At Georgia Tech, researchers have found a way to reduce how often AI chatbots reinforce delusions.