic yet critical human language function that could prove dangerous.